How Google AI Mode actually works, from query to answer

Type a question into the regular Google search box and, for most queries, you still get the familiar page: ads at the top, a possible AI Overview summary, then ten blue links. Tap the AI Mode tab next to it, or switch to it inside the Google app, and the page changes shape entirely. There is no ranked list of links waiting underneath a summary. There is a single scrollable answer, built specifically for the question you asked, with a text box at the bottom inviting a follow-up. That switch in layout is the clearest sign of what AI Mode is: not a feature bolted onto search results, but a separate mode of search with its own interface, its own retrieval behaviour, and its own reasoning model underneath it.

What AI Mode actually is inside Google Search

AI Mode is a dedicated tab inside Google Search that answers questions through a multi-step, conversational process instead of a ranked list of pages. Google introduced it in the United States in March 2025 as a Search Labs experiment, moved it to general availability without a Labs opt-in at Google I/O in May 2025, and by late 2025 had pushed it to more than 200 countries and territories in close to 100 languages. Structurally, it sits next to AI Overviews rather than replacing it. AI Overviews still generates the short AI-written summary that appears above organic results for select queries on the standard results page. AI Mode is the separate destination people go to when a question needs more than a summary — comparison shopping across a dozen criteria, planning a multi-day trip, or working through a technical problem that requires several linked sub-questions.

The distinction Google draws internally is about depth of reasoning, not just interface. A short factual query such as a capital city or a unit conversion barely touches AI Mode’s more elaborate machinery; the system recognises it as simple and answers close to instantly. A layered question — “which of these three laptops handles video editing best without going over budget, and which one has the better return policy” — triggers a longer process: breaking the question into parts, running searches for each part, checking the results against each other, and writing a single answer that references all of it. That longer process is what most of this analysis is about, because it is where AI Mode diverges most sharply from a traditional search engine and from a plain chatbot.

AI Mode is built on a version of Gemini tuned specifically for search, not a general-purpose chat model wearing a search engine’s clothing. Google has described it as a “custom version of Gemini” since the original March 2025 announcement, and successive updates have swapped in newer Gemini generations as they became available — Gemini 3 arrived in the back end during 2025, and Gemini 3.5 Flash became the global default inside AI Mode in May 2026. The practical effect of that tuning is that the model is trained and evaluated specifically on retrieval-linked reasoning: deciding what to search for, when to search again, and how to weigh five partially conflicting sources against each other, rather than purely on open-ended conversation.

It is worth being precise about what AI Mode does not do, because the myths accumulate quickly around any new AI product. It does not run arbitrary code on a user’s behalf the way a general coding assistant might; there is no exposed interpreter for users to invoke. It does not have unlimited memory of a person’s life by default — the deeper personal-context features are opt-in and clearly labelled, covered later in this analysis. And it is not, contrary to a common assumption, simply AI Overviews with a bigger text box. The two features share a model family and some retrieval infrastructure, but they are triggered differently, they draw on different depths of the fan-out process described in the next section, and Google treats them as genuinely separate products in its own documentation.

The single most useful mental model for AI Mode is a research assistant that is only as good as the questions it asks itself. A person doing serious research on, say, a home safe purchase does not type one search query and stop. They search for fire ratings, then insurance implications, then specific brands, then reviews, cross-referencing as they go. AI Mode’s underlying design goal is to compress that multi-step human research behaviour into a single conversational turn, running the equivalent of many searches in parallel and stitching the results into one coherent, cited answer. Whether it succeeds at that goal consistently — and where it does not — is the subject of most of what follows.

The three-stage pipeline behind every AI Mode answer

Strip away the marketing language and AI Mode’s process breaks into three stages that repeat, sometimes more than once, for every substantial query: understanding the question, retrieving information to answer it, and writing a response that synthesises what was retrieved. None of this is unique to AI Mode in the abstract — it is the same basic shape as retrieval-augmented generation, the technique underpinning most serious AI search products since 2023. What is specific to AI Mode is how aggressively it applies that shape, and how much of the retrieval step happens outside a single query.

The first stage is intent parsing. When a question arrives, Google’s systems run it through natural-language processing that tries to establish what kind of question it is, how complex it is, and what would count as a satisfying answer. This stage decides how hard the rest of the pipeline needs to work. A query like “capital of Spain” is recognised as a simple, single-fact lookup and does not trigger extensive downstream processing. A query like “how to optimise website performance for a small e-commerce store on a shared budget” is recognised as compound — it contains several distinct sub-questions bundled into one sentence — and that recognition is what activates the heavier machinery described in the next section.

The second stage is retrieval, and it is the stage that differs most from a conventional search engine. Rather than running the user’s literal words against Google’s index once, the system decomposes the question and issues many searches in parallel, pulling from the live web index, Google’s Knowledge Graph, and specialised structured data sources including the Shopping Graph. This is the retrieval step that Google calls query fan-out, treated in detail in the next section, and it is the part of the pipeline that has drawn the most attention from search engine optimisation professionals, because it means a single user question can pull content from dozens of different pages that never appeared in a conventional results page for that literal query.

The third stage is synthesis: taking everything retrieved and writing one answer. The Gemini model reads through the retrieved passages, resolves overlaps and contradictions where it can, and produces prose that reads as a single coherent response rather than a stitched-together list of excerpts. Citations are attached to specific claims as the response is generated, linking back to the pages the model drew on. This is also the stage where errors creep in — a model can misattribute a claim to the wrong source, smooth over a genuine disagreement between two sources as if it did not exist, or draw a conclusion the retrieved passages do not actually support. Later sections on accuracy and citation behaviour return to this in more depth.

A fourth, less visible loop sits on top of these three stages: the system can decide it needs to go back and search again. If the first retrieval pass leaves gaps — a sub-question with thin or contradictory results — later Gemini generations can trigger a second or even third wave of fan-out queries specifically targeting the gap before writing the final answer. Google’s engineers have described this as the system checking its own work mid-process, closer to how a careful researcher would notice a hole in their notes and go back to fill it rather than writing around it. This iterative retrieval is part of why some AI Mode answers to complex questions take noticeably longer to appear than a standard search results page — the system may be running several rounds of search behind a single spinner.

None of this pipeline is static across query types. Google has been explicit that the depth of processing scales with the perceived complexity of the question, precisely so the system does not burn unnecessary computation — and unnecessary time — on queries that a simple lookup would answer just as well. This scaling matters commercially as much as technically: running the full fan-out-and-synthesis pipeline on every single query at Google’s scale, which now handles well over a billion AI Mode sessions a month on top of billions of ordinary searches, would be an enormous and largely wasted computational cost if applied indiscriminately to “what time is it in Tokyo.”

Understanding this three-stage shape — parse, fan out, synthesise, with an optional second retrieval loop — is the foundation for almost everything else in this analysis. The features layered on top of it, from Personal Intelligence to agentic booking to the advertising formats now appearing inside AI Mode responses, are all extensions of one or another stage in this same underlying pipeline rather than separate systems bolted on beside it.

Query fan-out and how one question becomes dozens

Google’s own name for its core retrieval technique is query fan-out, and the company has been unusually open about describing it, both in technical documentation and on stage. At Google I/O 2025, Search lead Elizabeth Reid told the audience directly: “AI Mode isn’t just giving you information — it’s bringing a whole new level of intelligence to search. What makes this possible is something we call our query fan-out technique.” She went on to describe the mechanism in plain terms: “Under the hood, Search recognizes when a question needs advanced reasoning. It calls on our custom version of Gemini to break the question into different subtopics, and it issues a multitude of queries simultaneously on your behalf.”

Query fan-out is the process of expanding one user question into many separate sub-queries, run at the same time, each aimed at a different facet of the original intent. Take a query about headphones with several stated requirements — over-ear design, Bluetooth, long battery life. AI Mode’s model recognises three explicit facets in that sentence: design, connectivity technology, and battery performance. It then infers related facets the user did not state but likely cares about, such as charging speed or portability, based on patterns learned from how people search around that topic. Each of those facets becomes its own search, dispatched in parallel against the live web, the Knowledge Graph, and specialised indexes such as Google’s Shopping Graph.

The technique is not exotic in principle — query decomposition has existed in information retrieval research for years — but the scale at which Google applies it is unusual. Search Engine Journal reported in mid-2025, citing Google’s own Robby Stein, that AI Mode’s fan-out process can issue dozens or even hundreds of background queries for a single user question when the system judges that deep reasoning is warranted, a mode Google internally distinguishes as Deep Search and covers separately later in this analysis. Even outside that heaviest mode, ordinary complex AI Mode queries routinely spawn a double-digit number of parallel sub-queries that never surface individually to the user — they exist purely as retrieval steps feeding the final synthesised answer.

This has a direct and well-documented consequence for anyone trying to understand why their website appears, or does not appear, inside an AI Mode answer. Because the visible query and the queries Google’s system actually runs against its index are frequently different strings entirely, a page can be pulled into an AI Mode answer without ever ranking for the exact phrase the user typed. Conversely, ranking well for a head term is no longer a reliable predictor of being cited, because the citation might depend on ranking for one of a dozen fan-out sub-queries the SEO team never explicitly targeted. Digiday’s reporting on the technique in mid-2025 quoted several heads of search-engine-optimisation teams at major publishers describing this shift bluntly: the metric that used to matter, ranking position for a target keyword, has become a much weaker signal of visibility inside AI Mode than it remains for classic search.

Fan-out queries are not random expansions; Google frames them as modelling the follow-up questions a thorough human researcher would ask next. The system draws on semantic understanding of the topic, aggregate patterns in how real users have searched around similar questions in the past, and what Google calls logical information architecture — essentially, a model of how a topic decomposes into standard sub-parts that experienced researchers in that domain would naturally check. For comparative or multi-criteria questions, this fan-out is described by Google as especially useful, because those are exactly the query types where a single search result page historically failed to satisfy the underlying need and forced the user into several separate follow-up searches themselves.

The technique also scales with how confident the system is that it has covered the question adequately. If early rounds of fan-out retrieval turn up thin or conflicting information on a particular facet, later model generations can trigger additional waves of queries specifically targeting that gap — the multiplying fan-out mentioned in the previous section. This is one of the more consequential architectural choices in AI Mode’s design, because it means the retrieval process is not fixed in advance; it adapts mid-query based on what the first pass of searches actually returned, closer to how a human researcher decides whether to dig further based on what the first few search results show them.

Search Engine Journal quoted Google’s Robby Stein putting the scale of this in blunt terms during a mid-2025 presentation: AI-powered search experiences including query fan-out now serve approximately 1.5 billion users each month across Google’s properties, drawing on data sources that include the live web as well as real-time systems such as the Shopping Graph, which Stein said updates roughly two billion times an hour. He described Google Search overall, in that context, as “the largest AI product in the world” — a framing that positions query fan-out not as an experimental feature but as core retrieval infrastructure now running underneath a meaningful share of all Google queries, AI Mode and AI Overviews alike.

The custom Gemini model running underneath AI Mode

Every stage of the pipeline described above depends on a language model doing the reasoning — deciding how to decompose a query, judging which retrieved passages matter, and writing the final answer. That model has never been a generic, off-the-shelf Gemini deployment. Google has consistently described AI Mode as running on a custom version of Gemini, tuned specifically for the search task rather than for open-ended conversation, coding, or the other jobs Gemini performs elsewhere in Google’s product line.

The underlying model generation has changed multiple times since AI Mode’s March 2025 launch, and each change has been tied to a broader Gemini release. Gemini 2.5, released in the first half of 2025, was built on the Gemini 1.5 architecture with substantial gains in mathematics, reasoning, and code, and it became the workhorse behind Google’s AI features across Search, Workspace, and the standalone Gemini app during that period. Google DeepMind continued iterating rapidly afterward: Gemini 3 arrived later in 2025, and by I/O 2026 Google announced that Gemini 3.5 Flash would become the new default model inside AI Mode globally, chosen specifically for combining strong reasoning performance with the low latency that a conversational search product needs to feel responsive rather than sluggish. Google’s own benchmark figures at that announcement claimed Gemini 3.5 Flash outperformed the prior Gemini 3.1 Pro on agentic and coding benchmarks including Terminal-Bench 2.1 and MCP Atlas, evidence Google used to justify shipping a nominally smaller “Flash” model as the new default over a larger “Pro” variant — speed and cost efficiency at scale outweighing the marginal reasoning gains of a heavier model, at Google’s query volume.

Why a search-specific tune matters, rather than just plugging in the latest general-purpose Gemini release, comes down to what the model needs to be good at. A conversational assistant answering from its own training knowledge needs to be persuasive, fluent, and broadly knowledgeable. A model built to sit inside a retrieval pipeline needs a different, narrower skill set: judging when it does not know something and needs to search rather than guess, decomposing compound questions into well-formed sub-queries that actually retrieve useful results, resolving disagreements between multiple retrieved sources, and writing prose that stays traceable back to specific citations rather than blending retrieved facts with the model’s own unsupported assumptions. Google’s search-specific tuning targets precisely this second skill set, and it is one reason AI Mode’s outputs tend to read differently from a general chatbot’s — more clipped, more citation-dense, and more willing to present several options side by side rather than commit confidently to one.

The model also carries specific multimodal capabilities that general search never needed before. AI Mode processes text, voice, and image input within the same conversational thread, which is why a person can start a session by uploading a photo of a plant and follow up with a typed question about its care requirements without starting over. This multimodal grounding is inherited directly from Gemini’s broader architecture — the same family of models Google uses for Gemini’s image understanding in the standalone app and in Google Lens — rather than being a separate vision system stitched on for search specifically.

Compute cost is the constraint that shapes almost every visible decision about which model version runs where. Running a heavier reasoning model on every one of the billions of queries AI Mode now handles monthly, with query volumes that Google itself has said more than doubled every quarter since launch, would be financially and operationally unsustainable regardless of the marginal quality gain. This is the direct explanation for why Google has repeatedly favoured “Flash” model variants, engineered explicitly for speed and cost efficiency, as the default inside AI Mode, reserving heavier reasoning modes such as Deep Search for queries the system judges complex enough to justify the extra latency and computation. The architecture, in other words, is not simply “the best available Gemini” running on every question; it is a tiered system that routes different amounts of model capability to different questions based on an upfront complexity judgement, an economic decision as much as a technical one.

From a Search Labs experiment to a global product

AI Mode did not launch as a finished product. It launched, deliberately, as an opt-in experiment inside Google Labs, which is Google’s standard mechanism for testing search features on a subset of willing users before deciding whether to ship them broadly. That sequencing matters for understanding how the product evolved, because several of its most consequential features — Personal Intelligence, agentic booking, information agents — followed the same Labs-first pattern, and that pattern is a reasonable predictor of how Google will continue introducing new AI Mode capabilities going forward.

Google introduced AI Mode in the United States in March 2025, gated behind a Search Labs opt-in. This was the phase in which Google gathered the earliest usage data and feedback that would shape the wider rollout, a deliberate choice the company has made with most major search-facing AI features going back to the original AI Overviews rollout the year before. At Google I/O in May 2025, the company announced that AI Mode would become available to all signed-in users in the United States without requiring the Labs opt-in — a meaningful signal that Google had moved from testing to committing.

The removal of the Labs gate did not happen instantly; the rollout to the full US user base took roughly three weeks after the I/O announcement, according to contemporaneous trade reporting, landing in mid-June 2025. From there the pace of international expansion accelerated sharply. On 24 June 2025, Google launched AI Mode within Search Labs in India — again gated at first — before removing that Labs requirement weeks later so the feature was available to all Indian searchers. The United Kingdom followed in July 2025. By August and September 2025, Google had extended availability into additional Asian and European markets, and on 8 September 2025 the company added language support for Hindi, Indonesian, Japanese, Korean, and Portuguese, a deliberate choice Google described as building localised understanding rather than simple machine translation of the English-language product.

The single largest expansion came on 7 October 2025, when Google announced AI Mode in more than 35 new languages and over 40 additional countries and territories in one coordinated push, bringing cumulative availability past 200 countries and territories including, notably, most of the European Union. Hema Budaraju, Google’s Vice President of Product Management for Search, described this as the most extensive geographical expansion since the original March 2025 launch. By November 2025, Google reported that AI Mode had reached 75 million daily active users and 100 million monthly active users — figures the company disclosed itself, without independent third-party verification, a caveat worth holding onto given how central usage-scale claims have become to how Google frames AI Mode’s importance.

The growth accelerated further into 2026. At Google I/O on 19 May 2026, Google announced that AI Mode had crossed one billion monthly active users — roughly a tenfold increase from the November 2025 figure in about six months, with query volumes that Google said had more than doubled every quarter since the original launch. AI Overviews, the separate but related summary feature living on the standard results page, was reported at the same event to be reaching roughly 2.5 billion monthly users, giving a sense of the relative scale between the two products even as both continued growing in parallel. Again, these figures come directly from Google’s own disclosures rather than independent measurement, and the company has not broken down what counts as an active user by frequency or depth of engagement — a one-billion-user figure could describe anything from occasional curious taps on the AI Mode tab to habitual daily reliance, and Google’s public statements do not distinguish between the two.

The rollout has not been uniform, and the gaps are informative. France remained excluded from the initial European wave in October 2025 over an unresolved standoff with French publishers about neighbouring rights and compensation for AI-driven summarisation of news content. Google told French publishers, according to reporting relayed via Ouest-France, that both AI Overviews and AI Mode would finally launch in France by 23 September 2026, under terms that reportedly include a publisher opt-out mechanism, transparency commitments about AI-driven traffic impact, and compensation structured under neighbouring-rights law — terms that go further than what publishers have secured in most other markets, with the notable exception of the United Kingdom, where the Competition and Markets Authority separately ordered Google to provide an opt-out mechanism in June 2026. France’s position as the last major Western European market without either feature is a direct product of the regulatory and publisher-relations friction covered in more depth later in this analysis, and it demonstrates that AI Mode’s rollout has been shaped as much by legal negotiation as by technical readiness.

The country-by-country rollout and where AI Mode still does not reach

Availability is not a single on-off switch, and treating it that way leads to a distorted picture of who actually has access to AI Mode and under what conditions. Three separate variables determine whether a given person, in a given country, sees the feature at all: geographic availability, language support, and account eligibility, and these three variables have moved on different timelines throughout the product’s life.

Geographic availability reached more than 200 countries and territories by late 2025, following the sequence described in the previous section — the United States first, India and the United Kingdom next, then a large coordinated European and Asian expansion in October 2025. That figure describes where the feature exists at all, not where every resident automatically has it turned on, since account-level eligibility rules and regional legal requirements layer additional restrictions on top of raw geographic availability.

Language support has expanded roughly in step with geography but was not identical to it. The September 2025 addition of Hindi, Indonesian, Japanese, Korean, and Portuguese was framed by Google explicitly as more than translation — Google has stated that extending AI Mode into a new language requires building localised understanding of how people in that language and cultural context actually phrase questions and what sources are considered authoritative there, rather than running the English-language pipeline through a translation layer at the input and output stages. By mid-2026, one industry tracker put total language coverage at close to 100 languages, alongside the billion-plus monthly user figure disclosed at that year’s I/O.

Account-level eligibility is where the picture gets genuinely uneven, and where several of the product’s more advanced features remain gated well behind general availability. The core AI Mode experience — text-based conversational search — is free and available to any signed-in user in a supported market. But several of the capabilities covered later in this analysis are restricted to paid subscription tiers. Personal Intelligence, the feature that connects Gmail and Google Photos to AI Mode responses, launched exclusively for Google AI Pro and AI Ultra subscribers in the United States in January 2026, in English only, before Google expanded it to a free, near-global feature within about five months — a notably fast trajectory even by the standards of Google’s usual feature-rollout pace. Information agents, the persistent background-monitoring capability announced at I/O 2026, launched first to AI Ultra subscribers only, a tier priced at $99.99 or $199.99 per month in the United States depending on the plan, with availability to the less expensive AI Pro tier promised for later in the summer of 2026. Agentic booking capabilities, discussed in a later section, similarly rolled out first to paying subscribers before any wider extension.

The European Union and the United Kingdom occupy a distinct regulatory position that has shaped both the pace and the terms of rollout. AI Overviews rolled out in Europe initially restricted to signed-in users aged eighteen and over, a constraint tied directly to EU regulatory considerations around minors and data processing that did not apply in the same form in the United States. The Digital Markets Act has added a further layer: the European Commission’s July 2026 fine against Google, discussed later in the section on regulatory pressure, specifically targeted self-preferencing behaviour in search results, and the ongoing antitrust complaint from European publishers explicitly names both AI Overviews and AI Mode as products under scrutiny for potentially breaching Article 102 of the Treaty on the Functioning of the European Union. France’s delayed launch, tied to neighbouring-rights negotiations rather than technical limitations, is the most visible example of how European legal frameworks have directly determined where and when the product appears, rather than availability being purely a function of Google’s engineering and localisation timeline.

The practical upshot for anyone assessing AI Mode’s real reach is that headline figures like “200 countries” or “one billion monthly users” describe the outer boundary of availability, not a uniform product experience inside that boundary. A free user in the United States gets a materially different set of capabilities than a free user in a market where Personal Intelligence has not yet expanded, and a paying AI Ultra subscriber in the United States has access to agentic features that remain entirely unavailable to the vast majority of the reported billion-plus monthly users elsewhere.

AI Mode compared with AI Overviews

The confusion between AI Mode and AI Overviews is common enough that it is worth resolving carefully, because the two products share a technology base but serve genuinely different purposes, trigger under different conditions, and require different strategies from anyone trying to be visible inside either of them.

AI Overviews are the AI-generated summary boxes that appear automatically above the ranked list of links on a standard Google search results page. They launched broadly in the United States in May 2024, roughly ten months before AI Mode existed at all, evolving out of what Google had previously called the Search Generative Experience during its earlier experimental phase. AI Overviews trigger selectively — Google decides, query by query, whether it is confident enough that a synthesised summary will satisfy the user’s intent immediately, and if so, generates one automatically without the user requesting it. The user does not opt into AI Overviews the way they opt into AI Mode; it simply appears, embedded in the results they were already going to see, when Google’s systems judge the query calls for it.

AI Mode, by contrast, requires an active choice. A user has to tap the AI Mode tab or otherwise deliberately switch into the conversational interface; it does not auto-trigger based on the content of a typed query the way AI Overviews does. That single difference in activation model has real downstream consequences. Because AI Mode requires deliberate entry, the people using it are, on average, further into a research or decision-making process than the broader population hitting AI Overviews passively on ordinary searches. Industry analysis has consistently framed AI Mode as the surface where higher-intent, higher-consideration research happens — someone comparing five enterprise software platforms before recommending one to their employer, for instance — while AI Overviews handles the much larger volume of simpler, single-shot informational and transactional queries that make up the bulk of everyday search traffic.

The two products differ in reasoning depth as well as in triggering. AI Overviews is designed to answer with a single retrieval pass and a compact summary suitable for sitting above a results page without dominating it visually. AI Mode is designed for exactly the multi-step fan-out-and-synthesis pipeline described earlier in this analysis, capable of running many parallel sub-queries, following up with additional retrieval rounds, and sustaining a multi-turn conversation where later questions carry context from earlier ones in the same session. A query typed into AI Overviews gets one shot; the same question typed into AI Mode can be followed by “what about the cheaper option” or “does that apply in Europe too,” and the system carries the prior exchange forward rather than starting from zero.

Both features draw on the same underlying model family and much of the same retrieval infrastructure, including query fan-out itself — Google’s own documentation states plainly that both AI Overviews and AI Mode may use the fan-out technique to develop a response. This shared foundation is precisely why the two are so often conflated in casual usage, and why the confusion is understandable rather than simply a failure of public communication. The difference is one of deployment context and depth, not of two unrelated technologies wearing similar branding.

For anyone trying to optimise content visibility, the practical implication is that ranking well for AI Overviews inclusion and being cited inside AI Mode responses are related but distinct goals, and success at one does not guarantee success at the other. AI Overviews inclusion depends heavily on snippet-level factors — clear, extractable, concisely stated answers close to the top of a page, the kind of content that has always performed well in Google’s featured snippets. AI Mode citation depends more on how well a page answers the specific fan-out sub-queries the system generates around a topic, which, as covered earlier, are frequently different strings from the head-term query a page might already rank for. Marketing and SEO teams that treat the two as interchangeable targets, optimising only for one, are likely to see inconsistent results across the two surfaces — visible in AI Overviews on a given topic but absent from an AI Mode answer to a more elaborate version of the same question, or the reverse.

AI Mode versus AI Overviews at a glance

The table captures the split this section has argued for: AI Overviews is built for speed and low friction on ordinary queries, while AI Mode is built for depth on the smaller share of questions that genuinely need it, and the two increasingly work as connected stops in one continuous search session rather than as rival products competing for the same query.

A final structural difference worth noting is how the two interact within a single user session. Google has described a “seamless AI Search experience,” live globally by mid-2026, in which a user can flow from a query, to a results page carrying an AI Overview, to a follow-up question typed directly into AI Mode, all inside one continuous journey with links connecting the surfaces. That framing — AI Overviews as the lightweight entry point and AI Mode as the deeper destination a user graduates into — is closer to how Google itself now describes the relationship between the two products than treating them as competitors for the same query.

Multimodal input, Lens, and visual fan-out

AI Mode does not require typed text to begin a session. A person can open the camera, point it at an object, a plant, a piece of furniture, or an outfit, and start a research process from that image alone — a capability built on Google Lens and Circle to Search, both integrated directly into the AI Mode pipeline rather than existing as separate, disconnected tools.

The core capability is that AI Mode can decompose a single image into multiple distinct objects and run fan-out retrieval on each one simultaneously, rather than the older visual-search behaviour of identifying one dominant object per image. Google described this shift in detail through its own product blog: a major update to Circle to Search and Lens, rolled out in stages through late 2025 and early 2026, now allows the system to break down and search for multiple objects within a single photo at once. Using Circle to Search on an Android device around an entire outfit in a photo now returns results for every visible component of the look — the jacket, the shoes, the bag — rather than forcing a person into a slow, one-item-at-a-time visual search process, which was the state of the art before this update.

A concrete example Google has used to illustrate the feature involves a photo of several houseplants. Historically, identifying care requirements for each plant in a multi-plant photo meant asking about them one at a time, in sequence. With the updated pipeline, AI Mode identifies every plant in the frame, runs the necessary fan-out searches for each one’s specific care requirements using web results, and returns a single organised breakdown covering all of them together, often including suggested next steps such as repotting timing or seasonal watering adjustments. The underlying mechanism is identical to the text-based query fan-out described earlier — decomposition into sub-queries, parallel retrieval, synthesis into one answer — applied to visually identified entities instead of linguistically parsed sub-topics.

Visual search inside AI Mode is not limited to starting with an uploaded image; it can also branch mid-conversation from a result the system itself displayed. A person can begin with an entirely text-based query such as “visual inspiration for work outfits,” receive a set of image results, and then say something like “show me more options like the second skirt.” The system takes that specific image from its own prior response and begins a new fan-out process anchored to it, without the user needing to re-upload anything or restart the conversation. This blurs the line between text search and visual search in a way that a conventional search engine, built around a single text query box, never supported.

The multimodal capability draws directly from Gemini’s broader architecture rather than existing as a Search-specific vision system. Gemini’s image understanding — the same underlying capability Google uses in the standalone Gemini app, in Gemini in Chrome, and in various Workspace integrations — is what allows AI Mode to interpret an uploaded photograph, recognise the individual objects within it, and pass structured descriptions of each one into the fan-out retrieval pipeline described earlier in this analysis. This is a meaningful architectural point: Google is not maintaining a separate computer-vision stack purely for search; it is routing visual understanding through the same Gemini model family that handles text reasoning elsewhere in AI Mode, which is part of why new visual capabilities tend to appear across Google’s AI products in a similar window rather than showing up in Search months ahead of or behind the Gemini app.

The commercial motivation behind this investment is not subtle. Google’s own framing of the visual-search update leans heavily on shopping use cases — outfit components, product recommendations, price comparison across visually similar items — and the feature update arrived alongside, not separately from, the broader push toward monetising AI Mode through shopping ads and Direct Offers, covered in detail in a later section. Visual fan-out is, in practical terms, as much a commerce discovery mechanism as it is a convenience feature, and the two goals were clearly designed together rather than the shopping application being an afterthought bolted onto a purely descriptive visual-search tool.

Deep Search and the shift toward expert-level reports

Beneath AI Mode’s ordinary conversational responses sits a heavier mode, reserved for questions that genuinely require extensive research rather than a quick synthesis of a handful of sources. Google calls this Deep Search, and it represents the most extreme version of the fan-out-and-synthesis pipeline described earlier — the same mechanism, pushed to its upper limit of scale and patience.

Deep Search collates information from what Google’s own press materials describe as hundreds of websites, explicitly designed to connect and draw conclusions from information scattered across different, unrelated sources and contexts, producing what Google calls an “expert-level, fully-cited report” rather than a short conversational answer. Where an ordinary AI Mode response might run a dozen or so fan-out sub-queries and return an answer within a few seconds, Deep Search can issue dozens or even hundreds of background queries and take several minutes to complete — a deliberate trade of speed for depth that only makes sense for a specific category of question.

Google’s Robby Stein illustrated the feature with a personal example: researching home safes before a purchase. He described the task as involving unfamiliar variables that most buyers would not know how to weigh against each other unprompted — fire-resistance ratings, insurance implications tied to specific certification levels, and how different safe categories map onto different risk scenarios. Rather than running a handful of searches manually and cross-referencing them himself, he used Deep Search and reported that “it spent, I don’t know, like a few minutes looking up information and it gave me this incredible response. Here are how the ratings would work and here are specific safes you can consider.” That anecdote captures the intended use case precisely: questions where the value lies not in retrieving a single fact but in synthesising several types of specialised, cross-referenced information that a typical searcher would not otherwise know to look for separately.

The category of query best suited to Deep Search is defined by need for cross-domain synthesis rather than by topic. A comparative purchase decision involving technical specifications, safety certifications, and financial implications together is a natural fit. So is a research task spanning several distinct fields — medical information alongside insurance policy details, for instance, or a business decision that touches regulatory requirements, competitive positioning, and cost structure simultaneously. What unifies these cases is that no single source is likely to contain the full answer; the value Deep Search adds is specifically in connecting disparate pieces that would otherwise require a person to search several times, in several different specialist contexts, and hold the results in their head long enough to compare them.

The reporting format itself is structured differently from a standard AI Mode answer. Rather than a single flowing paragraph-style response, Deep Search output is organised more like a document — headed sections, comparative structure, and a citation trail dense enough to support the kind of scrutiny a genuinely consequential decision warrants. This format choice mirrors, deliberately or not, the structure long favoured in institutional research and consulting output, where the reader needs to be able to trace a specific claim back to a specific source without re-reading the entire piece.

Because Deep Search is by construction the most computationally expensive mode inside AI Mode, it has so far been used selectively — triggered by the system’s own judgement about query complexity rather than something users invoke as a labelled button in most markets, though Google has also framed it as a capability tied to its more advanced usage tiers. This selective triggering follows the same economic logic described earlier regarding model routing: Google is not willing to run hundreds of background queries against every ambiguous question, so the system reserves the heaviest processing for the cases where the marginal value of extra research time is highest, and the ordinary fan-out-and-synthesise pipeline handles everything else.

Deep Search is also the clearest evidence inside AI Mode’s public description of Google’s longer-term ambition for the product. A tool that produces an “expert-level fully-cited report” from a personal question about a household purchase is explicitly positioning conversational search as a substitute for tasks that previously required either extensive personal research effort or, in more consequential domains, a paid consultation with a specialist. Whether that substitution is reliable — whether a fully-cited report is also a fully accurate one — is a separate question addressed later in the section on where AI Mode gets things wrong, but the ambition itself is stated plainly in how Google talks about the feature.

Personal Intelligence and the Gmail and Photos connection

The feature that has drawn the most sustained privacy scrutiny of anything inside AI Mode is Personal Intelligence, which lets the system draw on a user’s own Gmail inbox and Google Photos library as context for its answers, rather than relying solely on publicly indexed web content and the immediate conversation.

Personal Intelligence launched on 22 January 2026, initially as a Labs feature restricted to Google AI Pro and AI Ultra subscribers with personal Google accounts in the United States, in English only. Robby Stein, Vice President of Product for Google Search, explained the intent in Google’s own announcement: “Personal Intelligence transforms Search into an experience that feels uniquely yours by connecting the dots across your Google apps. Starting today, Google AI Pro and AI Ultra subscribers can opt-in to securely connect Gmail and Google Photos to AI Mode. With this new experience, you can tap into your own personal context and insights to unlock even more helpful Search responses that are tailored to you.” The feature had been teased almost eight months earlier, at Google I/O in May 2025, as a category of “personal context” capability, and Google SVP Nick Fox confirmed in a December 2025 interview that it remained in internal testing with no public launch timeline at that point — the January 2026 rollout delivered what had effectively been delayed for the better part of a year.

The mechanism works by having AI Mode check connected apps for relevant information before generating a response, then citing exactly where that personal information contributed. Google’s own support documentation walks through a representative example: asking “what should I pack for my trip next week?” causes AI Mode to read a flight confirmation email in Gmail, identify the destination and travel dates from it, and cross-reference past trip photographs in Google Photos to produce packing suggestions built around that specific trip rather than generic advice. The interface signals this process transparently as it happens — showing “Checking Connected Apps” and “Applying Personal Intelligence” as visible steps — specifically so users can see when their personal data is being drawn into an answer rather than having it happen invisibly in the background.

Inline citations distinguish personal-context sources from ordinary web sources visually, and every personalised response carries a “Try without personalization” option that regenerates the answer using only public information, letting a user directly compare the personalised and non-personalised versions side by side — a built-in check Google added specifically so people could evaluate whether the personal-context layer had actually improved the answer or introduced an unwanted assumption.

The feature explicitly extends beyond travel planning into shopping and lifestyle recommendations. Google’s own example describes a family planning a weekend getaway: with Personal Intelligence enabled, AI Mode can reference a hotel booking confirmation found in Gmail and travel-related photos in Google Photos to suggest an itinerary built around the specific family — recommending, in Google’s example, an interactive children’s museum and an old-fashioned ice cream parlour, the latter suggestion drawn from noticing a pattern of ice-cream-themed photographs already saved in the user’s library. Robby Stein summarised the intended value proposition directly: “With Personal Intelligence, recommendations don’t just match your interests — they fit seamlessly into your life. You don’t have to constantly explain your preferences or existing plans, it selects recommendations just for you, right from the start.”

Expansion from the initial paid, English-only, US-only launch to a free, global feature happened remarkably fast. By around June 2026, roughly five months after the January launch, industry reporting described Personal Intelligence as available for free, in nearly 200 countries, in 98 languages, and integrated not just into AI Mode but also into the standalone Gemini app and Gemini in Chrome — described by one technology outlet as, in under five months, moving “from a U.S. paid-tier experiment to a free global feature.” That is an unusually rapid trajectory for a feature this sensitive, given how much scrutiny privacy-adjacent features typically draw before wide release, and it stands in sharp contrast to the far more cautious, staged rollout Google applied to information agents and agentic booking, both discussed in later sections, which remained restricted to the highest paid tier for considerably longer after their own launches.

Access to Workspace, business, enterprise, and education accounts was explicitly excluded from the initial rollout and, as of the feature’s early expansion, remained excluded — a deliberate boundary Google drew given the different data-governance obligations that apply to organisational Google accounts compared with personal consumer accounts, and one that meaningfully limits where Personal Intelligence can be relied upon in a professional context even as consumer access expanded quickly.

What Personal Intelligence stores, and what it says it does not

Connecting an AI system to an email inbox and a photo library raises an obvious question that Google has had to answer directly and repeatedly: does the model train on that content, and where does the personal information go once a query is finished with it?

Google’s stated position is that AI Mode does not train directly on a user’s Gmail inbox or Google Photos library. The company has said training is limited to specific prompts made inside AI Mode and the model’s own generated responses to those prompts, used to improve functionality over time — a distinction between using connected personal data as retrieval context for a single answer, which Google says happens, and using the raw content of that personal data as training material to improve the underlying model for other users, which Google says does not happen. This distinction between grounding and training is a meaningful one technically, though it depends entirely on trusting Google’s own characterisation of internal data flows, since the process is not independently auditable by outside researchers or regulators in real time.

The system is explicit and self-critical about its own error rate. Google’s documentation and public statements acknowledge plainly that Personal Intelligence “may not always personalize perfectly and can make mistakes,” including the specific failure mode of incorrectly connecting unrelated topics or misunderstanding context — for instance, surfacing a photo or email that happens to share a keyword with the query but has no genuine relevance to what the user actually meant. Google’s suggested remedy for this is behavioural rather than architectural: users are told to correct errors through a follow-up response inside the conversation, or to flag the issue with the thumbs-down feedback control, rather than any promise that the underlying matching logic itself would be fixed for that specific case going forward.

Connection and disconnection are framed as fully reversible and entirely user-controlled, at any time, not a one-way commitment. Google’s support documentation describes the ability to connect or disconnect Google content apps from Search services at will, alongside a specific memory-management interface that lets a user directly instruct the system in natural language — telling AI Mode “forget [specific detail]” to remove a piece of retained personal context, or providing an updated detail such as “I don’t run marathons anymore” to correct outdated context the system had previously inferred and retained. This natural-language memory control is a notable design choice: rather than requiring users to navigate a settings menu to manage what the system remembers about them, Google built the correction mechanism directly into the conversational interface itself.

Age-gating is explicit: Personal Intelligence is restricted to users eighteen or older, a restriction Google’s own support page states directly, consistent with the broader pattern of Google applying stricter default protections and eligibility gates around minors across its AI products, and consistent with the age restriction separately applied to the earlier AI Overviews rollout in the European Union.

None of this eliminates the underlying tension that privacy researchers and journalists raised as soon as the feature launched. A publication focused on security specifically framed the January 2026 announcement as tying “AI Search to Gmail and Photos, raising new privacy questions” — even a well-designed opt-in system with transparent citation and an easy undo mechanism still represents a large expansion of what a search engine can see about a person, compared with the purely public, anonymous query-and-response model that defined search for its first quarter century. The genuinely new element is not that Google collects data — Gmail and Photos data has always existed inside Google’s ecosystem, subject to Google’s existing privacy policies — but that AI Mode now actively reasons over that data in real time to shape the content of a specific answer, and does so with enough fluency that the connection can feel less like a database lookup and more like being personally known, an experience some users will welcome and others will find uncomfortable regardless of how clearly it is disclosed.

One further practical point deserves emphasis: opting in is a single decision covering both Gmail and Photos together in the initial design, though Google’s later documentation describes connecting them as separate, individually toggleable choices rather than an all-or-nothing bundle — a sensible-middle-path approach that lets a cautious user connect Photos alone, evaluate whether the personalisation feels worthwhile, and decide separately and later whether to also connect Gmail, rather than being forced into an immediate all-or-nothing decision the first time the feature is offered.

Agentic booking, Project Mariner, and the technology that replaced it

AI Mode’s ambitions extend beyond answering questions into completing tasks — booking tickets, filling in forms, and carrying out multi-step transactions on a user’s behalf. That capability has a specific, somewhat unusual origin story inside Google, tracing back through a standalone experimental product that Google ultimately shut down and folded back into the products it was originally meant to support.

Project Mariner began in December 2024 as an experimental, browser-based AI agent from Google DeepMind, designed to navigate websites autonomously — clicking buttons, filling forms, and completing multi-step tasks the way a human would, by taking screenshots, visually recognising interface elements, and translating that recognition into browser actions. Unlike an agent working through a clean, structured API, Mariner operated the way a person does: it saw the page, interpreted it visually, and acted on it, which meant it could work with essentially any website regardless of whether that site had built any special integration for automated agents. Google positioned it early on as a glimpse of “agentic checkout” — search evolving from finding information toward completing goals — and struck partnerships with platforms including Ticketmaster, Resy, and StubHub specifically to enable agent-driven commerce flows through it.

Google brought Mariner’s capabilities directly into AI Mode at I/O 2025, describing the integration as bringing “the agentic capabilities of Project Mariner into AI Mode, to help save you time with tasks like purchasing tickets.” The illustrative example Google used was concrete: a user could type “Find 2 affordable tickets for this Saturday’s Reds game in the lower level,” and AI Mode would trigger a query fan-out that looked across ticketing sites, analysing hundreds of potential ticket options against real-time pricing and inventory data, then handle the tedious work of filling in purchase forms — while still presenting the matching options to the user and leaving the final purchase decision and payment step in the user’s control on whichever site they preferred. This was explicitly framed as saving research and form-filling time rather than fully removing the human from the transaction.

Despite the early promise, Project Mariner as a standalone product had a short life: Google shut it down as of 4 May 2026, after roughly seventeen months of operation, folding its underlying technology into other Google products rather than discontinuing the capability itself. The Mariner landing page, after shutdown, carried a message thanking users and stating that its technology had “voyaged to other Google products.” Reporting from Wired in the preceding months had already flagged warning signs, describing Google reassigning staff off the dedicated Mariner team as the company responded to a shifting competitive field.

The stated reasons for the shutdown were architectural and economic rather than a judgement that the underlying idea had failed. Every action Mariner took required capturing a screenshot of the browser, sending it to cloud infrastructure, running it through a large multimodal model, interpreting the model’s response, and translating that back into a specific browser command — a round-trip that added real, noticeable latency to every single step of a multi-step task. Multiplied across millions of concurrent sessions, that compute overhead became difficult to justify economically, particularly as the broader industry moved decisively toward agentic tools that operate closer to the level of code and structured actions rather than visually interpreting a rendered page pixel by pixel, the way coding-focused agent products from other companies had begun to demonstrate as cheaper to run for many of the same categories of task.

Mariner’s core capabilities did not disappear; they were absorbed into Gemini Agent, into Google’s “Agent Mode” concept, into a new Chrome feature called Auto Browse, and, centrally for this analysis, into AI Mode itself. Auto Browse, which rolled out in Chrome in early 2026, lets the browser navigate complex, multi-step web flows without direct human input at each step, and reporting on the feature has noted the architectural resemblance to Mariner is difficult to miss even though Google has not explicitly drawn a direct line connecting the two products in its own communications. As of mid-2026, Auto Browse remained available only to AI Pro and AI Ultra subscribers in the United States, with no announced timeline for a UK or EU rollout — a gap reporting has connected specifically to unresolved questions about how browser agents that auto-fill forms or handle stored authentication credentials would need to operate under GDPR’s stricter consent and data-processing requirements.

The broader industry context helps explain both Mariner’s rise and its fall. OpenAI shipped its own agentic browsing product, ChatGPT Atlas, in October 2025, and Perplexity had launched its Comet browser even earlier, in July 2025, both offering broadly comparable agentic browsing to users well beyond the narrow group Google had initially given access to Mariner. Google’s decision to fold Mariner’s technology into existing, already-massive-distribution products — Search, Chrome, Gemini — rather than continue investing in it as a separate standalone brand reflects a strategic choice toward platform consolidation over maintaining a distinct agent product competing head-on for attention against rivals with larger existing user bases in that specific category.

Information agents and the era of Search agents

At Google I/O 2026, Google introduced a new category of AI Mode capability that shifted the product’s basic behaviour from purely reactive — answering a question when asked — toward proactive, continuous monitoring on a user’s behalf, without the user needing to return and ask again.

Google calls this category “Search agents,” and the first member of the category to ship is called information agents — persistent background processes that monitor the web around the clock and surface relevant findings through push notifications, without the user needing to initiate a new search each time. Elizabeth Reid, Google’s Vice President and head of Search, introduced the concept on stage as the start of “the era of Search agents,” a framing deliberately positioned as a successor to Google Alerts, the far simpler notification tool the company has offered since 2003, rebuilt now with the reasoning capacity of a frontier language model applied to deciding what is genuinely worth surfacing rather than matching on keywords alone.

The mechanism works by having a user describe an ongoing interest or need in natural language, phrased with trigger language such as “keep me updated on” or “alert me when,” which the system recognises and converts into a standing background task rather than a one-off query. Google’s own illustrative example asked the system to “keep me updated when any of my favorite athletes announce sneaker collabs or signature drops” — a request with no natural single answer, since the relevant information does not exist yet at the moment the request is made, and satisfying it requires ongoing monitoring rather than a single retrieval pass. The agent then reasons across blogs, news sites, social media posts, and Google’s own real-time structured data covering areas such as finance, shopping, and sports, watching continuously for changes relevant to the specific question it was given, and notifying the user through the surface it operates on when something matches.

A second illustrative example Google used, apartment hunting, demonstrates the intended breadth of application beyond entertainment trivia. A user can describe detailed housing requirements in one extended, unstructured message — Google described this explicitly as a “brain dump” of exact requirements — and the resulting agent continuously scans new listings against that full requirement set, notifying the user specifically when a listing that satisfies the stated criteria appears, sparing the user from manually repeating the same search on the same listing sites day after day.

Access rolled out deliberately narrow at first: information agents launched on 12 June 2026 exclusively for Google AI Ultra subscribers, the most expensive consumer subscription tier Google offers, priced at either $99.99 or $199.99 per month depending on the specific plan selected. Google stated explicitly that Search agents would extend to the less expensive AI Pro tier later the same summer, following the familiar staged-rollout pattern the company has applied to essentially every advanced AI Mode capability discussed in this analysis — narrow paid-tier launch first, wider paid-tier access second, free-tier access, if it comes at all, considerably later and not guaranteed.

Distinct from information agents but announced in the same wave, Google also described expanding agentic booking capabilities within Search to a broader range of tasks beyond the ticket-purchasing example described in the previous section, extending specifically into local experiences and services — a direct continuation and broadening of the Mariner-derived agentic transaction capability, now generalised well past its original ticketing and travel-booking use case.

The competitive framing behind this entire announcement wave is impossible to separate from the pressure Google is under from rival AI products. The same reporting that covered the information agents launch noted, for context, that OpenAI had reported ChatGPT reaching comparable or larger usage scale during the same period, and Google’s own executives have spoken candidly about a period of internal alarm within the company during late 2025 as ChatGPT’s search-adjacent features gained ground — reporting that specifically described Google’s “code red” response to OpenAI’s rise becoming, by the same account, “OpenAI’s code red” in turn once Google’s own AI Mode and Gemini investments began showing comparable momentum by early 2026. Search agents, in that context, read less as an isolated feature launch and more as one further move in an accelerating cycle of competitive escalation between the major AI search and assistant products, each pushing further into proactive, agentic territory in response to what the others are shipping.

The underlying technical shift this category represents matters regardless of the competitive framing: it moves AI Mode from a strictly session-bound tool, where the system has no existence or memory outside an active conversation, toward something closer to a standing service that persists and acts between sessions. That shift raises the same category of questions already discussed regarding Personal Intelligence — what is retained, for how long, and under what user control — but applied now to continuous, autonomous monitoring rather than a single query answered with the aid of connected personal context, a distinction that matters because a system continuously watching multiple data sources on a user’s behalf, even for a narrowly scoped task such as sneaker-drop alerts, is a meaningfully different privacy and control surface than one that only activates when explicitly asked a question.

The redesigned search box and what it changes about asking

Alongside the deeper reasoning and agentic capabilities covered so far, Google made a much more visible change at I/O 2026: it redesigned the search box itself, the single interface element that has defined how people interact with the product since its inception.

Elizabeth Reid described the update as “the biggest upgrade to our iconic search box since its debut over 25 years ago,” a description Google’s own communications repeated closely enough across multiple channels to signal it was a deliberately chosen framing rather than an offhand comparison. The redesigned box dynamically expands to accommodate longer, more conversational input, moving away from the historical assumption that a search query should be a short string of keywords and toward accepting the kind of extended, natural-language description a person might use when explaining a problem to another person rather than typing a query into a database.

The new box also incorporates an AI-powered suggestion system that Google describes as going beyond traditional autocomplete, attempting to actively anticipate a user’s underlying intent and help them formulate a more precise question rather than merely predicting the next few characters of what they have already started typing. This distinction matters conceptually: traditional autocomplete is a statistical prediction of likely completions based on what other users have typed; the redesigned system is described as trying to understand what the user is actually trying to accomplish and proposing a better-formed version of the question itself, closer to a research assistant helping someone sharpen a vague question into a specific, answerable one before the search even runs.

Multimodal input is built into the box directly rather than requiring a separate interface: it accepts text, images, files, videos, and even open Chrome tabs as input, in any combination within a single query. The inclusion of Chrome tabs as an input type is a particularly direct extension of Gemini in Chrome’s browsing-context capabilities into the core Search interface, letting a user reference something they are currently looking at in another tab without needing to describe it in words or take a screenshot manually first — the system can read the tab’s content directly as context for the query being composed.

Taken together with the other I/O 2026 announcements — information agents, Gemini 3.5 Flash as the new default model, expanded agentic booking — the redesigned search box completes a coherent narrative Google was clearly building deliberately for that keynote: that the traditional format of search, a short keyword string returning ten ranked links, is now explicitly treated internally as a transitional format rather than the permanent default. That said, Google’s own Vice President of Search, Liz Reid, was quoted directly pushing back against the strongest possible reading of that narrative in the immediate aftermath of the keynote, telling TechRadar plainly: “This new search box does not mean that you’ll only get AI responses.” That single sentence, from the executive most directly responsible for the product, closes the door on any reading of the announcement as an imminent, wholesale replacement of classic search results with AI-only answers, and instead frames the redesign as an expansion of what the box can accept and how deeply it can reason, layered on top of — rather than as an outright replacement for — the traditional results format that a large share of queries, particularly navigational and simple factual ones, continue to receive by default.

No public Google document between the original AI Mode launch in May 2025 and the I/O 2026 keynote a year later set any fixed date for retiring the classic search-results page, and the pattern of announcements across that entire period has consistently been lateral rather than substitutive — more countries, more languages, more input types, deeper integration with other Google products — rather than a single frontal replacement of the ten-blue-links format with an AI-only interface. Readers and businesses trying to plan around when, or whether, classic search disappears entirely should treat that question as genuinely unresolved rather than settled in either direction, based on everything Google has stated publicly through the period this analysis covers.

Advertising inside AI Mode and how Google monetizes conversation

For its first several months of existence, AI Mode carried no advertising at all, and Google made a point of saying so. That changed through 2026, in a sequence of announcements that revealed both how quickly Google intended to monetise the surface and how much backend advertising infrastructure had reportedly already been built in before any ads became visible to ordinary users.

AI Mode launched, and remained for roughly its first year, explicitly ad-free — Google described this as a deliberate choice to give users “a clean and quiet search space” that would help build trust in the product’s answers while the underlying model and retrieval quality were still being proven out, with the clear caveat that Google might add advertising later. That caveat became operative through 2026. Google Marketing Live in May 2026 introduced a set of new, conversationally-native ad formats explicitly designed for the AI Mode and AI Overviews context rather than simply porting over the display-ad or text-ad formats that had defined search advertising for two decades. Google’s own description of these formats emphasises that Gemini itself is now involved in constructing the ad content: “Our Gemini model evaluates and synthesizes information about a product or service, and displays that context alongside the advertiser’s creative,” with Google stating this approach is intended to preserve transparency and trust, and that the resulting formats “will also continue to be clearly labeled as ‘Sponsored.'”

Two named formats illustrate the design: Conversational Discovery ads and Highlighted Answers. Conversational Discovery ads are built to answer a person’s specific stated question directly, using Gemini to generate creative matched to the particular phrasing of that question rather than serving a generic, pre-written ad matched only on broad keyword overlap — Google’s own example imagines someone asking how to make their home smell like a spa or a rainy forest, with the ad format using Gemini to build creative highlighting features specifically relevant to that phrasing. Highlighted Answers are aimed at the research phase of a purchase decision, offering helpful, guidance-style suggestions from advertisers during exploratory queries rather than only at the point of an explicit transactional intent.

Independent technical analysis suggested the advertising infrastructure had been built and tested well before any of it became visible to users. A research group called Discovered Labs published a network-traffic analysis in January 2026, ahead of the May Marketing Live announcements, based on capturing 547 Google network flows and more than 1,300 total requests during live AI Mode sessions. Their central finding was that AI Mode was already running complete ad delivery, tracking, and attribution systems in the background months before any ad actually rendered visibly in a response — including a specific ad placement identified in the code as “AI Mode Bottom Ads” with the internal shorthand aimba, infrastructure the researchers said was ready to display ads below AI-generated answers, plus query-to-conversion attribution already tracking user journeys through AI Mode using a parameter the researchers identified as adview_query_id, and shopping and product-listing ad code present within the product’s JavaScript bundles well ahead of any public shopping-ad announcement.

The first ads to become publicly visible arrived even earlier than the May Marketing Live formats, and in AI Overviews rather than AI Mode specifically. Ad-intelligence firm Adthena reported detecting ads appearing inside AI Overviews at a frequency of 0.052 percent across a sample of 25,000 search-results pages analysed in November 2025 — a vanishingly small but real and independently measured early signal of the broader monetisation Google would expand through the following year.

Once AI Mode itself began carrying visible ads in 2026, the reaction from at least some users and industry observers was sharply critical, and the criticism centred specifically on placement inside the flow of the AI-generated answer itself rather than beside or below it. Entrepreneur Juozas Kaziukenas described the experience directly in a widely shared LinkedIn post after asking AI Mode for coffee-maker recommendations: “I asked for coffeemaker recommendations, and it inserted a sponsored listing in the middle of organic results,” framing the practice as breaking what he characterised as AI’s biggest implicit promise to users — that a synthesised answer would represent the system’s best independent judgement rather than a judgement partially shaped by which brands had paid for placement. He drew an explicit parallel to Anthropic’s Super Bowl advertising earlier in 2026, which had mocked the idea of ChatGPT introducing advertising into its own responses, arguing that Google had effectively done the same thing it had watched a competitor get mocked for. His summary framing was blunt: “AI Mode is where Google is first monetizing its AI.”

The scale context Google itself provided around the same period makes clear why this monetisation push mattered commercially, independent of the specific user reaction to any one ad placement. At the same I/O 2026 keynote where the ad formats were previewed, Google disclosed that AI Mode had surpassed one billion monthly active users, with query volumes that had more than doubled every quarter since the original March 2025 launch, and that total search volumes across Google’s properties had hit an all-time high in the first quarter of 2026 as a direct result of these AI integrations driving more overall query activity rather than cannibalising it. A surface generating that much engagement, with essentially no advertising revenue attached to it for its first year, represented a substantial and growing gap in Google’s core revenue model for exactly as long as it remained ad-free — a gap the company had strong commercial incentive to close as quickly as user tolerance and advertiser readiness allowed.

Shopping ads, Direct Offers, and the commerce layer

Beyond the conversational ad formats described in the previous section, Google built a distinct and more transactionally direct commerce layer into AI Mode, aimed specifically at the point where research turns into a purchase decision.

Google introduced a dedicated shopping ad format built specifically for AI Mode on 11 February 2026, letting sponsored product listings from named retailers appear directly beneath AI-generated product recommendations. An early public example showed AI Mode displaying sponsored listings from Target and B&H underneath a set of AI-generated headphone recommendations — a placement that puts paid, retailer-specific listings in close visual proximity to, but visually distinguished from, the model’s own independently generated suggestions. Google grounded the design of this format in commissioned consumer research: a study run between October and December 2025 involving 4,773 participants found, according to Google’s characterisation, that AI Mode provides “a more helpful shopping experience when users can easily compare a variety of brands and stores” — research Google used to justify the specific design choice of surfacing multiple sponsored retailers alongside a single product recommendation rather than a single sponsored placement standing alone. Google’s framing insists the ads are “clearly labeled as sponsored” and fit “naturally into the conversation” rather than disrupting the research process a user is in the middle of.

Direct Offers is a related but distinct format that launched earlier, on 11 January 2026, letting participating retailers surface time-limited discounts directly inside an AI Mode response when the system judges an offer relevant to the specific query and context. Google’s own illustrative example describes a user asking AI Mode for outdoor-furniture recommendations with specific stated requirements; a participating retailer whose product matches those requirements can surface a time-sensitive discount on exactly that matching product, appearing inside the response as a labelled “Sponsored deal.” The mechanism depends on retailers pre-configuring relevant offers within their existing campaign settings, with Google’s AI system then determining in real time when a given user’s specific query and stated context make displaying that particular offer appropriate, rather than the retailer choosing the trigger conditions directly themselves. Early access to the Direct Offers pilot was restricted to retailers meeting specific criteria around product feed quality and offer-management practices, a gating decision consistent with Google’s general pattern of restricting new advanced features to a qualifying subset of participants before wider release.

The advertiser-facing infrastructure underlying both formats draws on campaign types and Google Ads structures that advertisers were already broadly familiar with, a deliberate design choice intended to lower the barrier to adopting the new placements. Search Engine Land’s analysis of the situation, published in March 2026 under the direct framing that AI Mode was becoming “the next ads engine,” noted that Google’s documentation on advertising eligibility inside AI Overviews and AI Mode pointed advertisers toward campaign types they already knew how to run, rather than requiring them to learn an entirely new bidding or targeting system from scratch specifically for the conversational surface — a pragmatic choice that traded some potential for AI Mode-specific ad innovation against faster advertiser uptake industry-wide.

Beyond the two named formats already covered, Google has signalled further monetisation directions it expects to test, based on the direction of its own public statements and current product trajectory rather than confirmed launches. These include sponsored product comparisons that would let advertisers promote specific products directly within AI-generated comparison tables, interactive ad experiences letting users ask follow-up questions directly to an advertiser inside the AI Mode conversation itself, video ad integration for query types where video content is contextually appropriate, and local-business promotions surfaced specifically within location-based AI Mode queries. None of these should be treated as confirmed, shipped features; they represent Google’s own stated direction of travel rather than products currently live, and readers evaluating AI Mode’s advertising trajectory should distinguish clearly between the shopping ads and Direct Offers formats that have actually launched and this broader set of plausible next steps that remain, as of this analysis, unconfirmed.

A separate program, launched alongside these shopping-specific formats, is a Universal Commerce Protocol that Google introduced on 11 January 2026, intended to standardise how AI systems — Google’s own and, by design, potentially others — interact with merchant systems for tasks spanning product discovery through to checkout. This protocol-level investment matters because it signals Google’s ambition extends past simply placing ads inside AI Mode responses and toward building the underlying transactional plumbing that would let an AI system complete a purchase directly, a capability that connects the advertising layer covered in this section back to the agentic booking capabilities discussed earlier in this analysis — the commerce and the agentic-action stories inside AI Mode are, in Google’s own architecture, two expressions of the same underlying infrastructure rather than separate product lines developing independently of each other.

The Shopping Graph and real-time retail data

Underneath every shopping-related fan-out query, product recommendation, and sponsored listing described in the previous two sections sits a single specialised data source that Google has invested in for years and increasingly treats as central infrastructure for AI Mode’s commerce capabilities: the Shopping Graph.

The Shopping Graph is Google’s structured, continuously updated database of product listings, pricing, availability, and retailer information, distinct from the general web index that AI Mode’s fan-out queries otherwise draw on. Robby Stein disclosed a specific figure for its update frequency in mid-2025 remarks reported by Search Engine Journal: the Shopping Graph updates roughly two billion times an hour, a scale of continuous refresh that reflects how volatile retail data genuinely is — prices change, items go out of stock, promotions start and end, and a static, periodically re-crawled index of product pages would be systematically stale for a meaningful share of queries by the time a user actually searches.

Query fan-out draws on the Shopping Graph as one of its named specialised data sources, alongside the live web and the Knowledge Graph, specifically for queries where structured retail data answers the question better than a general web page would. A comparative shopping question about headphones, the example used earlier in this analysis to illustrate fan-out mechanics generally, depends on exactly this kind of structured data — current price, current stock status, and specific technical specifications pulled from a standardised schema rather than extracted through general-purpose text parsing of an arbitrary retailer’s product page, which would be slower and considerably less reliable at scale.

The Shopping Graph’s role expanded materially with the introduction of the agentic ticket-purchasing example discussed in the section on Mariner’s integration into AI Mode. Analysing “hundreds of potential ticket options with real-time pricing and inventory” for a specific event, as Google’s own illustrative example described, is only possible because that pricing and inventory data exists in a structured, queryable, continuously updated form in the first place; an agent attempting the same task purely through visually parsing dozens of individual ticketing websites in real time, the way the original Project Mariner architecture worked before its capabilities were absorbed elsewhere, would be dramatically slower and less reliable for exactly the kind of time-sensitive, inventory-dependent task ticket purchasing represents.

The commercial incentive behind maintaining infrastructure at this update frequency is direct and substantial: retail search is one of the highest-intent, highest-conversion-value categories of query that exists, and the Shopping Graph is the mechanism that lets AI Mode compete credibly on exactly the category of search where losing ground to a dedicated commerce-focused competitor would be most costly to Google’s advertising revenue. This is the same underlying commercial logic already discussed regarding the shopping ads and Direct Offers formats — the Shopping Graph is the data layer those advertising products are built on top of, and its continuous, high-frequency refresh is what makes real-time, contextually relevant sponsored placements technically viable inside a conversational answer in the first place, rather than an ad system reliant on data that might be hours or days stale by the time a user actually sees it.

For anyone running a retail or e-commerce business trying to understand why their product does or does not surface inside an AI Mode shopping answer, the Shopping Graph is a more direct lever to pull than general web SEO practices. Ensuring accurate, complete, and properly formatted product feed data — the same structured feed data that already powers Google Shopping listings on the conventional search results page — is the mechanism by which a retailer’s inventory becomes eligible to appear inside AI Mode’s fan-out retrieval for shopping-related queries, a more technical and infrastructure-dependent requirement than the content-quality and citation-worthiness factors that matter for AI Mode visibility on informational queries generally, covered in more detail in the practical guidance sections later in this analysis.

How AI Mode decides which sources to cite

A response inside AI Mode is not simply generated prose; it is prose with citations attached, linking specific claims back to specific retrieved pages. Understanding how that citation process actually works — what determines which sources get referenced and which do not — matters both for evaluating the reliability of any given answer and for anyone trying to be visible inside the product professionally.

Citation attachment happens during the synthesis stage of the pipeline described earlier in this analysis, as the model writes its response using the passages retrieved across all the fan-out sub-queries run for that question. The model is not choosing citations after the fact from a fixed shortlist; it is grounding specific sentences in specific retrieved content as it generates them, which is why a single AI Mode response commonly cites several different domains, each supporting a different claim or a different facet of a compound question, rather than leaning on one dominant source the way a traditional search result page might funnel attention toward the single top-ranked link.

Because query fan-out generates the retrieval pool the model draws from, and fan-out sub-queries frequently differ from the literal text a user typed, citation eligibility depends on a page’s performance across an entire cluster of related, machine-generated queries rather than its ranking for any single head term. This is the practical mechanism behind a pattern search-engine-optimisation professionals have widely reported: a page that ranks respectably, but not spectacularly, for its primary target keyword can still earn a citation inside an AI Mode answer if it happens to answer one of the specific fan-out sub-questions the system generated particularly well, while a page that dominates the rankings for the head term itself can be entirely absent from the synthesised answer if it does not address any of the specific sub-facets the fan-out process actually searched for.

Independent research into citation patterns across AI search systems more broadly — not limited to AI Mode specifically — has found that citations concentrate heavily among a relatively small number of domains rather than spreading evenly across the web’s full diversity of publishers. A 2025 academic study of AI search citation patterns, drawing on a large dataset from an AI Search Arena benchmark, found that news sources represented only about 9 percent of citations across the systems studied, and that a substantial share of all citations — reported estimates ranging from roughly 40 to 55 percent depending on the specific platform — flowed to fewer than 1,000 domains total. While that specific study covered several AI search systems rather than AI Mode in isolation, the underlying dynamic it documents is directly relevant: AI-generated answers reward a narrower set of sources more heavily than the long, gradually declining tail of visibility that characterised traditional search results pages, where thousands of individually smaller sites could each capture a meaningful trickle of traffic for their own specific niche.

Structured, extractable content appears to perform disproportionately well in this citation process, consistent with the broader pattern already discussed regarding AI Overviews inclusion. Content organised with clear headings, direct and concisely stated answers near the top of a page, and explicit factual claims that do not require the model to infer information the page states only implicitly, is easier for a retrieval-and-synthesis pipeline to extract cleanly and cite with confidence, compared with content that buries its actual answer inside long narrative paragraphs, however well-written those paragraphs might be by conventional editorial standards.

Google itself has been notably reluctant to share detailed data on citation patterns specific to AI Mode, which several heads of SEO at major publishers told Digiday in 2025 was forcing the industry to rely on third-party tools such as AlsoAsked and Profound, alongside adjacent Google features like People Also Ask, to infer visibility patterns indirectly rather than through any first-party reporting Google provides. Mike King, founder of the SEO agency iPullRank, was quoted in that same reporting framing query fan-out’s practical effect on the profession bluntly: the discipline is shifting focus from subintents behind a search query toward understanding and addressing the full cluster of questions the fan-out process is likely to generate, a materially different and more speculative task than the more measurable practice of optimising for a known, fixed set of target keywords that defined search-engine optimisation for the two decades before AI Mode existed.

A separate observation reported across multiple industry sources concerns citation halo effects between organic and paid visibility. Discovered Labs’ analysis of AI Mode’s ad infrastructure suggested that a brand’s presence in organic AI Mode citations likely creates a favourable effect for that same brand’s paid placements — users may respond more receptively to a sponsored suggestion from a brand the system has already organically recommended earlier in the same session, a dynamic that, if borne out at scale, would tie a brand’s paid advertising performance inside AI Mode more tightly to its organic citation performance than has historically been true on the classic results page, where paid and organic listings have generally functioned as more separable, independently measurable channels.

Why AI Mode sometimes gets things wrong

No description of how AI Mode works is complete without a direct account of where and why it fails, because the same architecture that makes the product powerful — synthesising many retrieved sources into one confident-sounding answer — is also the architecture responsible for its most consequential failure modes.

The most basic failure mode is retrieval quality: if the fan-out process pulls in a thin, outdated, or low-quality set of sources for a given sub-query, the synthesis stage has no way to compensate for that gap, because it can only work with what retrieval actually returned. A well-tuned model writing confidently over a poor retrieval pool produces exactly the kind of fluent, plausible-sounding, but substantively wrong answer that has become one of the defining reputational risks associated with AI-generated search summaries generally, and AI Overviews specifically drew sustained public ridicule in 2024 for a small number of viral errors — recommending glue as a way to help cheese stick to pizza, and suggesting eating rocks for mineral content, both traced back to satirical Reddit posts that the retrieval process had pulled in and the synthesis stage had failed to recognise as unreliable or non-literal.

A second failure mode is misattribution and false synthesis: the model can smooth over a genuine disagreement between two retrieved sources as though no disagreement existed, or attribute a specific claim to a source that did not actually make it, particularly when several retrieved passages discuss closely related but not identical facts. This risk grows, rather than shrinks, as the number of parallel fan-out sub-queries increases, because a heavier retrieval pool gives the synthesis stage more raw material to potentially blend incorrectly, even as it also, in the more common case, gives the model more corroborating evidence to draw a more confident and better-supported answer from. The net effect on any individual response depends heavily on the specific query and the specific quality of what was retrieved for it, which makes error rates difficult to characterise with a single general statistic that would apply meaningfully across the full range of query types AI Mode handles.

A third and more subtle failure mode concerns compound questions where the model treats each sub-facet independently rather than recognising an interaction between them. A question genuinely requiring integrated judgement across several linked factors — the kind of question Deep Search is specifically designed to address, as covered earlier — can still produce an answer that technically addresses each individual facet correctly in isolation while missing an important interaction effect between them, a failure mode considerably harder to catch through simple fact-checking of individual claims than an outright factual error would be, precisely because each component claim can check out true on its own even as the overall synthesis misses something a genuine domain expert would have caught by weighing the factors against each other rather than addressing them in sequence.

Personal Intelligence introduces its own distinct category of error, already discussed in an earlier section, where the system incorrectly connects unrelated personal context to a query — surfacing a photo or email that happens to share superficial features with the query’s subject matter but carries no genuine relevance, a failure mode Google itself has openly acknowledged rather than denied, framing correction as something the user actively participates in through follow-up prompts and feedback rather than promising the underlying matching logic to be reliably error-free from the outset.

Google’s own public position on accuracy has been to acknowledge the problem directly rather than deny it, a notable shift in public communication strategy compared with the company’s initial, more defensive response to the 2024 AI Overviews errors. Reporting in 2026 characterised this shift as Google “quietly admitting” ongoing issues with AI-generated responses producing inaccurate information, fabricated sources, incorrect citations, and outright hallucinated content, attributing these persistent problems to a combination of large-language-model synthesis limitations, genuine retrieval challenges on ambiguous or rapidly changing topics, and the fundamental difficulty of covering the near-infinite range of query phrasing and context that real users bring to the product at the scale AI Mode now operates at.

The practical implication for anyone relying on AI Mode for a genuinely consequential decision is straightforward and consistent with how any single information source, however sophisticated, should be treated: a fully-cited AI Mode answer, even one produced through the heavier Deep Search pipeline, is evidence to check against independent sources rather than a substitute for that checking, particularly for decisions involving meaningful financial, medical, legal, or safety stakes. This is not a criticism unique to AI Mode among AI search products — the same fundamental limitation applies to every retrieval-augmented generation system currently deployed at scale, across every major AI company building comparable products — but it is a limitation that AI Mode’s confident, fluently written, heavily cited presentation style can make easy to underestimate in practice, precisely because the format signals a level of rigour that the underlying retrieval-and-synthesis process does not always actually deliver.

The zero-click debate and what the traffic data actually shows

Nowhere has AI Mode’s impact been more contentious, or more heavily measured by outside researchers, than its effect on whether searchers ever click through to an external website at all. The data on this question is genuinely extensive by 2026, but it is also inconsistent across studies in ways that are worth surfacing honestly rather than collapsing into a single tidy number.

SparkToro, a research firm that has tracked zero-click search behaviour for years, reported in June 2026 that the share of Google searches generating at least one click fell 9.51 percentage points between 2024 and 2026, a 22.9 percent relative decline, using Similarweb desktop and mobile web panel data covering US searches from January through April 2026. That figure covers clicks to any destination, including organic results, paid ads, and Google’s own properties such as Maps and YouTube, and it excludes follow-up searches conducted within Google itself — a category SparkToro found had grown by 7.2 percentage points over the same period, a pattern the firm’s co-founder Rand Fishkin interpreted as reflecting Google’s growing ability to answer questions directly within the results page while still encouraging users to refine or continue searching inside Google’s own ecosystem rather than leaving it.

Other studies conducted around the same period report figures that are directionally consistent but numerically quite different, a divergence that traces mainly to differences in methodology, geography, and exactly which query types each study’s sample covers. Search Engine Journal, drawing on Similarweb data, reported zero-click searches making up 69 percent of all queries as of late 2025. A separate analysis published in mid-2026 put the figure specifically for AI Mode sessions at 93 percent — describing AI Mode as effectively replacing organic results entirely rather than layering a summary above them, meaning that for queries actually routed through AI Mode specifically, as distinct from the broader Google search ecosystem including AI Overviews and ordinary results pages, organic SEO visibility outside of earning a direct citation inside the AI-generated conversation has essentially no reach at all.

Not every methodology agrees on the direction of the most recent trend, and this disagreement deserves to be stated plainly rather than smoothed over. Datos’s Q1 2026 State of Search report, using what its authors describe as a strict clickstream methodology, found the US zero-click rate actually falling slightly, from 24.5 percent in December 2025 to 22.4 percent in March 2026 — the opposite direction from SparkToro’s headline trend, though the two figures are not directly comparable given substantially different underlying methodologies, sample definitions, and what each study counts as a “click.” One technology outlet summarising the broader body of research suggested the Datos finding might indicate that zero-click pressure is levelling off as AI Overviews prevalence stabilises, supporting a broader thesis that the sharpest phase of 2025’s disruption to publisher traffic had already passed by early 2026 — a genuinely open empirical question rather than one with a single settled answer as of this analysis.

Publisher-side traffic disclosures, where companies have named specific figures rather than relying on third-party panel estimates, paint a consistently severe picture regardless of which broader zero-click methodology one prefers. The Reuters Institute and Chartbeat’s January 2026 report found Google search traffic to publishers had fallen 33 percent globally in the year to November 2025, with US publishers hit harder, at a 38 percent decline, alongside a 21 percent fall in Google Discover referrals across a sample of more than 2,500 publisher sites. The same report found ChatGPT referrals to publishers growing more than 200 percent over the same period — a dramatic relative growth rate that, in absolute terms, still accounted for less than 1 percent of total publisher referral traffic, making explicit that chatbot-driven traffic is nowhere close to offsetting the scale of what has been lost from declining traditional search referrals. Named company disclosures cited in the same body of reporting include HubSpot estimating a 70 to 80 percent organic traffic decline, Business Insider losing 55 percent of its traffic between April 2022 and April 2025 alongside a 21 percent staff reduction, CNN reporting a decline in the 27 to 38 percent range, and Chegg reporting a 24 percent year-over-year revenue fall that the company explicitly attributed to AI Overviews in a lawsuit it filed against Google in February 2025.

Ahrefs’ own monthly tracker, covering more than 75,000 domains that have opted in to have their traffic metrics aggregated and published, found an 8 percentage point decline in Google-referred traffic share between June 2025 and May 2026 — roughly a 22 percent relative drop — across a panel SparkToro specifically noted is far from a neutral, representative sample of the web, since it consists overwhelmingly of sites with professional marketing teams actively working to grow their traffic rather than passive or declining sites, meaning the true average impact across the entire web could plausibly be either better or worse than this specific panel shows, depending on which direction professional SEO effort systematically biases the sample.

The effect is not remotely uniform across query categories or industry verticals, and this unevenness is arguably more useful for any individual business than the aggregate national or global figures. Ahrefs data cited in a 2026 analysis found AI Overview trigger rates and associated traffic impact varying enormously by sector — health-related queries showing heavy AI Overview presence with correspondingly steep organic traffic losses in that vertical, while e-commerce and shopping queries showed only around 3.2 percent AI Overview trigger rates, down from an initial roughly 29 percent, a pullback the analysis attributed to Google’s own recognition that AI-generated summary responses were not converting into actual sales for shopping-intent queries as effectively as a traditional results page with visible product listings did. That specific data point — Google apparently throttling back AI Overview presence specifically where it measurably hurt commercial outcomes — is a useful corrective to any assumption that AI-generated summaries are expanding uniformly and irreversibly across every category of query; Google’s own behaviour suggests a willingness to pull back selectively where the format demonstrably underperforms the alternative it replaced.

Publisher lawsuits and the European Commission investigation

The traffic decline documented in the previous section has not gone unanswered. Through 2025 and into 2026, publishers, publisher trade associations, and eventually formal regulators moved from complaint to litigation to active antitrust investigation, with both AI Overviews and AI Mode named specifically as the products under scrutiny.

The earliest formal complaint came from the Independent Publisher Alliance, which filed with the European Commission in mid-2025, alleging that Google’s “core search engine service is misusing web content for Google’s AI Overviews in Google Search, which have caused, and continue to cause, significant harm to publishers, including news publishers in the form of traffic, readership and revenue loss.” The complaint specifically highlighted the absence of a genuine opt-out mechanism: publishers wishing to prevent their content from being used in AI-generated summaries had no way to do so without effectively de-indexing from Google Search altogether, an outcome no commercially viable publisher could accept given how dependent most publishing business models remain on Google search traffic even as that traffic has been declining. The complaint was backed by two additional organisations, the Movement for an Open Web, a coalition of digital advertisers and publishers, and Foxglove Legal Community Interest Company, a UK-based technology-accountability nonprofit, and it requested interim measures from the Commission specifically to prevent what the complainants characterised as an urgent, ongoing competitive harm rather than a matter that could wait for the Commission’s normal multi-year investigatory timeline.

The European Commission escalated matters considerably on 9 December 2025, announcing a formal antitrust investigation into whether Google had breached EU competition rules by using web publishers’ content, as well as content uploaded to YouTube, for artificial intelligence purposes without adequate compensation or a meaningful ability for publishers to decline participation. EU Competition Commissioner Teresa Ribera framed the investigation’s purpose directly, stating that innovation “cannot come at the expense of the principles at the heart of our societies,” and describing the case as a “strong signal” of the European Union’s commitment to protecting online press and other rights holders as generative AI products expand.

The European Publishers Council, whose membership includes major publishing groups such as DMG Media, The Guardian, News UK, and The New York Times, filed its own separate, more formally structured antitrust complaint on 11 February 2026, alleging that Google was abusing a dominant market position in general search services in breach of Article 102 of the Treaty on the Functioning of the European Union specifically through the deployment of AI Overviews and AI Mode. The complaint’s core argument, as summarised in the EPC’s own executive summary, is that by embedding AI-generated summaries and chatbot-style responses directly into its dominant search interface, Google has transformed Search “from a referral service into an answer engine that substitutes original publisher content and retains users within Google’s own ecosystem” — language that captures precisely the zero-click dynamic documented in the previous section, reframed explicitly as a competition-law harm rather than merely a commercial disappointment for the publishing industry. The EPC’s specific requested remedies included mandating meaningful and enforceable publisher control over how their content is used by AI systems, transparency requirements around the traffic and revenue impact of AI-driven features, and a fair licensing and remuneration framework that would reflect the genuine economic value of the journalism being summarised.

Regulatory pressure has not been confined to formal antitrust complaints filed by outside parties; the European Commission has also begun issuing direct financial penalties against Google under the separate Digital Markets Act framework. On 23 July 2026, the Commission fined Google slightly over one billion dollars combined across two separate decisions — approximately 530 million dollars specifically for giving Google’s own shopping, hotel, transport, and sports listings more prominent placement in search results than competing third-party services, and a further roughly 495 million dollars for restricting how app developers on Google Play could direct customers toward cheaper offers available outside the Play Store. This was described as Google’s first penalty under the Digital Markets Act since the law took effect in 2022, and while the specific violations penalised in this decision predate the AI Mode self-preferencing allegations raised in the EPC’s own complaint, the ruling establishes a directly relevant legal and enforcement precedent for how EU regulators are prepared to treat self-preferencing behaviour by Google across its search products more broadly — precisely the kind of behaviour the EPC’s separate complaint alleges is occurring through AI Mode and AI Overviews specifically.

The United Kingdom’s Competition and Markets Authority has pursued a parallel but procedurally distinct track, confirming it had received a comparable complaint from publishers and, by June 2026, going further than the EU Commission had at that point by directly ordering Google to provide publishers with a genuine opt-out mechanism specific to AI-driven search features — an enforcement outcome that predates and appears to have directly informed the terms Google subsequently offered French publishers as a condition of finally launching AI Overviews and AI Mode in France, discussed earlier in the section on geographic rollout, where the reported terms included a publisher opt-out, transparency commitments, and neighbouring-rights compensation closely tracking the structure the CMA had already required in the UK.

Reuters Institute’s separate 2026 predictions report captured the industry’s own forward-looking expectation amid this regulatory activity starkly: publishers surveyed for the report expect search engine referral traffic to decline by a median of 43 percent over the following three years, with roughly one in five expecting declines greater than 75 percent — expectations that, if they materialise anywhere close to that scale, would represent a fundamental restructuring of how digital publishing generates audience and revenue, regardless of whatever legal remedies the various antitrust complaints and investigations described in this section ultimately produce, on whatever timeline European and other regulators eventually move at.

Google-Extended, opt-out controls, and what publishers can actually block

Given how central the absence of a meaningful opt-out mechanism is to the regulatory and legal actions described in the previous section, it is worth explaining precisely what technical controls actually exist, because there is widespread and persistent confusion — including among publishers themselves — about what each control actually does.

Google-Extended is a robots.txt token Google introduced in September 2023, well before AI Mode existed, that lets a website owner control whether their already-crawled content can be used to train and ground Google’s generative AI models, specifically Gemini and Vertex AI. It is not a separate crawler with its own distinct user-agent string; it works as a semantic control layered on top of the existing Googlebot crawl, letting a publisher decouple the crawling itself, which continues to feed ordinary Search indexing, from a specific downstream use of that crawled content for AI model training and grounding.

The single most important and most widely misunderstood fact about Google-Extended is that it does not affect a site’s inclusion in AI Overviews or AI Mode at all. Google states this directly in its own documentation: Google-Extended governs training and grounding use inside Gemini Apps and the Vertex AI API specifically, and a site’s eligibility to appear in AI Overviews or AI Mode depends only on the ordinary requirements of being indexed and snippet-eligible through Googlebot, entirely independent of whatever a publisher has set for the Google-Extended token. A December 2025 study analysing the top 100 news sites in the US and UK found that 79 percent block at least one bot specifically used for AI training, but that Google-Extended was, notably, the least-blocked training-specific bot among the ones the study examined — a pattern the researchers suggested reflects publishers being more cautious about antagonising Google specifically, given how much more of their traffic depends on Google Search than on any other single referral source, compared with their evident greater willingness to block equivalent training crawlers operated by other AI companies with far less power over publisher traffic through search referrals.

The controls that actually do affect AI Overviews and AI Mode inclusion are different, older tools originally built for an entirely separate purpose: managing conventional search snippets. The nosnippet meta tag instructs Google not to display any text excerpt from a page in search results at all, a directive that does apply to AI Overviews and AI Mode, but which comes with a large and often unacceptable trade-off, since it also removes that page’s ability to show a conventional text snippet in ordinary, non-AI search results — publishers cannot use nosnippet to opt out of AI summarisation specifically while preserving their normal snippet visibility in the ordinary results a majority of search traffic still involves. The max-snippet directive offers a partial, more granular alternative, letting a publisher specify a maximum character length Google may display from the page, which similarly constrains both conventional snippets and AI-generated summaries together rather than allowing separate control over each.

The most extreme available control, blocking Googlebot entirely through robots.txt, removes a site from all of Google Search — AI Overviews, AI Mode, and ordinary results alike — simultaneously and without distinction, an outcome essentially no commercially operating publisher can accept given how much of the web’s overall traffic still flows through Google Search in some form, however diminished that flow has become relative to its historical scale. This all-or-nothing structure, where the only fully effective opt-out from AI summarisation specifically also means complete removal from Search altogether, is precisely the mechanism the various antitrust complaints described in the previous section identify as the core competitive harm: publishers are, in practical commercial terms, not offered a genuine choice at all, only a choice between accepting AI summarisation of their content on Google’s terms or forfeiting the search visibility their business depends on.

In direct response to this sustained pressure, Google began piloting a more granular, product-level opt-out mechanism through 2026. A new toggle introduced inside Google Search Console lets a website owner decide, separately for each of three distinct surfaces — AI Overviews, AI Mode, and AI Overviews specifically within Google Discover — whether they want that property to appear in and help ground responses for that surface, with settings that can be configured independently or set to inherit from a parent-level default. Google’s own principal for product management, Ron Eden, described this as building on the company’s existing framework of open-standards-based controls, explicitly stating: “we’re now exploring updates to our controls to let sites specifically opt out of Search generative AI features. Our goal is to protect the helpfulness of Search for people who want information quickly, while also giving websites the right tools to manage their content.” As of the reporting available for this analysis, the new Search Console toggle remained available only to a limited group of website owners as Google tested it ahead of any wider rollout, and it applies at the whole-site level rather than allowing page-by-page control — Google has been given until March 2027, under terms tied to the broader regulatory pressure described in the previous section, to add genuinely page-level granularity to the mechanism.

The overall picture, even accounting for this new toggle, remains one of partial and still-evolving controls rather than a single clean, universally understood opt-out. A publisher genuinely trying to manage their exposure to AI summarisation across every relevant Google surface, as of this analysis, needs to understand and separately configure at minimum three distinct mechanisms — Google-Extended for AI training and grounding, the older snippet-control directives for summary-level display with their unavoidable side effects on conventional snippets, and the newer, still-limited Search Console toggle for product-level AI feature inclusion — a level of technical complexity that itself represents a meaningful barrier for the large number of smaller publishers who lack a dedicated technical SEO team capable of navigating and correctly configuring all three simultaneously.

Data privacy, retention, and training on personal content

The privacy questions raised specifically by Personal Intelligence, covered in an earlier section, sit inside a broader set of data-handling practices that apply to AI Mode as a whole, and it is worth separating what is specific to the personal-context feature from what applies to every AI Mode query regardless of whether Personal Intelligence is enabled.

Every AI Mode query, personal-context features aside, involves Google processing the query text itself, any uploaded images or files, and the retrieved web content needed to answer it, within Google’s existing search infrastructure and subject to Google’s general search privacy policies. This baseline layer of data handling is not meaningfully different in kind from how Google has always processed search queries; what has changed with AI Mode is the depth of processing applied to that query, and the fact that a single conversational session now carries context across multiple turns rather than treating each search as an entirely independent, stateless event the way the classic search box has traditionally operated.

The multi-turn, contextual nature of an AI Mode conversation is itself a meaningful privacy consideration distinct from Personal Intelligence specifically. Because AI Mode retains conversational context within a session — remembering an earlier question when interpreting a follow-up — the system necessarily holds a richer, more revealing picture of what a person is actually trying to figure out than a series of disconnected, one-off searches would ever reveal individually. A sequence of searches about symptoms, then treatment options, then insurance coverage, run as three separate stateless queries, reveals less about a specific ongoing personal situation than the same three questions asked as a single connected AI Mode conversation, where the system explicitly links them together as one coherent line of inquiry in order to answer the third question well.

Google’s stated position specifically regarding Personal Intelligence — that the system does not train directly on a user’s Gmail inbox or Google Photos library, with training instead limited to AI Mode prompts and the model’s own responses to them — has already been covered in detail in an earlier section, and it represents the clearest specific commitment Google has made regarding personal-content data flows. That commitment does not extend, at least not with the same explicit public framing, to the broader category of ordinary AI Mode query content that falls outside the Personal Intelligence feature specifically, and Google’s general search privacy policies, rather than any AI Mode-specific carve-out, continue to govern how ordinary query and interaction data is retained and potentially used more broadly.

The information agents feature discussed earlier introduces its own distinct privacy dimension, because a persistent background-monitoring agent, by design, holds and continuously acts on a standing description of a user’s stated interest between sessions, rather than only processing data at the moment of an active query. This is a structurally different data-retention pattern from either ordinary AI Mode queries or Personal Intelligence’s per-query context lookup, and it raises its own separate question, not addressed in Google’s currently available public documentation with the same specificity applied to Personal Intelligence, about exactly how long a standing information-agent task persists, what happens to the record of what it has already found and reported when a user eventually cancels it, and whether any of the intermediate search activity an information agent generates while monitoring on a user’s behalf is treated differently from an equivalent one-off query a user typed manually.

Regulatory attention to these questions has so far focused more heavily on the agentic-action and browser-automation side of AI Mode’s capabilities than on the personal-context features specifically, at least based on the reporting available for this analysis. The gap in Auto Browse’s rollout to the UK and EU, discussed earlier in the section on Mariner’s successor technologies, was explicitly tied by reporting to unresolved questions about how an agent that auto-fills forms or handles stored authentication credentials on a user’s behalf would need to operate to satisfy GDPR’s stricter consent and data-processing requirements — a regulatory friction point that has, at least as of this analysis, delayed a specific capability’s international rollout more directly and more visibly than any comparable friction has yet delayed Personal Intelligence’s own international expansion, which, as covered earlier, moved considerably faster from a narrow paid pilot to broad free availability than most of AI Mode’s other advanced features have.

For a person deciding how to use AI Mode’s more advanced, data-connected features, the practical distinction worth holding onto is between features that process personal data per query, with a visible on-off toggle and an easy undo mechanism, such as Personal Intelligence, and features that establish an ongoing, standing relationship with a user’s data and interests across time, such as information agents, where the privacy surface is structurally larger and the currently available public documentation on exactly how that standing data is retained and eventually discarded is considerably less detailed and less specific than what Google has published regarding the comparatively simpler, single-query Personal Intelligence feature.

How AI Mode differs from ChatGPT Search and Perplexity

AI Mode did not develop in isolation, and understanding it fully requires placing it against the two products most frequently used as points of comparison — OpenAI’s ChatGPT Search and Perplexity’s dedicated answer-engine product — because the architectural choices each company made reveal what each considers most important about conversational search, and the market-share data through 2026 shows those choices producing genuinely different competitive outcomes.

Market-share figures for AI-specific search and chat products vary considerably depending on methodology, and several credible research firms tracking the same period through 2026 report meaningfully different numbers, a divergence worth stating plainly rather than picking whichever single figure sounds most dramatic. One analysis using Similarweb data reported ChatGPT’s traffic share within the AI chatbot category falling from 86.7 percent in January 2025 to 64.5 percent in January 2026, alongside Gemini overtaking Perplexity for second place in the same category at somewhere between 18 and 21.5 percent chatbot share by mid-2026, with Perplexity’s own referral traffic share reportedly declining more than 40 percent from its own April 2025 peak as Gemini and ChatGPT Search both expanded around it. A separate research firm’s figures for the same general period reported ChatGPT holding closer to 74.78 percent of AI referral traffic in June 2026, illustrating just how much these headline percentages shift depending on exactly what underlying activity is being measured — total chatbot usage, AI-specific referral traffic to third-party websites, or absolute monthly active user counts — and readers should treat any single cited percentage in this space as one estimate among several plausible ones rather than a settled, universally agreed figure.

What the data sources broadly agree on, despite disagreeing on precise percentages, is the relative scale of the underlying markets: Google retains roughly 80 to 90 percent of total global search query volume across all query types, even as AI-native products capture a growing and increasingly consequential share specifically of informational, research, and comparison-shopping queries — the exact query categories that have historically driven the most organic traffic to independent websites and publishers. Google’s own defensive strategy against this pressure, as this analysis has documented throughout, has been to integrate Gemini as deeply as possible directly into Search itself through AI Overviews and AI Mode, effectively turning Google’s own dominant search product into a comparably capable AI answer engine rather than ceding the entire category of complex, research-oriented queries to a separate class of competitor products built from scratch around conversational AI.

Perplexity has built its entire product identity around a specific architectural choice that distinguishes it clearly from both AI Mode and ChatGPT Search: dense, prominent, numbered source citation as the core visible feature of every answer, typically five to fifteen numbered references per response, a design choice that industry analysis has consistently described as offering the highest citation yield per query among the major AI search products, making Perplexity a particularly attractive target for brands and publishers specifically seeking citation volume even though its absolute traffic scale remains far smaller than Google’s. Perplexity’s growth trajectory has nonetheless been genuinely dramatic in relative terms even as its absolute market share stayed comparatively small — one analysis put its year-over-year user growth at roughly 370 percent through early 2026, nearly triple ChatGPT’s growth rate of about 125 percent over the same window, alongside reported annualised revenue crossing 450 million dollars by March 2026 and a publicly discussed target IPO timeline around 2028.

ChatGPT Search, by contrast, inherits its core design philosophy from ChatGPT’s own conversational-assistant identity rather than building citation density as its primary distinguishing feature the way Perplexity does — its search integration functions more as one capability layered onto an already-dominant general-purpose conversational assistant than as a standalone, search-first product built around citation transparency as its core value proposition. This distinction in origin and design philosophy — search-first with heavy citation emphasis for Perplexity, general-assistant-first with search as an added capability for ChatGPT, search-engine-first with a new conversational layer added on top for AI Mode — helps explain why direct feature-for-feature comparisons across the three products can be genuinely misleading if they ignore what each company actually optimised for from the outset, rather than treating all three as interchangeable variations on a single, identical underlying product category.

How the three major conversational search products differ in design emphasis

None of these emphases is inherently superior; they reflect what each company chose to optimise for first, and the table is best read as a summary of design philosophy rather than a ranking, since the market-share figures earlier in this section already show all three products growing in absolute terms during the same period.

AI Mode’s most distinctive structural advantage over both rivals is one this analysis has returned to repeatedly throughout: its direct integration with Google’s existing, decades-deep search infrastructure, including the Knowledge Graph, the continuously updated Shopping Graph, and Google’s broader index of the web, none of which either ChatGPT Search or Perplexity can access natively at anything close to the same depth, since both depend more heavily on their own more recently built retrieval infrastructure and, in varying degrees, licensing or indexing arrangements with third-party data providers rather than decades of Google’s own accumulated, purpose-built search infrastructure. Whether that structural advantage translates into a durable long-term competitive lead depends on factors well beyond retrieval infrastructure alone — model quality, user trust, advertiser relationships, and regulatory outcomes across multiple jurisdictions all remain live, unresolved variables as of this analysis, and the sharply competitive “code red” dynamic between Google and OpenAI documented earlier in the section on information agents suggests neither company currently regards its own position as secure regardless of whatever infrastructure advantage either side may hold.

Where Google still leads and where it has lost ground

Pulling together the architecture, the market data, and the regulatory pressure documented so far in this analysis, a reasonably clear picture emerges of where Google’s position inside conversational search is genuinely strong and where it has demonstrably weakened, rather than treating “AI Mode’s competitive position” as a single, undifferentiated question with one answer.

Google’s clearest and most durable advantage is scale of infrastructure that took decades to build and cannot be replicated quickly by any competitor, however well-funded. The Shopping Graph’s roughly two-billion-updates-per-hour refresh rate, the Knowledge Graph’s decades of accumulated structured entity data, and Google’s index of the web itself, built continuously since the late 1990s, together give AI Mode a retrieval foundation that no AI-native competitor built from scratch since 2022 can match in comparable depth, regardless of how sophisticated that competitor’s underlying language model may be in isolation. This advantage shows up most clearly in exactly the categories where structured, rapidly changing, transactional data matters most — shopping, local business information, and real-time information generally — which is precisely why Google’s own commerce-layer investment, covered earlier in the sections on shopping ads and the Shopping Graph, concentrates so heavily in these categories rather than spreading evenly across all query types.

Google’s clearest weakness, by contrast, is trust and brand association specifically around advertising inside AI-generated answers, a weakness that is self-inflicted rather than structural. The sharp public reaction documented earlier to sponsored listings appearing mid-conversation inside AI Mode responses — the Kaziukenas criticism explicitly invoking Anthropic’s own Super Bowl advertising that had mocked the idea of ChatGPT introducing ads — reflects a specific reputational vulnerability tied directly to Google’s core revenue model. Google’s entire business depends on search advertising in a way that neither Perplexity’s nor, at least for now, ChatGPT’s core business model depends on comparable advertising revenue to the same degree, which means Google faces a structural pressure to monetise AI Mode through advertising that its two most-discussed competitors, at least as of this analysis, do not face with equivalent urgency — a pressure that shows up directly in how quickly and how visibly Google moved to introduce ads into AI Mode once the product reached sufficient scale, compared with a slower, more cautious approach either rival has so far taken toward advertising inside its own equivalent conversational answer product.

Google’s regulatory exposure is also structurally distinct from either rival’s, purely as a function of Google’s market position rather than anything specific to AI Mode’s technology. The European Commission’s antitrust investigation, the EPC’s Article 102 complaint, and the Digital Markets Act fine covered earlier in this analysis all depend on Google’s underlying dominance in general search — a level of market dominance neither Perplexity nor ChatGPT Search holds in the same category, and therefore a level of regulatory scrutiny neither currently attracts to a comparable degree specifically because dominance, not simply having an AI feature, is what triggers this specific category of EU competition-law exposure. This creates a genuine strategic asymmetry: Google’s very success and market position is itself a source of ongoing legal and reputational risk that its smaller AI-native competitors, purely by virtue of their smaller scale, do not currently share to nearly the same extent.

On raw usage growth, the picture through 2026 shows genuine, well-documented momentum for Google specifically, even while overall AI-native competitors were also growing rapidly in parallel — this is not, based on the data covered in this analysis, a zero-sum contest where Google’s gains necessarily came at a rival’s direct expense, or vice versa. AI Mode’s own growth from 100 million monthly users in November 2025 to over one billion by May 2026 represents genuine, substantial expansion in absolute terms, occurring during the same general window in which Perplexity was also reporting user growth in the hundreds of percent and ChatGPT continued adding users at a large absolute scale. The most defensible reading of the available data is that the overall category of AI-assisted search and research is expanding rapidly enough, across the entire market, that multiple major players have been able to report substantial simultaneous growth without one company’s gains being drawn primarily from another’s losses — though this reading could shift as the market matures and growth rates for the category as a whole inevitably slow from their current early-stage pace.

The most honest overall assessment, synthesising everything covered in this analysis, is that Google holds a genuine and probably durable structural advantage in retrieval infrastructure and breadth of integration across its existing product ecosystem, carries a genuine and largely self-inflicted reputational vulnerability specifically around advertising trust, and faces a genuine and structurally distinct regulatory exposure tied directly to its underlying market dominance — three separate and only loosely related competitive dimensions, none of which alone determines whether AI Mode ultimately succeeds or fails as a long-term product, and all three of which will likely continue evolving independently of each other rather than converging toward one single, clean verdict on Google’s overall competitive position in conversational search.

Business impact by sector

The effects documented throughout this analysis do not land evenly across the economy. Five sectors illustrate how differently AI Mode’s fan-out retrieval, agentic booking, and shopping infrastructure reshape competition depending on what a business actually sells and how its customers research before buying.

Retail and e-commerce face the most direct structural change, because AI Mode’s shopping-specific infrastructure — the Shopping Graph, shopping ads, and Direct Offers, all covered earlier in this analysis — was built specifically around this sector’s query patterns. A retailer with clean, complete, properly formatted product feed data becomes eligible for both organic fan-out citation and paid shopping placements inside AI Mode responses; a retailer without that structured data simply does not exist inside the shopping-specific retrieval pool, regardless of how well its website otherwise ranks in conventional search. The comparatively low AI Overview trigger rate on shopping queries documented earlier — around 3.2 percent, down from an initial roughly 29 percent, after Google found the format converting poorly for commercial intent — suggests Google itself recognises that shopping decisions still benefit from visible, comparable product listings rather than a single synthesised summary, a nuance that cuts against the assumption that every category of query is moving uniformly toward AI-generated answers.

Travel and hospitality sit at the centre of the agentic-booking capabilities discussed earlier, through the Mariner-derived ticket and reservation partnerships with platforms including Ticketmaster, Resy, and StubHub. A travel business integrated into this agentic booking layer becomes directly reachable by an AI Mode session that fans out across real-time pricing and inventory on a user’s behalf; one that is not integrated risks losing the booking step entirely to a competitor that is, even if the business would otherwise have won the customer’s attention through conventional search visibility. Personal Intelligence adds a further layer specifically relevant to this sector, since travel bookings sitting in a user’s Gmail are one of the explicit examples Google has used to illustrate how personalised recommendations get built.

Publishing and digital media face the most severe and best-documented negative impact of any sector covered in this analysis, detailed extensively in the earlier sections on the zero-click debate and the resulting antitrust complaints. The named traffic declines already cited — HubSpot’s estimated 70 to 80 percent organic decline, Business Insider’s 55 percent loss and associated staff cuts, CNN’s 27 to 38 percent decline, Chegg’s 24 percent revenue fall tied explicitly to AI Overviews in litigation — represent a sector whose historical business model depended on exactly the referral-traffic mechanism that AI Mode and AI Overviews together have most directly undermined, with the regulatory remedies discussed earlier still unresolved and, even where they exist, applying unevenly across different jurisdictions.

Local services — plumbers, contractors, medical practices, restaurants, and similar businesses reliant on local search discovery — face a more mixed picture than publishing does. Location-based queries and shopping-adjacent local intent have, per the sector-level data cited earlier in this analysis, seen comparatively less aggressive AI Overview penetration than informational categories such as health content, and Google’s own local-business-promotion plans, mentioned in the advertising section as a stated future direction rather than a confirmed feature, suggest the company sees continued commercial value in preserving a recognisable local-business discovery experience rather than fully collapsing it into pure AI synthesis. That said, a local business entirely dependent on ranking for informational, how-to-style content adjacent to its core service — a plumber’s blog explaining how to unclog a drain, for instance — faces the same broad category of zero-click pressure documented earlier for informational queries generally, even if the core transactional “plumber near me” query type has so far proven more resistant to full AI-answer replacement.

B2B software and enterprise services occupy a distinctive position specifically because AI Mode is explicitly positioned, as covered in the comparison with AI Overviews earlier in this analysis, as the surface for high-consideration, multi-criteria research — precisely the kind of research a B2B buyer does before recommending a software platform to their organisation. Industry analysis covered earlier framed this directly: a buyer comparing five competing enterprise platforms across a dozen criteria is a canonical AI Mode use case, which means earning a positive citation inside that comparative research process carries outsized weight relative to a single click on a conventional results page, precisely because the AI Mode session represents a buyer already deep inside a serious evaluation process rather than a casual, early-stage information gather. This makes AI Mode visibility a genuinely different, and in some ways higher-stakes, marketing objective for B2B software vendors than conventional SEO ranking ever was, since a single well-earned citation inside a comparative AI Mode research session may influence a purchase decision more directly than ranking position ever reliably did on a page the buyer would have scanned quickly and then clicked through several competing options from anyway.

Across all five sectors, the pattern that recurs is that AI Mode rewards businesses that have already invested in structured, machine-readable, clearly organised information about themselves — product feeds, booking-system integrations, clearly stated comparative facts — and penalises businesses whose competitive advantage previously rested on winning a click through compelling but unstructured page content, a distinction this analysis returns to directly in the practical guidance that follows.

Practical guidance for earning a citation inside AI Mode

Given everything documented so far about how query fan-out actually retrieves and cites content, several concrete practices follow directly from the mechanics rather than from speculation, and they are worth stating plainly for any business or publisher trying to act on this analysis rather than simply understand it.

Map the fan-out cluster around a topic, not just the head keyword. Because a fan-out process for a compound question generates a dozen or more parallel sub-queries that rarely match the literal phrase a user typed, the practical research task shifts from identifying one target keyword to anticipating the full set of related sub-questions a thorough human researcher would naturally ask around the same topic — the design and connectivity and battery-life facets of the headphone example covered earlier in this analysis, applied to whatever specific topic a given business needs to be visible for. Content that answers several of those sub-facets directly, on the same page or across a clearly interlinked cluster of pages, has more surface area for fan-out retrieval to draw on than content narrowly optimised for a single head-term phrase alone.

Write extractable, self-contained answers near the top of a page rather than requiring the reader to infer the answer from surrounding narrative. The citation-behaviour section earlier in this analysis noted that structured, clearly stated content performs disproportionately well specifically because a retrieval-and-synthesis pipeline can extract and cite it with more confidence than content whose actual answer is buried inside long, otherwise well-written prose. This does not mean abandoning depth or narrative quality — it means placing a direct, factually complete answer early, and reserving elaboration, nuance, and supporting detail for what follows it, rather than building toward the answer gradually across several paragraphs the way traditional long-form editorial writing has often done.

Treat structured data as infrastructure, not decoration, particularly for any business with a commerce or booking component. The Shopping Graph section of this analysis was explicit that accurate, complete, properly formatted product feed data is the direct mechanism determining shopping-query eligibility, independent of a page’s general content quality; the same logic extends, less formally documented but observably true in practice, to schema markup more broadly, since a retrieval system extracting facts at scale benefits from machine-readable structure the same way a human researcher benefits from a well-organised reference table over an equivalent fact buried in unstructured prose.

Diversify visibility measurement beyond conventional keyword-ranking tools, since Google has been notably reluctant to share first-party AI Mode citation data directly. The earlier section on citation behaviour noted that SEO teams at major publishers have turned to third-party tools such as AlsoAsked and Profound, alongside adjacent features like People Also Ask, specifically because no reliable first-party AI Mode visibility report currently exists from Google itself; any business serious about tracking its AI Mode presence needs to build measurement practice around these indirect signals rather than assuming conventional rank-tracking tools, built for the pre-AI-Mode search era, will capture what actually matters now.

Do not treat Google-Extended as an AI Mode opt-out, and do not assume blocking it protects anything relevant to AI Mode visibility, in either direction. The opt-out section of this analysis was explicit that Google-Extended governs training and grounding for Gemini specifically, not AI Overviews or AI Mode inclusion — a business trying to manage its AI exposure needs to separately evaluate the newer, still-limited Search Console toggle, the older snippet-control directives with their side effects on conventional search visibility, and Google-Extended, understanding that each governs a genuinely different downstream use of crawled content rather than functioning as three overlapping versions of the same control.

Recognise that citation and advertising visibility inside AI Mode are becoming linked rather than separate channels, per the halo-effect observation covered in the citation-behaviour section, and plan accordingly rather than treating organic AI Mode presence and paid AI Mode advertising as entirely independent budget lines the way organic SEO and paid search advertising have traditionally been managed as separate disciplines within most marketing organisations.

Structured data, entities, and what actually helps visibility

The previous section touched on structured data as a practical recommendation; this section explains the underlying mechanism in more technical depth, because the reasoning behind why structure matters is more useful than the recommendation alone, and it connects directly back to how the retrieval pipeline described early in this analysis actually processes a page.

A retrieval-and-synthesis system fundamentally operates over entities and their relationships to each other, not over raw, undifferentiated text — a product, a specification, a price, a location, a date, and the specific relationships connecting them, such as which product has which price or which business operates at which location. Schema markup, the structured-data vocabulary that has existed since long before AI Mode did, exists precisely to state these entity relationships explicitly and unambiguously, in a format a machine can parse directly rather than having to infer probabilistically from surrounding natural-language text. A page stating a product’s price only within a sentence of flowing prose forces the retrieval system to infer that the specific number mentioned is, in fact, the price, and infer which specific product it attaches to if several products are discussed on the same page; a page with explicit Product schema markup states that relationship directly, removing the inference step and the error risk that comes with it.

This distinction matters more for AI Mode specifically than it did for classic search, because classic search’s ranking algorithm has always been reasonably tolerant of ambiguity that a human reader would resolve easily just by reading the page in context, while AI Mode’s synthesis stage is actively extracting and recombining specific facts across many different retrieved sources at once, at a speed and scale that makes the individual inference judgement calls a human reader makes instinctively considerably more error-prone when performed automatically, particularly across a fan-out retrieval pool spanning a dozen or more distinct pages simultaneously feeding into one synthesised response. A single ambiguous fact on a single page is a minor and probably harmless imperfection for a human reader who can use surrounding context, tone, and general world knowledge to resolve it correctly; the same ambiguity, multiplied across the entire retrieval pool feeding a single synthesised AI Mode answer, becomes a more consequential source of the misattribution and false-synthesis errors already documented in the earlier section on why AI Mode sometimes gets things wrong.

Entity clarity extends beyond formal schema markup into more general content-organisation practice: consistent naming, clear disambiguation when a business or product shares a name with something unrelated, and explicit statement of relationships between related entities a page discusses together, all reduce the inference burden on a retrieval system in the same underlying way schema markup does, even without using the formal schema vocabulary itself. A page comparing three products by name, but referring to them inconsistently by nickname in some paragraphs and full model number in others, forces exactly the kind of disambiguation inference that a retrieval-and-synthesis pipeline is more likely to get wrong at scale than a careful human reader working through the same page slowly would be.

None of this is a fundamentally new discipline invented for AI Mode — structured data, clear entity disambiguation, and unambiguous factual statement have been considered SEO best practice for years, well before AI Mode existed, precisely because search engines have always benefited from exactly this kind of clarity to varying degrees. What has changed with AI Mode’s arrival is the stakes attached to getting it right: a classic search ranking algorithm tolerating some ambiguity might simply rank a page slightly lower than a clearer competitor; an AI Mode synthesis process working with ambiguous source material can produce an outright factually incorrect statement inside a fully-cited, confident-sounding answer, a failure mode with materially different consequences for both the business being inaccurately represented and the end user relying on that answer to make a decision.

What Central European businesses should watch for

Everything covered in this analysis so far applies globally, but several details are specifically relevant to businesses operating in Slovakia and the wider Central European region, given the specific timeline and terms under which AI Mode reached this part of Europe.

AI Mode’s arrival in the European Union came through the large coordinated expansion of 7 October 2025, covered earlier in this analysis, which brought the feature into more than 40 additional countries and territories including most of the EU simultaneously, rather than through a separate, later, region-specific rollout the way France’s delayed launch, discussed extensively in the section on publisher lawsuits, eventually required. For a Slovak business, this means AI Mode’s fan-out retrieval, its citation behaviour, and its shopping infrastructure have all been operating against Slovak and broader Central European search behaviour for roughly a year by the time of this analysis — long enough that the visibility patterns and practical guidance covered earlier in this piece already apply directly, rather than being a preview of something not yet relevant locally.

The age restriction applied to AI Overviews in the European Union specifically — limiting the feature initially to signed-in users eighteen and over, tied directly to EU regulatory considerations that did not apply in the same form in the United States — is a detail worth understanding precisely because it illustrates a broader pattern: EU-specific regulatory requirements have shaped, and will likely continue to shape, exactly which AI Mode capabilities are available in this region and under what conditions, sometimes with a materially different feature set or a different rollout timeline than what US-based coverage of AI Mode describes. A Slovak business researching AI Mode primarily through US-focused technology coverage should treat any specific claim about feature availability with some caution until confirming it applies in the EU specifically, given how many of this analysis’s own sections have documented EU-specific gates, delays, or additional legal conditions layered onto features that launched with fewer restrictions in the US.

The Digital Markets Act enforcement actions documented earlier in this analysis, including the roughly one-billion-dollar combined fine issued in July 2026 for self-preferencing behaviour in search results, apply with particular direct relevance to any Slovak business competing against Google’s own shopping, travel, or comparison services inside search results — a local hotel comparison site, for instance, competing against Google’s own travel listings that the DMA fine specifically targeted for receiving preferential placement. Businesses in exactly this competitive position should track DMA enforcement developments closely, since the remedies the European Commission ultimately imposes could directly affect how prominently a competing local service is permitted to appear relative to Google’s own equivalent product inside both conventional search results and, potentially, AI Mode responses specifically, given that the EPC’s separate antitrust complaint names AI Mode explicitly as a surface where this same self-preferencing dynamic is alleged to be occurring.

Central and Eastern Europe’s broader economic position in technology manufacturing, though not the specific focus of this analysis, provides useful context for why EU-level AI search regulation carries disproportionate weight for the region’s businesses generally: the region hosts a substantial concentration of automotive electronics manufacturing, assembly, and testing operations across Slovakia, the Czech Republic, Hungary, and Poland, sectors that depend heavily on B2B research and comparison processes of exactly the kind AI Mode is positioned to intermediate, as discussed in the earlier section on business impact by sector. A regional supplier bidding for automotive electronics contracts, evaluated by a procurement team that increasingly does its comparative vendor research through AI Mode rather than a traditional search results page and a series of manually opened supplier websites, faces a genuinely new visibility challenge that overlaps directly with the structured-data and entity-clarity guidance covered in the two preceding sections — a supplier whose technical specifications, certifications, and comparative advantages are not stated in clearly extractable, machine-readable form risks becoming effectively invisible inside exactly the AI-mediated research process an increasing share of procurement decisions now run through.

The practical recommendation for any Central European business following this analysis is straightforward: do not assume US-focused AI Mode coverage describes your actual feature availability, timeline, or regulatory context precisely, and treat the EU’s distinct and generally more protective regulatory posture toward both AI features and publisher rights as a genuinely different operating environment rather than a slightly delayed version of the US rollout.

Open questions and where AI Mode goes next

Enough of AI Mode’s architecture and impact is now documented well enough to describe with confidence, as this analysis has tried to do throughout. Several genuinely open questions remain, though, and honesty about what is not yet settled is worth as much as confidence about what is.

Whether classic search results disappear entirely, or persist indefinitely alongside AI Mode as a permanent parallel option, remains unresolved by anything Google has stated on the record. Liz Reid’s direct statement following the I/O 2026 keynote — “this new search box does not mean that you’ll only get AI responses” — is the clearest available signal, and it points toward continued coexistence rather than replacement, but it is one executive’s characterisation at one point in time rather than a binding commitment, and the entire pattern of AI Mode’s rollout documented throughout this analysis has consistently been lateral expansion — more countries, more features, more surfaces — rather than the kind of frontal substitution that would definitively answer this question either way.

How the various open antitrust complaints and investigations covered in this analysis will actually resolve, and on what timeline, is genuinely uncertain. The European Commission’s formal investigation opened in December 2025 remains active as of this analysis, the European Publishers Council’s Article 102 complaint filed in February 2026 had not produced a final Commission decision by the time this piece was written, and EU antitrust investigations of this scale have historically taken years rather than months to reach enforceable conclusions — the Digital Markets Act fine issued in July 2026 concerned separate, already-established self-preferencing violations rather than resolving the AI-Mode-specific allegations the EPC’s complaint raises. Readers should expect this regulatory story to remain unresolved for a considerable further period rather than anticipate a near-term, decisive ruling one way or the other.

Whether the zero-click traffic decline documented extensively in this analysis has already reached its floor, or continues deepening, is contested even among the researchers actively measuring it. The direct disagreement between SparkToro’s June 2026 finding of continuing decline and Datos’s Q1 2026 finding of a slight reversal, both covered in detail earlier in this analysis, has not been resolved by any single subsequent study definitive enough to settle which trend accurately describes where publisher traffic is actually heading through the remainder of 2026 and beyond.

The long-term viability of the publishing business models most affected by this shift remains an open economic question rather than a settled one, regardless of whatever specific remedies eventual regulatory action produces. Even a publisher opt-out mechanism as strong as the one the UK’s Competition and Markets Authority has already ordered, or the comparable terms reportedly negotiated for France’s delayed launch, addresses the fairness and consent dimension of the problem without necessarily restoring anything close to the referral-traffic volume publishers depended on before AI-generated summaries became the default way a large share of searchers get their answer — a distinction between winning a fairer negotiating position and actually recovering lost revenue that several of the sources cited in this analysis note explicitly.

Whether AI Mode’s information agents and expanding agentic booking capabilities will extend from their currently narrow, expensive-subscription-tier availability to the free tier on a timeline comparable to Personal Intelligence’s unusually fast expansion, or instead remain gated behind paid tiers considerably longer, is not yet determinable from the pattern established so far, since the two features have followed genuinely different trajectories to date for reasons Google has not explained in detail. Personal Intelligence moved from a narrow US paid pilot to a free global feature in around five months; information agents, as of this analysis, remained restricted to the most expensive AI Ultra tier with only a promise of eventual AI Pro access and no stated free-tier timeline at all, and nothing in Google’s public communication explains why these two comparably important features have been treated so differently in their rollout pacing.

What can be said with confidence, pulling together everything this analysis has documented, is that AI Mode is no longer an experiment by any reasonable definition of the word. A product used by more than a billion people monthly, generating advertising revenue through named, launched formats, large enough to draw a formal European Commission antitrust investigation and a separate publisher-led Article 102 complaint naming it specifically, and central enough to Google’s own strategic communication that the company built an entire I/O 2026 keynote narrative around search agents and a redesigned search box, has moved well past the pilot phase it occupied in Search Labs in March 2025. What remains genuinely uncertain is not whether AI Mode matters, but exactly how the competing pressures documented throughout this analysis — advertiser revenue against user trust, publisher survival against Google’s own retrieval economics, rapid feature expansion against regulatory constraint — ultimately settle into whatever the product looks like several years from now.

Reader questions about how Google AI Mode actually works

Author:
Jan Bielik
CEO & Founder of Webiano Digital & Marketing Agency

This article is an original analysis supported by the sources cited below

Search Io 2026 Google’s own product blog post announcing the Google I/O 2026 Search updates, including Gemini 3.5 Flash as the new default AI Mode model, the redesigned search box, and information agents.

100 things we announced at Google I/O 2026 Google’s full roundup of I/O 2026 announcements, covering Gemini 3.5 Flash’s benchmark performance, the seamless AI Search experience, and the introduction of Search agents.

Personal Intelligence: how AI Mode personalizes responses Google’s own announcement of Personal Intelligence in AI Mode, including Robby Stein’s statements on how the feature connects Gmail and Google Photos context to search responses.

Personal Intelligence: How AI Mode in Search personalizes responses for you Google Search Help’s official support documentation covering how to connect, disconnect, correct, and manage Personal Intelligence, including the age restriction and memory-control commands.

Google’s AI Mode ‘query fan-out’ technique A detailed technical breakdown of how query fan-out expands a single query into multiple sub-queries across the live web, Knowledge Graph, and specialised data sources.

WTF is “query fan-out” in Google’s AI mode? Digiday’s reporting on query fan-out from the perspective of publisher SEO teams, including commentary from Mike King of iPullRank on shifting optimisation strategy.

Query Fan-Out Technique in AI Mode: New Details From Google Search Engine Journal’s coverage of Google’s Robby Stein describing Deep Search and disclosing usage-scale figures for AI-powered search experiences.

Google expands AI Mode to over 40 countries and territories Coverage of the October 2025 international expansion of AI Mode, including Hema Budaraju’s statements on the scale of the rollout.

Google launches AI Mode in 180 countries and territories Search Engine Land’s account of AI Mode’s earlier rollout history, from the March 2025 Search Labs launch through the mid-2025 India and US expansions.

Google AI Mode: 1B Users, Gemini 3.5 Flash Reporting on the I/O 2026 disclosure that AI Mode had surpassed one billion monthly active users, with quarterly query-volume growth figures.

Google I/O 2026: AI Mode and end of Search An analysis of AI Mode’s adoption curve from November 2025 through May 2026, including Liz Reid’s direct statement that the redesigned search box does not mean only AI responses.

Google AI Mode Tracker 2026: Availability, Features & SEO Impact A continuously updated tracker of AI Mode’s rollout milestones, including the France launch terms tied to publisher neighbouring-rights negotiations.

Google Expands AI Web-Browsing Agent Project Mariner Coverage of Project Mariner’s mid-2025 expansion, including its Ticketmaster, Resy, and StubHub partnerships for agent-driven commerce.

Why Google killed Project Mariner, its web-browsing AI agent Reporting on the May 2026 shutdown of Project Mariner and the integration of its capabilities into Gemini Agent and AI Mode.

Google shut down Project Mariner and folded its tech into Gemini Detailed reporting on the reasons behind Mariner’s shutdown, including its screenshot-based architecture’s latency and cost issues, and the Auto Browse feature that succeeded it.

Google ‘Sponsored’ Ads Emerge In AI Mode MediaPost’s coverage of user reaction to sponsored listings appearing inside AI Mode responses, including Juozas Kaziukenas’s criticism of mid-conversation ad placement.

Google unveils shopping ads in AI Mode Coverage of the February 2026 shopping ad format launch, including Google’s commissioned consumer research on comparison shopping inside AI Mode.

A new generation of ads for the AI era of Search Google’s own announcement of Conversational Discovery ads and Highlighted Answers, describing how Gemini constructs advertising creative matched to a user’s specific query.

How Google AI Mode ads work today Discovered Labs’ network-traffic analysis of AI Mode’s backend advertising infrastructure, including the identification of the “AI Mode Bottom Ads” placement and query-to-conversion attribution parameters.

Google Search Hits $63B: AI Mode Ads & Monetization in 2026 Coverage of the Direct Offers advertising program launched in January 2026, including its retailer-configured, context-triggered discount mechanism.

Google zero-click searches reach 68% in early 2026: Study Search Engine Land’s coverage of SparkToro’s June 2026 zero-click study, including Rand Fishkin’s analysis of which query categories remain resistant to the trend.

In 2026, Less than One Third of Google Searches Still Send a Click SparkToro’s own detailed report on declining click-through rates, including the Ahrefs 75,000-domain traffic-share tracker data.

AI SEO Statistics (2026): 57+ Data Points A compiled dataset of zero-click and AI Mode citation statistics drawn from named primary studies including Pew Research, Ahrefs, and Seer Interactive.

Google AI Overviews Statistics 2026: 60+ Data Points A statistics compilation including the Reuters Institute and Chartbeat publisher traffic findings and the Datos Q1 2026 State of Search report.

European Publishers Council files formal antitrust complaint against Google The European Publishers Council’s own statement on its February 2026 Article 102 complaint against Google over AI Overviews and AI Mode.

European Publishers File Antitrust Complaint Against Google AI Additional reporting on the EPC complaint, including its requested remedies around publisher control, transparency, and licensing.

EU fines Google $1 billion over Digital Markets Acts violations Coverage of the European Commission’s July 2026 Digital Markets Act fine against Google for self-preferencing in search results and Play Store restrictions.

Google’s AI Overviews Trigger EU Antitrust Complaint Coverage explaining the all-or-nothing opt-out problem publishers face and how the Digital Markets Act framework applies to AI Overviews.

Google faces EU antitrust complaint over AI overviews: report TechCrunch’s original reporting on the Independent Publisher Alliance complaint, including its description of the lack of a viable opt-out mechanism.

Google’s AI Overviews facing EU antitrust probe Coverage of the European Commission’s December 2025 formal investigation announcement, including Commissioner Teresa Ribera’s statement on protecting online press.

Opt Out of AI Overviews: Should You Take Google’s Offer? A detailed walkthrough of the new Search Console opt-out toggle, its site-wide limitation, and the March 2027 deadline for page-level controls.

Google-Extended: What It Controls and What It Does Not A technical explainer distinguishing Google-Extended’s training-and-grounding function from the separate controls that actually affect AI Overviews and AI Mode inclusion.

Google Introduces New Publisher Opt-Out Controls for AI Overviews Coverage of the December 2025 BuzzStream study on publisher AI-crawler blocking rates, including the finding that Google-Extended is the least-blocked training bot.

Google AI Mode vs. AI Overviews: Key Differences Explained An SEO-focused comparison of AI Mode and AI Overviews covering activation triggers, multimodal capability, and content strategy implications.

AI Search Market Share 2026: ChatGPT vs Perplexity vs Others A market-share analysis of the conversational AI search category, including Google’s own AI market-share estimate and category growth projections.

AI Search Market Share & Growth Trends (2026 Report) A detailed 2026 market-share compilation covering ChatGPT, Gemini, Perplexity, and Claude referral-traffic and user-growth figures.

Perplexity AI vs Google Market Share 2026 Coverage of Perplexity’s citation-dense design philosophy, its user-growth trajectory, and its reported revenue figures through early 2026.

Google adds your Gmail and Photos to AI Mode TechCrunch’s original reporting on the January 2026 Personal Intelligence launch, including Robby Stein’s statement on context-aware recommendations.

Google ties AI Search to Gmail and Photos, raising new privacy questions Help Net Security’s privacy-focused analysis of the Personal Intelligence launch and its data-handling implications.

Google launches always-on information agents in Search at I/O 2026 TheNextWeb’s coverage of the information agents announcement, including Elizabeth Reid’s framing of the feature as the start of the era of Search agents.

Google AI Mode starts rolling out Search agents 9to5Google’s reporting on the June 2026 information agents rollout to AI Ultra subscribers, including pricing and trigger-phrase details.

This article was prepared with the assistance of artificial intelligence tools. The content underwent expert human review, and Webiano Digital & Marketing Agency assumes editorial responsibility for its final version and publication.