Strong Google rankings no longer guarantee a citation in AI answers

The Slovak phrase that opens this piece translates to a simple, comforting idea: if your SEO is good, your GEO is good too. It has spread through marketing teams for a reason. It lets a business that already invested a decade in search engine optimization believe the money still works, that generative engine optimization is not a new budget line but a relabeling of the same discipline. The claim is attractive because it is partly true, and dangerous for exactly the same reason.

The claim behind “good SEO is good GEO” and why it caught on

Here is the honest starting position, supported by the data reviewed in this article: strong classic SEO remains the single best foundation for visibility in AI answers, but the correlation between ranking well and getting cited by an AI system has weakened sharply through 2025 and into 2026, and it varies so much by platform that treating “AI visibility” as one number is now a reporting error rather than a shortcut. A page can rank first on Google and be absent from ChatGPT. A page that never cracks Google’s top 100 can be quoted by Perplexity every day. Both facts are now routine.

The idea that good SEO equals good GEO gained traction in late 2024 and early 2025 for a defensible reason. In that period, the overlap between the pages Google cited in its AI Overviews and the pages ranking in the organic top 10 was genuinely high. Ahrefs measured it at around 76 percent in its July 2025 study. When three quarters of AI citations came from top-ranking pages, optimizing for Google’s blue links was, in practice, optimizing for the AI answer that sat above them. A marketer could keep doing the same work and call it GEO.

That period is ending. By early 2026, separate studies put the same overlap between 17 and 38 percent, depending on methodology and data source. The number did not just drift. It roughly halved, and in some datasets fell further. The comfortable equivalence between rankings and citations that made the “good SEO is good GEO” slogan feel obvious has been eroding month by month, and the erosion is measurable rather than anecdotal.

The reason the slogan still holds any water is that the foundation did not move. AI systems that answer questions still lean on the open web, still need to crawl and parse pages, still reward clarity, credibility, and topical depth, and still treat authority as a real thing. A site that is technically broken, thin, slow, or untrustworthy fails in both worlds. What changed is everything above the foundation: how sources are selected, how many are cited, which off-site signals matter most, and how little a single ranking position now guarantees.

This article works through the claim carefully rather than accepting or rejecting it wholesale. It sets out what GEO is, where the shared machinery sits, what the measured overlap actually shows across the major studies, why those studies contradict each other, and which specific SEO habits carry into AI search cleanly, which fail quietly, and which now backfire. It ends with a working answer that a practitioner can act on, not a slogan. The short version, stated up front so the rest can earn it: good SEO buys you a ticket into AI search. It no longer buys you a seat.

Generative engine optimization defined against classic SEO

Search engine optimization is the practice of improving a website so it appears higher in a ranked list of results and earns unpaid clicks. The output is a position. The reward is a click. The whole discipline, for twenty years, has been built around moving up a list and capturing the traffic that a higher position brings.

Generative engine optimization is the practice of improving content so that generative AI systems mention it, cite it, or summarize it inside an answer. The output is not a position on a list. It is inclusion inside a synthesized paragraph. When someone asks ChatGPT, Perplexity, Gemini, or Google’s AI Mode for the best project management tools for a distributed engineering team, they do not receive ten links. They receive a written answer that names two or three options, explains why, and sometimes cites its sources. If your brand is not inside that paragraph, you are not on the shortlist, and the reader may never reach the search results your SEO was designed to win.

The distinction that matters is the unit of success. Traditional SEO succeeds when it earns attention through position. GEO succeeds when it earns representation inside an answer. One competes for a slot on a page. The other competes for inclusion in a sentence. That difference sounds subtle and is not. A ranked list has ten organic positions. An AI answer typically names a handful of sources, and on some platforms fewer than that. The competition for a place inside the answer is tighter, and the winners are chosen by a different process.

The terminology multiplied faster than the clarity. GEO is the term from the original academic work. AEO, answer engine optimization, is used for the same idea with an emphasis on direct-answer formats. LLMO and LLM SEO appear in vendor decks. The acronyms describe overlapping practices rather than distinct disciplines, and the proliferation of labels has been useful mainly to people selling services. For a working practitioner, there are two surfaces that matter: the ranked list, and the generated answer. GEO is the work of being visible in the second.

The retrieval process underneath the two surfaces is where the real divergence lives. A classic search engine crawls, indexes, and ranks pages, then shows a list. A generative engine does something more layered. It usually starts from a search index, retrieves candidate documents, and then a language model reads, evaluates, and synthesizes them into an answer, choosing which sources to represent and which to ignore. That extra synthesis step is where a top-ranked page can be retrieved and then left out of the answer, and where a lower-ranked page with a cleaner, more extractable passage can be pulled in.

Two consequences follow directly. First, being retrievable is necessary but not sufficient; the model still has to choose you. Second, the qualities that help a language model choose a passage, such as clarity, a direct answer near the top, specific numbers, and clear attribution, are partly different from the qualities that historically moved a page up a ranked list, such as keyword targeting and link volume. The overlap between those two sets of qualities is the entire subject of this article, and it is smaller than the slogan suggests.

The Princeton study that named GEO and what it actually measured

The term generative engine optimization comes from a single academic paper. Pranjal Aggarwal and colleagues at Princeton, working with collaborators from Georgia Tech and IIT Delhi, published “GEO: Generative Engine Optimization” and presented it at KDD 2024, the ACM’s data mining conference, in Barcelona. The paper matters because most of the industry’s confident claims trace back to it, often in distorted form, and reading what it actually found corrects several myths that the slogan depends on.

The researchers built GEO-bench, a benchmark of roughly 10,000 queries across nine domains, along with the web sources needed to answer them. They then tested a set of content changes to see which ones increased a source’s visibility inside a generative engine’s answer. The headline result, repeated everywhere, is that GEO methods can lift visibility by up to 40 percent. That number is real but frequently misquoted. The average improvements across all queries were closer to 41 percent on a position-adjusted word count metric and 28 percent on a subjective impression metric, and the 40 percent figure describes the strongest methods on the queries where they worked best, not a universal guarantee.

The specific methods matter more than the headline. The strongest performers were adding statistics, citing sources, and adding quotations. Injecting relevant, specific numbers into content produced some of the largest gains. Adding attributable quotes from credible third parties helped, because models trained to recognize attribution treat quoted, sourced material as more citable. Adding citations to your own claims, moderately useful alone, became substantially stronger when combined with the others. The paper also tested fluency improvement, a purely stylistic change that added no new information, and found it produced roughly 28 percent gains on the word-count metric, which tells you that clear prose is easier for a model to parse, summarize, and attribute than dense or convoluted writing.

The finding that undercuts the “good SEO is good GEO” slogan sits quietly in the same paper. Keyword stuffing, the classic over-optimization tactic, scored about 8 percent below the untouched baseline. A method that once helped pages rank actively reduced visibility in a generative engine. The researchers put it plainly: traditional SEO tactics were largely ineffective for improving visibility inside generative responses, while GEO-specific methods worked. This is the earliest hard evidence that the two disciplines are not the same, and it came from the paper that named the field.

Two caveats keep the study honest. The generative engines of 2024, tested in a controlled benchmark, are not the AI Overviews, AI Mode, ChatGPT Search, and Perplexity of 2026, which are larger, faster-changing, and partly opaque. And the study measured visibility inside a synthesized answer under experimental conditions, not commercial outcomes in the wild. A more recent line of work, including a 2025 arXiv paper on transformer-based content optimization and an E-GEO testbed for e-commerce from Columbia and MIT researchers, has extended the idea, but the core lesson from Princeton still stands: the changes that most improve generative visibility are the ones that increase the intrinsic quality and credibility of the content, not the ones based on mechanical repetition or lexical tricks.

For a practitioner, the paper is more useful as a philosophy than a checklist. It says that models reward content that reads as authoritative, specific, and well-evidenced, and that the manipulative end of the old SEO toolkit does not transfer. That is a partial endorsement of the slogan, not a full one. Good content SEO, the kind built on genuine expertise and clear writing, carries over. Good technical SEO carries over. The gaming layer that many sites still rely on does not.

Shared machinery underneath search and generative answers

The reason the slogan is not simply wrong is that generative answers are not built on a separate internet. They are built on the same web, reached through much of the same infrastructure, and they inherit its constraints.

Start with retrieval. Google’s AI Overviews and AI Mode are produced by Gemini models sitting on top of Google’s existing search index. When an AI Overview appears, it is drawing on the same crawled, indexed corpus that produces the ranked results below it. Perplexity runs its own real-time retrieval across a very large URL pool, but that retrieval is a search step in everything but name. ChatGPT Search and the web-connected modes of other assistants also perform searches when a query calls for fresh or specific information. The generative answer is a synthesis layer bolted onto a retrieval layer, and the retrieval layer is search.

That single fact carries most of the shared machinery. A page that cannot be crawled cannot be retrieved, and a page that cannot be retrieved cannot be cited. Robots directives that block AI crawlers, rendering that hides content behind JavaScript the crawler does not execute, broken canonical tags, slow or unstable servers, and thin or duplicated pages all damage a source’s chances in both worlds. The pipeline has stages, and each stage can quietly disqualify a page: crawl access, index eligibility, retrieval for the query, extractability of a usable passage, and finally selection by the model. A failure at any stage produces the same visible outcome, which is absence, and the cause is often a plain technical problem that classic SEO already knows how to fix.

Authority is the second piece of shared machinery. Search engines spent two decades learning to treat some sources as more trustworthy than others, using links, brand signals, and behavioral data. Generative engines inherit a version of that judgment, partly because they sit on search indexes that already encode it, and partly because their training data reflects which sources the web treats as credible. A domain that the web references heavily, that is mentioned across reputable sites, and that carries strong entity signals tends to do better in both ranked results and generated answers. The mechanism is not identical, and the specific signals reweight, but the direction holds: earned authority helps everywhere.

Clarity is the third. A page organized with clear headings, consistent terminology, direct answers, and a logical structure is easier for a ranking system to understand and easier for a language model to extract from. The same document properties that help Google match a page to a query help a model pull a clean, attributable passage into an answer. This is why so many of the sources cited by AI systems share characteristics long associated with solid SEO: they are accessible, readable, well-structured, and clearly attributed.

The shared foundation is real, and it is why a business with strong SEO fundamentals starts the GEO race ahead of a business without them. The mistake is to assume the foundation is the whole building. Above crawlability, authority, and clarity sits a selection process that reweights signals, rewards different content properties, and behaves differently on every platform. The next sections measure exactly how much of the old advantage survives that selection process, and the answer is: less every quarter.

Overlap between top rankings and AI Overview citations, measured

The cleanest way to test the slogan is to measure how often the pages Google cites in an AI Overview are the same pages ranking in the organic top 10 for that query. If the overlap is high, ranking well is close to citation. If it is low, they are separate games. Several firms have run this measurement, and the trajectory across their studies tells the story better than any single number.

The early readings were high enough to justify the slogan. A late-2024 study found that around 75 percent of AI Overview citations came from pages in the top 12. Ahrefs, analyzing search rankings against AI citations in July 2025, found that 76 percent of cited pages also ranked in the top 10 for the same query, with the three URLs in an AI Overview showing a median organic ranking of 3. In that world, holding a top-three position gave you a strong chance of citation, and optimizing for one was close to optimizing for the other.

Then the readings fell, and fast. When Ahrefs re-ran the analysis with a larger dataset of 863,000 keywords and 4 million AI Overview URLs, the share of cited pages ranking in the top 10 had dropped from 76 percent to 38 percent. The remaining citations split almost evenly between pages ranking 11 to 100 and pages ranking beyond 100. A Brandlight analysis cited in a 5W research report put the collapse even more starkly, from around 70 percent to under 20 percent. Studies from other firms landed between 12 and 38 percent for AI assistants and lower AI Overview figures than the 2025 peak.

Ranking-to-citation overlap, study by study

These figures are not directly comparable because they use different query sets, different definitions of overlap, and different citation-detection methods, which is exactly why they scatter. The pattern within any single provider’s own series is the reliable signal, and within Ahrefs’s series the direction is unambiguous: the overlap roughly halved between mid-2025 and early 2026.

The contradictions are real and worth stating rather than smoothing over. BrightEdge’s 16-month study found the overlap growing, from 32 percent to 54.5 percent, as Google committed to favoring ranked content. seoClarity found extremely high overlap when it counted the top 20 rather than the top 10, reporting that around 94 percent of AI Overviews cite at least one source from the top 20 organic results. Originality.AI found that a slim majority of citations, 52 percent, came from pages that do not appear in the top 100 at all and therefore have no rank to compare.

Reconciling these requires care, and the reconciliation is the point. Two things can be true at once. AI Overviews still draw heavily on ranked content, so if you count broadly, count the top 20, and ask only whether at least one citation overlaps, the overlap looks large and even growing. But if you ask the sharper question, whether the specific page you rank number one with is the page that gets cited, the answer is increasingly no. The slogan lives in the gap between those two questions. “Ranked content gets cited” is largely true. “My top ranking guarantees my citation” is increasingly false. A business that hears the first and acts on the second will be surprised when its best-ranking pages go uncited.

The overlap numbers disagree violently, and the reasons matter

A reader looking at overlap figures ranging from under 20 percent to 94 percent could reasonably conclude the whole field is noise. It is not noise. The disagreement has specific, understandable causes, and understanding them is more useful than picking a favorite number.

The first cause is detection methodology. Parsing an AI Overview to extract exactly which URLs it cites is harder than it sounds, and the tools improved rapidly through 2025 and 2026. Ahrefs was explicit that its citation detection got better between its July 2025 study and its 2026 study, which means the two datasets are not directly comparable and some of the apparent drop in overlap reflects better detection of citations that were previously missed, rather than a pure change in Google’s behavior. When your measuring instrument changes between readings, some of the change you see is in the instrument.

The second cause is the overlap threshold. A study can count an overlap when any single cited URL appears anywhere in the top 10, or it can measure how many of the cited URLs rank, or it can ask whether the highest-ranked page is cited. These produce wildly different numbers from the same underlying data. seoClarity’s 94 percent figure counts whether at least one citation overlaps the top 20, a generous test that almost any AI Overview passes. Ahrefs’s 38 percent counts the share of cited pages that rank in the top 10, a stricter test. Both are correct; they answer different questions.

The third cause is query mix. AI Overviews behave very differently across query types. In high-stakes health and finance topics, the YMYL categories, Google leans hard on established, high-ranking, authoritative sources, and the overlap with organic rankings can exceed 75 percent. In commercial or informational queries with more diverse acceptable answers, the overlap is much lower. A study weighted toward one query type will report an overlap that a study weighted toward another will contradict, and neither is wrong about its own sample.

The fourth cause, and the most structural, is query fan-out, covered in detail in the next section. When Google splits a query into many sub-queries and cites pages that perform well across that wider cluster, the cited page may rank well for a sub-query you never tracked while ranking poorly for the exact query you measured. The overlap against the original query looks low, but the citation is still coming from ranked content, just ranked for something adjacent. This mechanism alone can explain much of the gap between “AI Overviews use ranked content” and “my ranking did not get me cited.”

Timing is the fifth cause. Google rolled out Gemini 3 as the default model for AI Overviews globally in late January 2026. Any study straddling that change is measuring two different systems. Model updates, index refreshes, and product launches all shift citation behavior, and the field moves fast enough that a six-month-old study describes a system that no longer exists in the same form.

The practical takeaway is not to distrust the data but to read it correctly. The reliable finding is the direction within any single provider’s consistent methodology, and the direction is toward weaker coupling between exact-query rankings and citations. The absolute numbers are contested; the trend is not. A practitioner should treat “my rankings and my AI citations are drifting apart” as an established fact and treat any specific overlap percentage as an estimate with wide error bars.

Query fan-out and the end of one-keyword-one-page thinking

The single mechanism doing the most to break the old equivalence between ranking and citation is query fan-out, and any serious account of GEO has to explain it because it changes what the target even is.

When a user asks a question and a generative answer is triggered, Google does not simply search the exact query. It breaks the query into multiple related sub-queries, runs searches for each, retrieves content across that wider set, and synthesizes a single answer that cites the sources it finds most useful across the whole cluster. In AI Mode, this can mean issuing hundreds of sub-searches behind one user question. The technical name is retrieval-augmented generation at scale, and the practical effect is that the answer is assembled from many small retrievals rather than one ranked list.

This wrecks the mental model that dominated SEO for twenty years: one keyword, one target page, one ranking to win. Under fan-out, the system is judging whether your content covers a topic deeply enough to satisfy a spread of related sub-questions, not whether you hold one position for one phrase. A page that ranks first for the head term but says nothing useful about the adjacent sub-questions may lose citations to a page that ranks lower for the head term but answers the whole cluster well. Ahrefs pointed to fan-out as a likely driver of the overlap collapse for exactly this reason: brands are judged less on holding one position and more on topical breadth across related queries.

Two consequences follow for anyone who wants to be cited. The first is that topical authority beats single-page optimization. A site that covers a subject comprehensively, with interlinked pages that each answer a distinct facet cleanly, gives the fan-out process more chances to pull it into an answer. A site with one heavily worked page and thin coverage everywhere else has fewer entry points. This rewards the content strategy that good SEO already favored, which is one of the genuine points of continuity between the disciplines, but it raises the bar. Depth that was optional for ranking is close to mandatory for citation.

The second consequence is that your tracked keywords no longer describe your exposure. If you monitor your rankings for a set of head terms, you are watching the wrong surface. The sub-queries that fan-out generates are often long-tail, conversational, and specific, and they may never appear in your keyword tracker. A page can be cited for a sub-question you did not know existed, or fail to be cited despite ranking well for the term you obsess over. This is why measurement in the AI era has to shift from tracking positions to tracking inclusion in answers across a set of representative prompts, a change covered later in this article.

Fan-out also explains one of the more confusing empirical findings. BrightEdge reported that most of the growth in overlap between citations and rankings was coming not from the top 10 but from positions 21 to 100, and described that band as the sweet spot, with only around 16.7 percent of citations coming from the top 10 in its data. Under a one-query model that makes no sense. Under fan-out it makes perfect sense: a page ranking 30th for the head term may rank in the top handful for a specific sub-query, and fan-out surfaces it there. Google appears to be seeking diversity within ranked content rather than repeatedly citing the same top-three pages.

For the slogan, fan-out is the decisive complication. It does not sever the link between SEO and GEO, because the cited pages are still, largely, pages that rank for something. But it moves the target from a position to a topic, and it means that good SEO in the narrow sense of winning specific rankings is a weaker predictor of citation than good SEO in the broad sense of owning a subject. The businesses that adapt fastest are the ones that already thought in topics rather than keywords. The ones still optimizing page-by-page for head terms are optimizing for a game the AI answer no longer plays.

Perplexity behaves like search, ChatGPT does not

The most important practical fact in GEO is that “AI search” is not one thing, and the clearest way to see it is to compare how Perplexity and ChatGPT choose sources. They sit at opposite ends of the spectrum, and the slogan is true for one and false for the other.

Perplexity is the most SEO-aligned major AI system. It was built to cite, retrieves in real time, and consistently favors content that ranks well in Google. Ahrefs’s study of 15,000 prompts found that nearly one in three of Perplexity’s citations, around 28.6 percent, pointed to pages ranking in Google’s top 10 for the target query, several times higher than the other assistants. A Semrush and SE Ranking analysis put the alignment even higher, finding Perplexity citing top-10 pages around 91 percent of the time. Perplexity also cites densely, averaging somewhere between 16 and 22 sources per answer depending on the study, far more than its rivals. For Perplexity, the slogan mostly holds: rank well in Google, structure your content for clean extraction, and you have a strong chance of being cited.

ChatGPT is the opposite. The same Ahrefs work found only around 8 percent of ChatGPT’s citations overlapping Google’s top 10, and a Semrush and SE Ranking study put the overlap near 14 percent. ChatGPT constructs answers far more from the parametric knowledge baked into its training data than from live web retrieval, and when it does retrieve, it is selective, citing fewer sources per answer, around 7 on average in one comparison. It leans heavily on a narrow set of trusted sources; Wikipedia alone accounts for a large share of its citations, with figures reported from around 8 percent up to nearly 48 percent of top-cited sources depending on the study and period. A striking finding from Ahrefs is that roughly 67 percent of the top 1,000 pages ChatGPT cites are in practice off-limits to brand SEO, sites like Wikipedia, government and educational domains, and major news media, and that around 28 percent of ChatGPT’s most-cited pages have zero Google organic visibility.

For ChatGPT, the slogan largely fails. Ranking well in Google does little to earn a ChatGPT citation, because ChatGPT is not reading the ranked list. What earns ChatGPT visibility is entity strength built over time: frequent, consistent mention of your brand across authoritative third-party sources, co-occurrence with your category terms, and presence in the encyclopedic and editorial sources ChatGPT trusts. Ahrefs found that brand web mentions correlate with AI citation at roughly 0.664, around three times stronger than backlinks at 0.218, and the effect is strongest for exactly this kind of parametric, entity-driven system.

Gemini and Copilot sit in between and closer to ChatGPT in the overlap studies, citing top-10 pages only single-digit percentages of the time in the 15,000-prompt analysis. Google’s own AI Overviews and AI Mode, being built directly on the search index, are the most ranking-correlated of the Google-branded surfaces, but even there the correlation has weakened as described above.

The operational lesson is blunt. A single GEO strategy cannot win every engine, because the engines do not share a source-selection process. Optimizing for Perplexity is close to optimizing for Google. Optimizing for ChatGPT is close to public relations and brand-building. A team that treats these as one task will over-invest in the tactics that work for one platform and stay invisible on another. The platform you are not tracking is usually the one where your biggest gap sits.

Cross-engine citation overlap sits near eleven percent

If the same content behaved similarly across AI engines, a single strategy would be defensible and the slogan could survive by treating “AI search” as one target. The data destroys that assumption. When researchers compare which domains different engines actually cite for similar prompts, the overlap is tiny.

The recurring figure is around 11 percent. An analysis of 680 million citations found that only about 11 percent of domains were cited by both ChatGPT and Perplexity. A separate study of 118,000 AI responses across ChatGPT, Perplexity, Google AI Mode, and Claude found the same: only about 11 percent of cited domains appeared across multiple platforms. A three-engine study from Passionfruit landed at around 12 percent. Even the two Google-owned surfaces disagree with each other; Google AI Overviews and Google AI Mode were found to cite the same URLs only about 13.7 percent of the time, despite reaching similar conclusions.

The magnitude of the divergence is easy to underestimate. One cross-platform analysis documented citation-volume variance of up to 615 times for the same brand between platforms. A company that dominates Perplexity’s citation pool can be almost absent from ChatGPT, and the reverse happens too. A study of 300,000-plus citations across six B2B SaaS brands, tracking identical content across ChatGPT, Perplexity, Gemini, Claude, AI Overviews, and AI Mode for 90 days, found the per-platform profiles so different that the same brand looked like six different companies. Claude gave brands the highest owned-citation share in that study at around 9 percent, Perplexity gave around 7 percent, and ChatGPT was consistently the weakest for brand visibility across every company studied.

This is the strongest single argument against “good SEO is good GEO” as a general rule. Even if good SEO reliably produced good visibility on one engine, that visibility would not transfer to the others, because the engines build answers from fundamentally different source pools. Treating AI visibility as one number is, as one analyst put it, like measuring your Google ranking and assuming it applies to Bing, except the gap between AI engines is far wider than the gap between the two classic search engines ever was.

The differences are not random; they follow each engine’s architecture. Perplexity is retrieval-first and cites densely from a pool that leans toward community sources like Reddit and, increasingly, YouTube, with a notable share of academic and institutional domains. ChatGPT is training-data-heavy and cites a narrow set of encyclopedic and editorial sources. Google’s surfaces inherit the search index and lean on ranked, authority-weighted content plus heavy use of YouTube. Each architecture produces a distinct citation fingerprint, and the fingerprints barely overlap.

The reason cross-engine overlap matters so much for strategy is that it converts GEO from a single optimization problem into a portfolio problem. A brand needs a core authority hub on its own site, a supporting layer of original evidence, and a third-party validation layer of mentions and reviews, and it needs to know which engines its buyers actually use before deciding where to concentrate. A B2B software company whose buyers live in ChatGPT should invest in entity-building and third-party presence. A publisher whose audience uses Perplexity should invest in freshness and clean, rankable, extractable pages. The same budget spent against the wrong engine buys almost nothing.

There is one consolation in the data. The Ahrefs 75,000-brand study found that while the specific cited URLs diverge, the brands that all three AI surfaces tend to mention overlap heavily, with an output-overlap correlation of around 0.779. The engines disagree about which pages to cite but largely agree about which brands are worth mentioning. That points to the deeper truth of GEO, developed in the next section: at the brand level, the game is authority, and authority is built off-site in ways that classic on-page SEO never fully captured.

Brand mentions beat backlinks by roughly three to one

For twenty years, the currency of off-site SEO was the backlink. Links were how PageRank measured authority, how sites voted for each other, and how agencies justified expensive link-building programs. In AI search, that currency has been partly replaced, and the replacement is the unlinked brand mention.

The evidence is unusually clear for a field this noisy. Ahrefs analyzed 75,000 brands to find which signals best predict inclusion in Google’s AI Overviews. Branded web mentions came out on top with a correlation of 0.664. Traditional backlinks correlated at just 0.218. That is close to a three-to-one gap, and it is a structural finding rather than a rounding artifact. Brand anchors correlated at 0.527 and branded search volume at 0.392. The top three signals were all off-site brand signals, and all of them beat link metrics. Brands in the top quartile for web mentions earned up to ten times more AI Overview mentions than the next quartile down.

The distinction between a link and a mention is the whole point. A backlink tells a system where to go. A brand mention, even without a hyperlink, tells a system what the web trusts and talks about. For a language model absorbing the entire text of the web during training, an unlinked mention of your brand in a credible article is a signal it can learn from directly, because the model reads the sentence, not the link graph. In the context of AI citation, being talked about matters more than being linked to. This inverts one of the oldest priorities in SEO, where an unlinked mention was often dismissed as wasted PR.

A related finding lands even harder. The number of pages on a site showed almost no relationship with AI visibility, correlating at around 0.194 in the same study. Publishing volume, the tactic behind a decade of content-mill SEO, does not build AI visibility. Earned presence does. A thin brand that publishes constantly stays invisible in AI answers, while an authoritative brand that publishes rarely but is discussed widely gets mentioned. Seer Interactive’s separate work confirmed the pattern, finding weak correlations for domain rank and backlinks and stronger relationships with brand-level signals when predicting ChatGPT brand visibility.

The platforms weight these signals differently, which matters for where to invest. Google’s AI Mode showed the highest correlation with classic brand-authority signals, with branded anchors around 0.628 and branded search volume around 0.466, which makes AI Mode harder for young brands because it rewards the demand and market presence that take years to build. ChatGPT showed weaker correlations with several traditional metrics, including domain rating around 0.27, which paradoxically means a smaller brand can sometimes earn ChatGPT visibility by becoming the clearest, best-described entity for a narrow problem before it has an incumbent’s search demand.

The obligatory caveat from the researchers is that correlation is not causation. High-visibility brands also tend to have high cross-platform presence; opening a YouTube channel does not mechanically trigger citations. The underlying driver is brand authority as a composite signal that training data and retrieval both absorb. But the practical instruction is hard to argue with. A dollar spent earning genuine brand mentions across credible sources buys more AI visibility than a dollar spent acquiring backlinks. This is where GEO diverges most sharply from the link-obsessed version of SEO, and where the slogan is most misleading, because a business with excellent backlink-driven rankings but a weak off-site brand footprint can rank well and still lose the AI-citation race to a more-talked-about competitor.

The rise of YouTube as an AI visibility signal

The most surprising finding in the 2026 brand-visibility research is that the strongest single predictor of AI visibility is not a website signal at all. It is YouTube.

When Ahrefs expanded its study to cover ChatGPT, Google AI Mode, and AI Overviews across 75,000 brands, YouTube mentions correlated with AI visibility at roughly 0.737, the highest figure in the entire dataset, beating even branded web mentions, which had previously topped the list at 0.664. YouTube mention impressions, the same mentions weighted by video views, correlated almost as strongly at around 0.717. A YouTube mention here means a brand name appearing in a video title, description, or transcript. The relationship held across every platform tested, including ChatGPT, which is owned by OpenAI rather than Google.

The mechanism becomes obvious once the training pipeline is visible. YouTube is one of the most crawled, most consistently formatted, and most heavily trained-on corpora on the web. Both Google and OpenAI have trained models on YouTube transcripts; reporting indicated OpenAI trained GPT-4 on over a million hours of YouTube transcription. When a brand is discussed in a well-watched video in its niche, that discussion enters the training data and the retrieval pool simultaneously, teaching the model to treat the brand as a known entity. Google’s own AI surfaces cite YouTube more than almost any other domain, and YouTube has emerged as one of the most-cited sources in AI Overviews and a rising share of Perplexity’s citations after its Reddit citations dropped following a legal dispute in late 2025.

For the slogan, YouTube is another crack. A brand can have flawless on-site SEO and rank well across its category, and still be under-represented in AI answers because it has no presence on the video platform that the models weight most heavily. Video was treated for years as a distribution channel that sat outside the SEO discipline. In the AI era it has moved to the center of visibility, and a brand that ignores it is leaving the strongest measured signal on the table.

The practical response is not to chase view counts. The correlation data suggests that the volume of mentions matters slightly more than their reach, which points toward consistent, category-relevant presence rather than one viral hit. A software company should aim to be named and explained in tutorials, reviews, demos, and expert commentary across many videos, not to produce a single polished ad. The goal is to become a brand that gets talked about on video by other people, in the same way that earned brand mentions across text sources build authority. This is closer to public relations and community presence than to keyword targeting, and it sits well outside the traditional SEO toolkit.

A sensible caution applies. YouTube’s dominance in these correlations partly reflects that AI systems trained heavily on its transcripts, and if the training and citation weighting shift, so will the signal’s value. The finding is a strong operating clue, not a permanent law. But for now, any GEO audit that ignores a brand’s YouTube presence is missing the single strongest measured correlate of AI visibility, and any team building a GEO plan should at least measure it before deciding whether the category warrants investment. For businesses whose buyers research through demos, reviews, and expert videos, it is no longer optional.

The parts of SEO that transfer cleanly to GEO

Having spent several sections on where the slogan breaks, it is worth being precise about where it holds, because a large part of good SEO does transfer, and pretending otherwise would be as wrong as the slogan itself. The continuity is real, and it is why a business with strong fundamentals is not starting from zero.

Technical crawlability transfers completely. Every generative engine relies on retrieval, and retrieval requires access. Clean HTML, fast and stable servers, sensible internal linking, correct canonical tags, working sitemaps, and no accidental blocks in robots directives all matter in both worlds, for the same reason: a page that cannot be fetched and parsed cannot be cited. The technical SEO checklist that keeps a site indexable keeps it eligible for AI answers. This is the least glamorous part of SEO and the part that transfers most cleanly.

Content quality and topical depth transfer. The Princeton study found that clearer, more fluent, better-evidenced content earned more visibility in generative engines, and the fan-out mechanism rewards sites that cover a subject comprehensively. Good content SEO, built on genuine expertise and thorough coverage rather than keyword density, is close to good GEO. A site that already owns a topic through deep, interlinked, well-written content is well positioned to be pulled into answers across the sub-queries that fan-out generates.

Authority and trust signals transfer, with reweighting. Search engines and generative engines both reward credibility, and the E-E-A-T principles that Google formalized, experience, expertise, authoritativeness, and trust, map onto what makes a source citable in an answer. The specific signals reweight, with brand mentions rising and raw backlinks falling, but the underlying goal of being a source the system trusts is shared. A brand that built genuine authority for SEO carries most of that authority into AI search.

Structured, extractable content transfers and becomes more important. Clear headings, direct answers, comparison tables, definition-first sentences, and FAQ blocks help ranking systems understand a page and help language models extract a clean passage. The formatting discipline that good SEO already encouraged is close to mandatory for GEO, because a model that cannot find a self-contained, quotable passage will pass over the page even if it ranks.

Structured data and schema transfer. JSON-LD markup that identifies articles, authors, organizations, and their relationships helps machines understand what a page is and who stands behind it, which supports more accurate citation. Unlike some newer tactics, schema has demonstrated impact, and it is a low-cost investment that serves both ranked results and AI answers.

Search intent understanding transfers. Knowing what a user actually wants from a query, and building content that answers it directly, matters as much for earning a citation as it did for earning a click. The fan-out era rewards content that anticipates the cluster of sub-questions around an intent, which is an extension of intent-driven SEO rather than a departure from it.

The honest summary is that the fundamentals transfer and the gaming does not. The parts of SEO built on making a site genuinely accessible, credible, thorough, and clear carry into AI search almost unchanged. This is a large inheritance, and it is why the slogan is not simply false. A business that did this work is genuinely ahead. The trouble starts with the parts of SEO that were about manipulating a ranking rather than serving a reader, and with the new off-site and platform-specific work that the fundamentals do not cover, which the next section addresses.

Tactics that fail to transfer, and a few that backfire

The other half of the honest answer is that a real slice of the SEO toolkit does nothing for AI visibility, and a few tactics that once helped rankings now hurt citations. A business that assumes its whole playbook carries over will keep spending on work that no longer pays.

Keyword stuffing backfires. The Princeton study measured this directly: loading content with repeated query terms scored around 8 percent below the untouched baseline in generative engines. Google’s AI Mode has been described as penalizing over-optimization, and the operative metric is information density, not keyword density. A page that reads as manipulated is less likely to be extracted and cited. The tactic that once nudged rankings now actively reduces the chance of appearing in an answer.

Exact-match, single-keyword targeting delivers less. Under query fan-out, building a page to win one head term while ignoring the surrounding topic leaves most of the citation opportunity uncaptured. The page may rank for the term and still be passed over in favor of a page that answers the whole cluster. Effort spent perfecting one page for one phrase is worth less than effort spent building topical coverage, and a team still working keyword-by-keyword is optimizing for a target the answer no longer uses.

Raw backlink volume delivers less. With brand mentions correlating three times more strongly than backlinks for AI visibility, a link-building program judged purely on link count is misallocated for GEO. Links still help, and they still support classic rankings, but a program that acquires links while generating no genuine brand conversation buys little AI visibility. The money moves better toward earning mentions.

Thin, high-volume content publishing backfires twice. Page count showed almost no correlation with AI visibility, and thin content risks E-E-A-T problems in classic search as well. AI-generated content pushed out at scale without human expertise, original data, or a real point of view is exactly what models are trained to recognize and discount, and it will not be cited. The content-mill strategy that some sites still run produces volume that neither ranks nor gets cited.

Manipulative technical tricks fail. Cloaking, doorway pages, and other schemes aimed at gaming a crawler add nothing in a system where a language model reads the actual content and a human-legible passage has to survive extraction. There is no ranking to trick when the outcome is inclusion in a written answer.

The most cautionary example of a tactic that failed to transfer is one the industry invented specifically for AI: the special file at a site’s root meant to guide AI systems, covered in the next section. It is worth flagging here because it belongs in the category of work that feels like GEO and produces nothing. The pattern to watch for is any tactic that targets the machine rather than the reader and the wider web. Machine-targeting worked in the era of ranked lists because rankings were a machine’s judgment that could be nudged. In the era of generated answers, the machine reads the content the way a person would and weighs the brand the way the web does, so the tactics that survive are the ones that improve genuine quality and genuine reputation.

There is a subtler failure mode worth naming. Some SEO teams responded to AI search by adding a layer of AI-specific technical artifacts, custom files, AI sitemaps, special robots blocks, on top of otherwise sound sites, and billed the work as GEO. Audits of such sites have found the traffic impact to be zero, because the files addressed a problem that did not exist. The tactics that fail to transfer are not just neutral; when they consume the budget that should have gone to content depth, brand mentions, and technical fundamentals, they carry a real opportunity cost. Recognizing them is as important as knowing what works.

The llms.txt distraction and the fundamentals it obscured

No single topic captures the gap between GEO folklore and GEO evidence better than llms.txt, and the story is worth telling because it is a near-perfect case study in how the industry manufactures work that feels like optimization and changes nothing.

The file itself is simple. An llms.txt file is a markdown document placed at a site’s root, listing the site’s most useful pages so that an AI system could, in principle, find them efficiently. It was proposed in September 2024 by Jeremy Howard, an AI researcher, to solve a real and narrow problem: language model context windows are too small to ingest a full website, and converting cluttered HTML into clean text is wasteful, so a curated markdown index would help. The proposal had nothing to do with search visibility. It was infrastructure for developer documentation and agents.

The SEO industry repackaged it as a visibility play, and the evidence never supported that framing. The major AI search crawlers overwhelmingly ignore the file. One firm analyzed over 500 million AI bot traffic events and found that GPTBot, ClaudeBot, PerplexityBot, and Google-Extended almost never fetch llms.txt, crawling standard HTML instead; only a statistically negligible fraction of requests touched it. A controlled test over 90 days and 60,000 bot visits found around 0.1 percent touched the file. A six-month log study across 57 bots and 180,000 requests found a clean zero from the major AI crawlers.

The platforms have said so on the record. In July 2025, Google’s Gary Illyes confirmed Google does not support llms.txt and is not planning to, and John Mueller compared it to the long-discredited keywords meta tag. Google’s published guidance now states plainly that Search ignores llms.txt and it neither helps nor harms rankings. As of mid-2026, no major AI vendor, including OpenAI, Google, Anthropic, Meta, and Mistral, had publicly committed to using the file in a production answer system. Independent studies from SE Ranking and others found no link between having the file and being cited more often. Adoption sits at roughly 10 percent of sites, spread evenly across large and small domains rather than concentrated among leaders, which is what folklore-driven adoption looks like.

There are honest counterpoints, and they matter. Google does crawl and index llms.txt files at scale; a Wix analysis found tens of thousands and later over 100,000 such files indexed, and a small share even ranked for keywords, which corrects the flat claim that Google ignores the file entirely. But crawling a file is not the same as using it to shape an answer, and the firm making that argument sells the feature it defends. The genuinely valid use case is developer tooling: AI coding assistants like Cursor, GitHub Copilot, and Claude Code do read llms.txt to fetch documentation efficiently, and for documentation-heavy sites it is worthwhile as agent infrastructure. That is a business-to-agent play, not an AI-search-visibility play.

The episode matters beyond one file because of what it reveals. Google published its first official guidance on optimizing for generative AI features in May 2026, and added a dedicated llms.txt subsection in June 2026 after community questions, both of which pointed back to the same message: the brands that win in AI search win through the fundamentals, not a text file at the domain root. Genuine topical authority, consistent mentions across high-quality sources, structured content that answers questions directly, and strong entity signals do the work. The llms.txt saga consumed a year of industry attention and produced, for search visibility, nothing. The lesson for anyone evaluating a GEO tactic is to ask for the crawler logs and the controlled study before spending, because the field is full of confident advice that the data does not support.

Google’s stance that GEO cannot exist without SEO fundamentals

Google has taken a public position on the slogan, and it is worth quoting accurately because it is both the strongest support for the “good SEO is good GEO” idea and, read carefully, a narrower claim than the slogan makes.

At Google Search Live in Zurich in late 2025, John Mueller put it directly: AI systems rely on search, and there is no such thing as GEO or AEO without doing SEO fundamentals. The message, repeated across Google’s communications since, is that AI-powered search experiences depend on the same basics that search engines have relied on for years, and that AI does not replace SEO so much as build on it. Google reinforced this in May 2026 by publishing its first official guide to optimizing for generative AI features, filed under a new generative-AI fundamentals section of its documentation, which framed AI visibility as an extension of established practice rather than a new discipline.

Google’s position is credible on its own terms, and it aligns with the shared-machinery argument made earlier. A site that fails the fundamentals fails everywhere. If a page cannot be crawled, is not indexed, carries no authority, and answers nothing clearly, it will neither rank nor be cited. In that sense there is no GEO without SEO, because the retrieval layer under every generative answer is a search layer, and the entry requirements are the same. A business that treats GEO as a reason to abandon SEO fundamentals is making a serious mistake, and Google is right to warn against it.

The interpretation to resist is the stronger one that marketers often hear. “No GEO without SEO fundamentals” does not mean “good SEO is sufficient for good GEO.” It means the fundamentals are necessary, not that they are enough. Google has an obvious institutional interest in this framing, because it wants site owners to keep doing the work that keeps Google’s index healthy and to keep viewing Google’s ecosystem as the center of visibility. That interest does not make the claim false, but it does mean the claim should be read for exactly what it says. Google is describing a floor, the minimum required to be eligible, not a ceiling, the work required to actually win citations.

The data reviewed in this article fills in what Google’s framing leaves out. Above the shared fundamentals sit the divergences: the collapse in exact-query ranking-to-citation overlap, the near-total lack of cross-engine agreement, the dominance of off-site brand mentions and YouTube presence over on-page and link signals, and the platform-specific behavior that makes ChatGPT and Perplexity almost opposite targets. None of these contradict Google’s floor. All of them describe the building that sits on top of it, which Google’s fundamentals-first message does not address.

There is also a jurisdictional point in Google’s message. Google speaks for its own surfaces, AI Overviews and AI Mode, which are the most search-correlated of all the AI systems because they are built directly on the search index. For those surfaces, “no GEO without SEO fundamentals” is close to a complete description, and the slogan is closer to true. For ChatGPT, which reads the training data more than the live index and cites Wikipedia and editorial sources over ranked pages, Google’s framing simply does not apply, because Google does not run it. The slogan’s truth value depends on which engine you mean, and Google is, understandably, describing the engine it controls.

The synthesis is that Google is right about the floor and quiet about the ceiling, which is exactly what a platform owner would say. A practitioner should take the fundamentals message seriously, keep doing sound SEO, and then add the off-site, brand-level, platform-specific, and video work that the fundamentals do not cover. Doing the first without the second produces eligibility without citations, which is the precise failure mode the slogan encourages.

Structured content, extractability, and the first thirty percent

If there is one on-page discipline that pays off across nearly every AI engine, it is writing content that a model can extract cleanly. This is where GEO gives concrete, testable instructions, and where good SEO habits translate almost directly if they were built around readers rather than crawlers.

The core finding is about position within a page. Research indicates that around 44 percent of all LLM citations come from the first 30 percent of a page’s content. Models front-load their attention, and a page that buries its answer beneath a long preamble gives the model nothing to extract early. The instruction that follows is concrete: lead with a direct, self-contained answer in the first 80 words of a page or a section, before context and nuance. This is the passage an AI system is most likely to lift. The old SEO habit of a slow, keyword-laden introduction is now a liability, because it delays the extractable answer past the zone the model reads most closely.

Structure at the section level matters as much as position. AI systems preferentially extract from content organized into clear, self-contained blocks: a question as a heading, a direct answer in the first sentence or two, then supporting detail. Comparison tables, step-by-step lists, definition-first sentences, and FAQ blocks are all formats that models pull from readily, because each is a compact, complete unit that can be quoted without surrounding context. One practical guideline circulating in the field is to write self-contained sections of roughly 75 to 225 words that each answer one question completely. A page built this way is easier to rank and far easier to cite.

Length is not the lever many assume. Ahrefs found near-zero correlation between content length and citation probability, measured at around 0.04, and reported that more than half of AI Overview citations, around 53 percent, went to pages under 1,000 words. Concision and directness serve citation better than word count. This does not contradict the value of topical depth across a site; it means each individual answer should be tight even as the site’s coverage is broad. Depth comes from covering many questions well, not from padding any single page.

Terminology consistency helps models form clear associations. Using the same name for the same concept throughout a site, rather than varying it for stylistic reasons, helps a model understand what a page is about and match it to a query. This connects to entity clarity, covered later, and it is a discipline that classic SEO encouraged for keyword consistency and that GEO rewards for a related reason: machine comprehension.

Structured data reinforces all of this. Marking up articles with author, publication date, and publisher, marking up organizations with links to their verified profiles, and using breadcrumb and FAQ schema where appropriate all help a retrieval-augmented system understand and trust a page. Schema has demonstrated impact on featured-snippet and AI-mention rates in field tests, and it is inexpensive to implement.

The reason this discipline transfers so well is that it was always about serving the reader, and a language model reading a page behaves, for these purposes, like an impatient reader who wants the answer first and the evidence attached. A team that already writes clear, well-structured, answer-first content for humans is most of the way to writing extractable content for models. The teams that struggle are the ones whose content was built to satisfy a keyword algorithm rather than a reader, because that content tends to bury its answers and vary its terminology in ways that both readers and models find hard to use.

Freshness, decay, and the speed of AI citation pools

One property separates AI citation from classic ranking more than almost any other: speed. Content enters and leaves AI answer pools far faster than it climbs or falls in search rankings, and this changes both the tactics and the maintenance burden.

The entry speed is the first surprise. New content can enter AI citation pools within three to five days, according to analysis cited in the 5W report, while ranking well in Google typically takes three to six months. A generative engine that retrieves in real time, like Perplexity, can cite a page almost as soon as it is published and indexed, with no need to wait out the slow climb up a ranked list. For time-sensitive topics, this is an advantage that classic SEO never offered, and it rewards publishers who can produce credible, well-structured content quickly.

The decay is the second surprise. The same analysis found that content shows measurable decline in AI citation frequency after roughly 13 weeks without a refresh. A page that earns citations can quietly lose them within a quarter simply by aging, even if nothing about its quality changed, because the engines favor recent material and newer competitors enter the pool. Studies of citation patterns found that a large majority of AI citations went to content published within the last two years, with a substantial share from the most recent year, and Perplexity in particular rewards very recent content, with reports of content under 30 days old earning several times more citations.

This freshness bias reshapes the maintenance model. Under classic SEO, a well-ranked evergreen page could sit untouched for years and keep earning traffic. Under GEO, content is closer to a subscription than an asset: it keeps earning citations only while it is kept current. A practical response is to build a refresh cadence for the pages that matter most for AI visibility, updating data, dates, and claims on a schedule rather than treating publication as the end of the work. The pages worth refreshing are the ones targeting the prompts a business most wants to win, identified through monitoring rather than guesswork.

The freshness effect also interacts with the brand-authority findings in a useful way. A brand with strong entity signals and wide mention coverage can weather the freshness bias better than a thin site, because the model’s underlying trust in the brand persists between refreshes. Freshness helps a page get retrieved and considered; authority helps it get selected. The two work together, and a strategy that chases freshness while neglecting authority produces content that is retrieved and then passed over, while a strategy that builds authority while neglecting freshness produces a trusted brand whose specific pages age out of the citation pool.

There is a caution against over-reacting. Freshness is one signal among many, and it does not override quality, structure, or authority. Republishing a page with a new date and no real update is the kind of manipulation that models are increasingly able to discount, and it risks the same credibility problems as any other trick. The useful version of a freshness strategy is genuine maintenance: revisiting content to keep it accurate, adding new evidence as it emerges, and retiring or consolidating pages that no longer serve a purpose. Freshness rewards real upkeep, not cosmetic date changes, and it is one more area where the honest version of the tactic works and the gamed version fails.

Zero-click search and the economics of being cited without a click

The hardest part of the transition from SEO to GEO is not technical. It is economic. The entire business case for SEO rested on the click, and the click is disappearing, which forces a rethink of why a business would want AI visibility at all when it often produces no traffic.

The zero-click trend predates AI but has accelerated sharply. SparkToro, using Similarweb clickstream data, found that in the first four months of 2026, around 68 percent of Google searches ended without a click to any external site, up from around 60 percent in 2024, the fastest acceleration of the phenomenon in a decade. When an AI Overview is present, the zero-click rate rises to around 80 to 83 percent. In Google’s AI Mode, roughly 93 percent of searches end without an outbound click, because the interface is designed to answer in place. For a business built on informational traffic, this is not a headwind; it is a structural change in whether search sends traffic at all.

The click-through numbers underneath are just as stark. Seer Interactive’s study, tracking 53 brands across 5.47 million queries and 2.43 billion impressions, found that organic click-through rate on queries with an AI Overview fell from a pre-AIO baseline of 1.76 percent to 0.61 percent by mid-2025, a 61 percent decline, before bottoming at 1.3 percent in December 2025 and partially recovering to 2.4 percent by February 2026. Ahrefs, analyzing 300,000 keywords, found the position-one click-through rate dropping 58 percent when an AI Overview is present, nearly double the 34.5 percent suppression the same team measured eight months earlier. A page can hold position one and lose most of its traffic, because the answer sits above it.

The economics of citation versus the click

This is the economic core of the GEO case, and it explains why being cited matters even when clicks fall. Being the cited source is now the prize that a ranking used to be. Seer found that brands cited inside an AI Overview earn around 35 percent more organic clicks than uncited brands on the same query, and around 120 percent more clicks per impression, along with 91 percent more paid clicks. The click did not vanish entirely; it concentrated on the cited sources. In a zero-click world, citation is the new ranking, and the gap between cited and uncited is the gap that used to separate page one from page two.

The strategic reframing is uncomfortable but clear. A business can no longer justify content purely by the clicks it earns, because on informational queries most of those clicks are gone regardless of how well the content ranks. The justifications that survive are being the cited source that captures the shrinking pool of clicks, building brand recognition through appearing in answers even without a click, and supporting transactional pages through internal linking and authority rather than direct traffic. The informational blog post as a pure traffic machine is a declining model. The informational content that survives does so as an authority-builder and a citation target, and it has to be measured that way, which the section on measurement addresses.

Traffic collapse, CTR suppression, and the recovery curve

The traffic story deserves its own treatment because the headline numbers are frightening and the underlying picture is more workable than the panic suggests. Understanding the shape of the collapse, and the recovery inside it, is what separates a useful response from a despairing one.

The collapse is real and broad. Gartner projected a 25 percent decline in traditional search volume through 2026 and organic search traffic to websites falling by 50 percent or more by 2028 as generative AI search scales. News publishers have projected losses on the order of 43 percent of organic traffic by 2029. Individual cases are brutal: large publishers have reported traffic declines of 70 to 80 percent and click-through drops approaching 89 percent on affected queries. The pattern that alarms teams most is that rankings can hold steady while traffic falls off a cliff, because the loss is driven by the AI answer sitting above the result, not by a drop in position. The classic diagnostic, stable or rising impressions in Search Console combined with collapsing clicks on informational queries, is now the signature of an AI-driven loss rather than a ranking problem.

The recovery inside the collapse is the part that gets less attention and matters more for planning. Seer Interactive’s data showed the AIO click-through rate rebounding sharply from its December 2025 floor of 1.3 percent to 2.4 percent by February 2026, an 85 percent recovery, which means the realistic scenario is not permanent decline but a new, lower structural baseline that pages can compete inside. The competition inside that baseline is won by citation: cited pages earn roughly 120 percent more clicks per impression than uncited ones even though both trail the pre-AIO world. The strategic question is no longer whether traffic fell but whether a brand’s priority pages sit in the cited cohort or the uncited one.

There is also a measurement mirage inside the collapse worth naming. Some of the apparent traffic loss is attribution failure rather than lost visits. When an AI Mode side panel or an AI Overview citation sends a user to a page, analytics tools often bucket that visit as direct or unattributed rather than as search referral, so the search channel looks worse than the actual traffic. One clickstream analysis even found the raw US zero-click share ticking down slightly in early 2026, from 24.5 percent to 22.4 percent, as AI Overview citations generated clicks that the old attribution model failed to credit to search. The traffic still flowed; the attribution model for it broke. This does not undo the real losses on informational queries, but it means part of the reported decline is a reporting problem, and teams should separate genuine loss from attribution loss before concluding their content failed.

Query type determines exposure, and this is the most actionable part of the traffic picture. Informational queries, the “what is” and “how to” content that filled the top of many funnels, are the hardest hit; nearly all informational keywords now trigger an AI Overview, and those overviews answer the question in place. Transactional and navigational queries are far less affected. “Buy,” “best X for Y,” and “price of X” queries trigger AI Overviews far less often, and they sit closer to conversion. A business whose content skewed informational faces the steepest losses and the clearest instruction: shift weight toward comparison content, product-level pages, case studies with real numbers, and transactional queries that AI is less able to answer and less inclined to intercept.

The recovery posture that follows is neither denial nor surrender. It accepts that informational traffic will not return to its old level, works to become the cited source that captures the concentrated clicks that remain, fixes attribution so the real picture is visible, and rebalances the content mix toward the query types that still send traffic and sit near revenue. Teams that internalized this in 2026 stopped mourning the old CTR and started competing for citations. Teams that kept measuring total organic sessions and waiting for a rebound kept watching a number that is not coming back.

AI referral traffic converts far better than it looks

The counterweight to the traffic collapse is a finding that surprises almost everyone the first time they see it: the small amount of traffic that AI systems do send converts dramatically better than traditional organic traffic, which changes the value calculation even at low volume.

The volume is genuinely small. Ahrefs found that visitors from AI search platforms accounted for around 0.5 percent of total traffic in its data, and Google still sends on the order of 190 times more traffic to websites than ChatGPT does. Anyone expecting AI referrals to replace lost organic volume in the near term is going to be disappointed; the raw numbers are not close. AI search drives well under 1 percent of referral traffic for most sites.

The conversion rate is where the picture inverts. Ahrefs found that visitors from AI search generated around 12.1 percent of signups despite being only 0.5 percent of traffic. Seer Interactive and Semrush data put AI referral conversion at roughly 4.4 to 9 times the rate of traditional organic traffic. Other analyses reported AI search visitors converting at multiples of organic, with figures for individual platforms ranging from around 10 to 17 percent conversion versus organic’s 1.76 percent. Perplexity visitors in particular have been reported converting at roughly 11 times the rate of standard organic traffic. A visitor who arrives from an AI answer has usually already read a summary, evaluated options, and self-qualified, so the click that survives is a high-intent one.

The reason is straightforward once the funnel is visible. In the old model, a search user clicked early and did their evaluation on your site, which meant a lot of low-intent traffic that mostly bounced. In the AI model, the evaluation happens inside the answer, and the user only clicks through when they are ready to go deeper or act. The AI system does the top-of-funnel filtering for you, and it sends you the users who made it through. Raw click volume becomes a misleading metric, because the surviving clicks are worth far more each than the clicks that disappeared.

This reframes the value of citation for a business that sells something. Being cited in an AI answer does two useful things even when it sends no click: it puts a brand in front of a buyer at the exact moment of research, building recognition that carries into the next search and the eventual purchase, and when it does send a click, that click converts several times better than the organic traffic it replaced. A B2B buyer who sees a brand recommended by an AI engine carries that signal into their next demo request and vendor shortlist, an effect that shows up as lifted branded search and paid performance rather than as direct AI referral traffic.

The honest limit is that high conversion on tiny volume is still tiny absolute numbers today, and a business cannot yet fund itself on AI referrals alone. The finding matters as a trajectory and a valuation lens, not a present-day traffic source. As AI search share grows, the high-quality end of search traffic will increasingly arrive through AI answers, and the brands that established citation presence early will be positioned to capture it. The right way to hold both facts at once is that AI referral is a small, high-value channel now and a growing one later, and its value is badly understated by any metric that counts clicks rather than outcomes. A team that dismisses AI visibility because the referral traffic is small is measuring the wrong thing, in the same way a team that panics about the traffic collapse is measuring the wrong thing. The volume is small; the quality is exceptional; the direction is up.

Sector by sector, where the SEO-to-GEO bridge holds

The slogan holds to different degrees in different industries, and averaging across all of them hides the variation that actually determines a business’s exposure. Five sectors show the range clearly, and a business should locate itself among them rather than reason from the aggregate.

B2B software and SaaS is where the divergence between SEO and GEO is widest and the stakes are highest. B2B buyers have moved research into AI tools early; surveys put weekly ChatGPT use among B2B researchers at around 68 percent, and B2B tech queries now trigger AI Overviews around 82 percent of the time, up from around 36 percent a year earlier. Because ChatGPT is the platform B2B teams care most about and the one where brand visibility is consistently weakest, the SEO-to-GEO bridge is shakiest here. A SaaS company can rank well and be nearly invisible in ChatGPT, and the fix is entity-building, third-party mentions, review-platform presence, and co-occurrence with category terms rather than more ranking work. The high conversion rate of AI referrals makes this the sector where citation is most worth chasing despite low traffic volume.

E-commerce sits at the opposite end for informational risk and a different end for opportunity. Product and transactional queries trigger AI Overviews far less often than informational ones, with e-commerce AI Overview rates reported in the low single digits for buying queries, which means the traffic collapse hits product pages less hard than it hits guides. The GEO opportunity in e-commerce is being the source AI systems pull from for product comparisons and recommendations, which rewards structured product data, genuine reviews, and clear specifications. The bridge holds reasonably well because product SEO and product GEO both reward clean, structured, trustworthy product information, though e-commerce also faces the emerging complication of agentic shopping, where an AI agent rather than a human evaluates the options.

Publishers and media face the most direct existential pressure and the clearest freshness advantage. News and established publishers dominate AI citations across platforms, taking a large share, but the zero-click pattern means the citation often does not translate into a visit, and projected traffic losses for publishers are severe. The bridge holds in that authority and quality still earn citations, but the business model built on click volume is under the most strain. Publishers with strong brands and fast, credible production benefit from the freshness bias, entering citation pools within days, while those dependent on evergreen informational traffic face the steepest decline.

Local and service businesses are the least disrupted by the informational collapse and the most dependent on a different surface. Local intent queries, “near me” searches, and service bookings behave more like transactional queries and are less fully answered by a generative overview. The bridge holds well because local SEO fundamentals, accurate business information, reviews, and consistent presence across directories, are also what AI systems use to represent a local business in an answer. The main GEO work here is ensuring the brand’s information is consistent and credible across the sources AI systems trust.

Health, finance, and other YMYL sectors show the tightest coupling between SEO and GEO, which makes them the sectors where the slogan is most nearly true. In high-stakes topics, Google leans hard on established, high-ranking, authoritative sources, and the overlap between top rankings and AI citations can exceed 75 percent. A YMYL business that earned strong rankings through genuine expertise and authority carries most of that advantage into AI answers, because the engines are deliberately conservative about which sources they trust for consequential information. The bridge is strongest here precisely because the engines are least willing to experiment.

The pattern across sectors is that the SEO-to-GEO bridge holds best where trust is paramount and query intent is transactional or local, and weakest where the dominant AI platform reads training data over the live index and where informational content carried the business. A B2B SaaS company relying on ChatGPT visibility and a publisher relying on informational click volume face the hardest transitions. A YMYL business and a local service business face the gentlest. Every business should place itself on this map before deciding how much of its SEO advantage it can expect to carry into AI search, because the honest answer ranges from most of it to very little depending on where it sits.

Impact on in-house teams, agencies, and freelancers

The people who do this work face a professional transition, and it is worth being concrete about how the job changes, because the slogan is comforting partly for career reasons: it implies existing skills are enough. They are a strong start and not enough.

In-house SEO teams face the widest scope expansion. The job was ranking pages; the job is now visibility across ranked results and a set of AI surfaces that behave differently and demand different tactics. An in-house lead now has to coordinate with public relations for brand mentions, with the video team for YouTube presence, with content for extractable structure, and with analytics for a measurement model that tracks citations rather than positions. The skill of moving a page up a list still matters but is now one skill among several, and the teams that adapt fastest are the ones that already worked cross-functionally. The teams that struggle are the narrow technical shops that treated SEO as an isolated on-page discipline, because AI visibility is substantially an off-site, brand-level problem that on-page work cannot solve alone.

Agencies face both the biggest opportunity and the biggest credibility risk. The opportunity is real: businesses need help finding their way through a genuinely changed environment, and the agencies that build real expertise in platform-specific citation behavior, brand-mention strategy, and AI-visibility measurement have a real service to sell. The risk is that the field’s immaturity has made it easy to sell work that does nothing, from llms.txt files to AI sitemaps to vague GEO retainers, and the reputational damage from that is accumulating. An agency that wants to survive the next few years has to be able to show crawler logs, controlled tests, and citation-tracking data rather than confident slides, because clients are starting to notice when the promised AI visibility does not arrive.

Freelancers and small operators face a barrier and an opening. The barrier is that some of the highest-return GEO work, earning brand mentions across authoritative sources, building YouTube presence, and establishing entity authority, takes resources and time that a small operator may not have, and Google’s AI Mode in particular rewards the market presence that incumbents already own. The opening is that ChatGPT’s weaker correlation with classic authority metrics means a small, focused operator can sometimes earn visibility by becoming the clearest, best-described source for a narrow problem before a larger competitor bothers to compete for it. A freelancer who owns a specific niche completely can be cited alongside brands many times their size.

The uncomfortable truth for the profession is that the parts of the job that were most mechanical are the parts that transfer least, and the parts that were most about judgment, expertise, and genuine authority are the parts that transfer most. A practitioner whose value was executing repeatable on-page tasks faces the hardest transition, because those tasks matter less. A practitioner whose value was understanding a subject deeply, writing clearly, and building genuine authority faces the easiest, because that is what AI visibility rewards. The field is moving from optimization-as-technique toward authority-as-substance, and that reshuffles who is worth paying.

There is a practical career instruction in this. The skills worth building now are cross-channel: understanding how brand mentions form and how to earn them, understanding how each AI platform selects sources, building measurement systems that track citation across engines, and writing content that both ranks and extracts. The skills worth de-emphasizing are the narrow, gameable techniques that neither rank nor get cited. The professionals who thrive will be the ones who treated SEO as a means to genuine visibility rather than as a bag of tricks, because the tricks are exactly what the new environment discards.

Measurement, attribution, and the reporting trap

The fastest way to make bad GEO decisions is to measure it with SEO tools, and most teams are doing exactly that. The metrics that defined SEO success describe the wrong things in an AI world, and the reporting habits built around them actively mislead.

The first trap is collapsing AI visibility into one number. Because cross-engine citation overlap sits near 11 percent, a single “AI visibility” score averages across platforms that behave nothing alike and hides where a brand is winning and losing. A brand that dominates Perplexity and is absent from ChatGPT has a mediocre average and two very different realities underneath it, and only the disaggregated view is actionable. Measurement has to be per-platform, tracking citation share on each engine separately against a set of representative prompts, because the strategies to improve each are different.

The second trap is tracking rankings as a proxy for citations. With the exact-query overlap collapsing and fan-out generating sub-queries that no keyword tracker captures, a ranking report no longer describes AI exposure. A page can rank first and never be cited, or rank poorly and be cited constantly for a sub-question. The measurement that matters is inclusion in answers, monitored by running the prompts a business wants to win through the engines regularly and recording whether and how the brand appears. Tools like Ahrefs Brand Radar and various AI-visibility trackers now do this across ChatGPT, Gemini, Perplexity, and Google’s surfaces, and they measure the thing that ranking reports cannot.

The third trap is attribution failure in analytics. As covered earlier, AI-driven visits often land in analytics as direct or unattributed rather than as search or AI referral, so the channels look worse than the traffic warrants. A team reading GA4’s default buckets will undercount AI’s contribution and may conclude its content failed when it is quietly working. Fixing this requires deliberate categorization: separating zero-click cited losses, zero-click uncited losses, side-panel attribution failures, and genuine ranking losses into distinct buckets, because each demands a different response. Zero-click cited losses recover through brand and schema work; uncited losses recover only through content rebuilt for the target prompts; attribution failures are not losses at all and need fixing in the analytics layer, not the content.

The fourth trap is the blind-spot rate. Surveys found that while around 43 percent of marketers said they optimized for AI search, only around 14 percent actually measured it, leaving a large majority working blind. A team that changes content for AI visibility without tracking citation cannot tell whether the work helped, which is how ineffective tactics like llms.txt persisted for a year. Measurement is not a reporting afterthought in GEO; it is the only way to distinguish tactics that work from tactics that merely feel like work.

The measurement model that fits the AI era tracks a small set of honest metrics: citation share per platform against a defined prompt set, brand mention volume and quality across the web, YouTube presence for categories where it matters, AI referral traffic and its conversion rate held separately from volume, and the cited-versus-uncited split for priority pages. These metrics measure representation and authority, which is what AI visibility actually is, rather than position and click volume, which is what SEO used to be. A team that rebuilds its dashboard around them can see the new game clearly. A team that keeps reporting average position and total organic sessions is watching the old game end without seeing the new one begin.

The deeper point is that measurement discipline is the antidote to the whole field’s credibility problem. The reason snake oil sells in GEO is that buyers cannot easily verify claims, and the fix is to demand the evidence: crawler logs for technical claims, controlled tests for tactical claims, and per-platform citation tracking for visibility claims. A business that measures rigorously is protected from both its own wishful thinking and its vendors’ confident slides, which in a field this immature is worth more than any single tactic.

The GEO snake oil problem and the tells that give it away

The GEO field has a credibility problem serious enough that respected practitioners have started warning about it publicly, and a business evaluating GEO services needs to recognize the pattern because the cost of getting it wrong is a wasted budget and a false sense of security.

The warnings are not fringe. Lily Ray’s analysis found that around 30 sites that publicly boasted about GEO wins in 2024 and 2025 went on to suffer heavy organic traffic losses afterward, suggesting the touted tactics either did nothing or coincided with decline. Broader critiques have argued the SEO industry is getting AI search dangerously wrong in ways that hurt the sites paying for the work. One practitioner titled a column bluntly warning about generative engine optimization snake oil. These are the working professionals with the most reputational skin in the game, and their consensus is that a large share of what is sold as GEO is invented work.

The mechanism that lets snake oil thrive is information asymmetry. A buyer who hears on a conference stage or a LinkedIn post that Google released a new AI standard, or that a special file is now required for AI visibility, has no simple way to verify the claim, and the vendor selling the fix has every incentive not to correct it. The llms.txt episode is the archetype: a genuine but narrow developer proposal was repackaged as something every site supposedly needed, generated a year of billable work, and produced nothing measurable for search visibility, with audits finding thousands of dollars of such work delivering zero traffic impact. The vendors were not necessarily dishonest; they were operating in a market where invented work is easier to sell than real work, and where buyers who do not read primary sources have no defense.

The tells that give snake oil away are recognizable once named. A tactic that targets the machine rather than the reader or the wider web is the first tell, because the tactics that survived the shift to generated answers are the ones that improve genuine quality and reputation, not the ones that try to signal a crawler. A tactic sold with confident certainty about how an opaque, fast-changing system works is the second tell, because the honest state of GEO knowledge is probabilistic and contested, and anyone claiming to know exactly how ChatGPT selects sources is overselling. A tactic that cannot produce crawler logs, controlled tests, or citation-tracking data to support its claims is the third tell, because real GEO effects are measurable and fake ones are not.

There is a related failure that is not quite snake oil but wastes budget the same way: applying real tactics to the wrong platform. A team that pours effort into Perplexity-style extractable formatting while its buyers live in ChatGPT, where entity strength matters more, is doing legitimate work that does not move its actual target. The cross-engine divergence means that even honest, working tactics are platform-specific, and spending them against the wrong engine buys almost nothing. This is why the measurement discipline from the previous section is the real protection: a business that tracks per-platform citation can see which work moved which engine, and a business that does not is vulnerable to both fraud and honest misallocation.

The constructive version of this warning is that the tactics that survive scrutiny are the unglamorous fundamentals plus the off-site brand work: genuine topical authority, consistent mentions across quality sources, structured and extractable content, strong entity signals, and platform-appropriate presence, all measured rigorously. None of these are proprietary secrets, none require a special file, and all of them can be verified. A GEO program built on them will look less exciting than one promising a novel technical hack, and it will be the one that actually works. The excitement is usually the tell.

Entity clarity and the semantic layer models actually read

Underneath the tactics sits a concept that explains much of how AI systems decide what to trust, and understanding it clarifies why brand mentions and consistency matter more than links: the entity. AI systems reason about the world in terms of entities, the people, organizations, products, and concepts that the web discusses, and a brand’s visibility depends heavily on how clearly it exists as an entity in the machine’s understanding.

An entity is a thing the system recognizes as a distinct, identifiable subject with known attributes and relationships. Google built a knowledge graph around entities years ago, and language models absorb a related understanding from their training data, learning which brands exist, what they do, what category they belong to, and which other entities they are associated with. A brand that exists clearly as an entity, consistently described across many sources, is one the model can confidently mention. A brand with a fuzzy or inconsistent entity presence is one the model hedges on or omits, because it cannot be sure what the brand is. This is why the strongest AI-visibility signals are all about how widely and consistently a brand is discussed, rather than how many links point to its site.

Entity clarity is built through consistency and association. Using the same brand name, description, and category terms across a site and across the web helps the model form a stable understanding. Being mentioned alongside the category’s key terms and problems teaches the model what the brand is for, so that a brand consistently discussed in the context of, say, project management for engineering teams becomes an entity the model associates with that need. Structured data reinforces this by explicitly stating a brand’s identity and its relationships to verified profiles. Presence on the sources models trust most, encyclopedic references, editorial media, and authoritative industry sites, strengthens the entity because the model weights those sources heavily.

This reframes several earlier findings into a single mechanism. Brand mentions beat backlinks because a mention is a piece of entity information the model reads directly, while a link is a navigation instruction the model may not need. YouTube mentions correlate so strongly because video titles, descriptions, and transcripts are entity-rich text that the models trained on heavily. ChatGPT rewards entity strength because it reasons from training data where entities are learned, not from a live ranked list. The freshness and authority effects both feed the entity: fresh mentions keep the entity current, and authoritative mentions make it trusted. Entity clarity is the unifying idea beneath GEO, and it is largely an off-site, reputation-driven property that classic on-page SEO never directly addressed.

For the slogan, entity clarity is the concept that best explains both its truth and its limits. Good SEO built some entity clarity as a byproduct, through consistent naming, structured data, and the authority that rankings both required and signaled, which is why strong SEO brands start ahead. But good SEO never made entity-building its central goal, because rankings could be won page by page without a coherent brand entity. GEO makes the entity central, and a brand that ranks well through page-level work but has a weak entity presence will find that its SEO advantage does not fully convert. The work of GEO, stripped to its core, is the work of making a brand a clear, trusted, current entity in the machine’s understanding of its category, and that work is more about reputation across the web than optimization on a site.

Practical steps for a site that already ranks well

A site with strong rankings holds a real advantage and a specific risk: the advantage is that it clears the fundamentals and starts ahead; the risk is complacency, the belief that good rankings will carry into AI answers automatically. The steps below are ordered for a site that already ranks and wants to convert that position into AI citations.

Start by finding out where you actually stand, because rankings will not tell you. Run the prompts you most want to win through ChatGPT, Perplexity, Gemini, and Google’s AI surfaces, and record whether your brand appears, how it is described, and which of your pages, if any, get cited. Do this per platform, since the answers will diverge sharply. This audit usually surprises well-ranked sites, revealing strong Perplexity presence and near-total ChatGPT absence, or citations for pages that rank poorly and silence on the pages that rank first. The gap between your rankings and your citations is the map for everything that follows.

Fix extractability on your best pages next, because it is the cheapest high-return work for a site that already has authority. Move the direct answer to the top of each page and each section, into the first 80 words, ahead of context and preamble. Break content into self-contained blocks that each answer one question completely, with a clear heading and an answer-first sentence. Add comparison tables, definition sentences, and FAQ blocks where they fit the content honestly. Keep individual answers tight even as the site’s coverage stays broad. A well-ranked page that buries its answer is leaving citations on the table that a structural edit can capture.

Audit and strengthen your entity presence, which for a well-ranked site is usually the biggest gap. Check that your brand is described consistently across your site and the web, with the same name, category terms, and positioning. Add or correct structured data that states your organization’s identity and links to verified profiles. Identify where your brand is discussed across the web and where it is absent, and treat earning credible mentions as a priority equal to or above link acquisition. For most well-ranked sites, the constraint on AI visibility is not on-page quality but off-site brand presence, and this is where the budget should shift.

Address the platform where you are weakest deliberately. If your audit showed ChatGPT absence, the work is entity-building and third-party presence rather than more on-page optimization, because ChatGPT reads reputation over rankings. If it showed weak Perplexity presence despite good rankings, the work is likely extractability and freshness. Do not spread effort evenly; concentrate it on closing the specific gap the audit revealed, on the platform your buyers actually use.

Build a refresh cadence for the pages that matter most, since AI citations decay within a quarter without maintenance. Schedule updates to data, dates, and claims on your priority pages rather than treating publication as the end of the work, and add new evidence as it emerges. This keeps well-ranked pages in the citation pool that would otherwise age out even while their rankings hold.

Finally, rebalance the content mix toward the queries that still send traffic. A well-ranked informational library is the most exposed asset in the AI era, and the response is to keep it as an authority-builder and citation target while adding comparison content, product-level pages, and transactional content that AI intercepts less. A site that already ranks well should treat AI search as a conversion of existing authority into citations, not as a new discipline bolted on, and the fastest wins come from extractability and entity work rather than from anything exotic. The advantage is real; it just does not cash itself.

Practical steps for a site that ranks poorly but wants citations

A site without strong rankings faces a harder road and one genuine opening. The road is harder because the fundamentals that AI search requires are the same fundamentals that produce rankings, so a site failing at SEO is usually failing at the floor that GEO also demands. The opening is that some AI platforms, ChatGPT most of all, weight classic authority metrics weakly enough that a focused newcomer can earn visibility before it earns rankings.

Fix the fundamentals first, because there is no shortcut around them. A site that is not reliably crawlable, indexable, fast, and technically sound cannot be retrieved, and a page that cannot be retrieved cannot be cited on any platform. This is the least glamorous work and the non-negotiable prerequisite. Before any GEO-specific tactic, ensure the site can be fetched, parsed, and understood, because everything else is wasted effort on a site that fails here. Google’s fundamentals-first message is exactly right for a site in this position.

Choose a narrow territory and own it completely, rather than competing broadly against established brands. The one real advantage a small or low-ranking site has in AI search is that ChatGPT’s weak correlation with domain authority means being the clearest, best-described, most-cited source for a specific problem can earn visibility even without incumbent-level search demand. A newcomer that covers one narrow subject with genuine depth, clarity, and original insight can become the entity the model associates with that subject. Spreading thin across many topics forfeits this advantage; concentration is the strategy.

Produce original evidence, because it is the most citable asset a small site can create and the hardest for competitors to replicate. Original data, internal case studies, proprietary benchmarks, and firsthand analysis give AI systems something they cannot get elsewhere, and the Princeton findings on statistics and citations suggest models reward exactly this kind of specific, evidenced content. A site that reports its own numbers is more citable than one that only summarizes others, and this is a place where effort beats budget.

Build entity presence off-site from the start, since AI visibility is substantially a reputation property. Earn mentions in credible sources within your niche, appear in relevant videos and podcasts, and get discussed in the community spaces your category uses. For a small site this is slow and unglamorous, but it is the work that AI systems reward, and it compounds. Because brand mentions correlate far more strongly with AI visibility than backlinks, a small site should prioritize being talked about over being linked to, which is often more achievable for a genuine expert than a link-building campaign.

Write for extraction from the first page you publish, building the answer-first, well-structured, self-contained format into your content habits rather than retrofitting it later. A small site has the advantage of no legacy content to fix, and content built correctly from the start is both more rankable and more citable.

Measure per platform from the beginning, running your target prompts through the engines and tracking whether you appear, so you can tell which work is moving which engine. A small site cannot afford wasted effort, and measurement is what prevents it. A low-ranking site’s realistic path is to fix the floor, own a narrow niche with original evidence, build entity presence off-site, and target the platforms that reward focus over incumbency, while accepting that Google’s AI Mode, which leans hard on established authority, will be the last surface to open up. The opening is real but narrow, and it rewards depth and patience rather than breadth and speed.

Risks, limits, and the things the evidence cannot settle

Honesty about GEO requires admitting how much remains unsettled, because the field’s confidence outruns its evidence, and a practitioner who mistakes current findings for permanent laws will be caught out when the systems change, which they do constantly.

The largest limit is that the systems are opaque and moving. Generative engines are black boxes whose source-selection functions are not disclosed, and they change frequently: Google made Gemini 3 the default for AI Overviews in late January 2026, and any study straddling such a change measures two different systems. Every correlation in this article is a snapshot of a moving target, and correlations measured in one quarter may not hold in the next. The finding that YouTube mentions correlate most strongly with AI visibility, for instance, partly reflects that current models trained heavily on YouTube transcripts, and if training and weighting shift, so will the signal. Treating any specific finding as durable is a mistake; treating the directions as informative is reasonable.

Correlation is not causation, and the field’s headline findings are correlational. The Ahrefs researchers were explicit that high-visibility brands also have high cross-platform presence, which does not prove that building that presence causes visibility. It is plausible that brand mentions cause citations, and plausible that both are caused by an underlying authority that a business cannot manufacture quickly. A team that reads “brand mentions correlate at 0.664” as “buy brand mentions and get cited” is over-reading the data. The honest position is that the correlated signals are strong operating clues about where to invest, not proven causal levers.

The overlap numbers, as discussed, are genuinely contested, and no single figure should be trusted. Studies disagree because of detection methodology, overlap thresholds, query mix, fan-out, and timing, and the disagreement is not going to resolve into one clean number, because the underlying reality is that overlap varies by query type, platform, and moment. The reliable claim is the direction within a consistent methodology; the unreliable claim is any specific percentage stated without its methodology.

The traffic and conversion findings carry their own uncertainty. The high conversion rate of AI referral traffic is measured on tiny volumes, and whether it holds as volume scales is unknown; it is possible that early AI-search users are unusually high-intent and that the conversion advantage shrinks as adoption broadens. The recovery in click-through rates from the December 2025 floor is real but recent, and whether it represents a stable new baseline or a temporary rebound is not yet clear. Projections of 50 percent organic traffic declines by 2028 are extrapolations from measured trajectories, not certainties, and the trajectories could bend.

There are open questions the evidence simply cannot yet settle. Whether agentic browsing and AI agents that act rather than answer will reshape the picture again is unknown, though early signals suggest a further shift. Whether the cross-engine divergence will narrow as the platforms mature or widen as they differentiate is genuinely uncertain; the current data shows it widening. Whether Google’s AI surfaces will keep pulling from ranked content or drift further toward independent selection is unresolved, with different studies pointing different ways. The intellectually honest stance is that GEO is a young discipline built on correlational snapshots of opaque, fast-changing systems, and the right posture is to act on the strongest current evidence while measuring continuously and expecting the specifics to change. Anyone offering certainty in this field is either selling something or has not read the data carefully enough to know how contested it is.

Realistic scenarios for search and AI answers through 2027

Forecasting a field this volatile is hazardous, but a few scenarios are worth laying out because they frame the decisions a business faces now, and the differences between them matter for how much to invest and where.

The most likely near-term scenario is continued fragmentation. The cross-engine divergence has been widening, not narrowing, and the platforms are differentiating deliberately: ChatGPT leaning on entity strength and training data, Perplexity on real-time retrieval and dense citation, Google’s surfaces on the search index and authority signals. In this scenario, the single-strategy dream stays dead, and GEO remains a portfolio problem requiring per-platform work and measurement. A business that builds broad entity authority, structured extractable content, and platform-specific presence is positioned for this scenario, which is the one the current data most supports.

A second scenario is partial reconsolidation around Google. If Google’s AI Mode and AI Overviews keep gaining share, and they are gaining fast, with AI Mode reportedly crossing 100 million users and processing over a billion queries a month by early 2026, then the most search-correlated AI surfaces become the most important ones, and the slogan recovers some truth. In this world, strong classic SEO carries further because Google’s AI surfaces are the ones built on the search index, and a business heavily invested in Google-ecosystem visibility benefits. The risk in betting on this scenario is that it depends on Google’s surfaces continuing to favor ranked content, which the overlap-collapse data suggests they are doing less over time.

A third scenario is the rise of agentic search, where AI agents act on behalf of users rather than presenting answers for humans to read. Early signals point this way: coding assistants and agents already consume machine-readable interfaces, and the industry is beginning to build for agents rather than pages. If this accelerates, the relevant optimization shifts again, toward being usable by agents that evaluate and act, and toward standards like machine-readable interfaces that are currently niche. A business in this scenario would need to think about being agent-legible, a discipline barely formed today. The llms.txt saga, for all its failure as a search-visibility play, hinted at this direction: its genuine use is agent infrastructure, not answer visibility.

A fourth scenario, less discussed and worth holding, is a credibility correction. If the GEO field’s snake-oil problem keeps producing sites that pay for AI-visibility work and get nothing, and if respected practitioners keep documenting the failures, the market may correct toward the boring fundamentals, deflating the acronym economy and returning attention to genuine authority and quality. In this scenario, the businesses that quietly did the fundamentals and the off-site brand work win, and the ones that chased tactical hacks lose the money they spent chasing them.

These scenarios are not mutually exclusive; the realistic future is probably fragmentation with Google gaining share, agentic search rising at the edges, and a slow credibility correction underneath. The strategy that survives all four is the same: build genuine topical authority, earn broad and consistent brand mentions, structure content for extraction, maintain it for freshness, establish clear entity presence, and measure citation per platform. None of these depend on a specific scenario, none require a proprietary trick, and all of them compound. The strategies that fail in most scenarios are the ones that bet on a single platform, a single tactic, or a slogan, which is the deeper argument against treating “good SEO is good GEO” as a plan rather than a partial truth.

From position zero to the cited answer, a short history

The shift to AI answers did not arrive from nowhere. It is the latest and largest step in a decade-long move by search engines to answer questions in place rather than send users elsewhere, and seeing that history clarifies why the SEO-to-GEO transition feels both sudden and familiar.

The first step was the featured snippet, the boxed answer that appeared above the ranked results and became known as position zero. It gave a page the top spot and often the answer, which meant users increasingly got what they needed without clicking through. SEO adapted by optimizing for the snippet, structuring content to win the box, and the discipline learned that being the answer could matter more than being the first link. The knowledge panel extended the pattern, pulling entity information directly into the results and answering factual queries in place. Both were early forms of the same idea that AI answers now embody: the search engine keeping the user on its page by resolving the query itself.

The zero-click phenomenon grew alongside these features. SparkToro first documented in 2019 that around half of Google searches ended without a click, and the figure climbed steadily as features multiplied, reaching around 60 percent by 2024 before AI Overviews accelerated it toward 68 percent by 2026. The trend predates generative AI by years; AI Overviews did not create zero-click search, they poured fuel on a fire that had been burning since featured snippets. A practitioner who lived through the featured-snippet era already experienced a smaller version of the current disruption: rankings holding while clicks fell, and the prize shifting from the link to the answer.

What generative AI changed was the scale and the mechanism. A featured snippet pulled one passage from one page and attributed it clearly. An AI answer synthesizes many sources into original prose, names a handful of them, and can answer questions no single page fully addresses. The featured snippet was extractive; the AI answer is generative. That difference is why the old optimization partly transfers, the answer-first, extractable structure that won snippets also helps win citations, and why it partly does not, because winning a synthesized answer built from many sources through fan-out is a different and harder problem than winning one box from one page.

The historical view offers a useful correction to the panic. The search world has been moving toward answering in place for a decade, SEO has adapted to each step, and the businesses that treated content as a genuine service to users rather than as bait for clicks weathered each transition better. The AI answer is the biggest step in a long-running direction, not a break from it, and the response that worked before, building genuine authority and serving the user’s actual need, is the response that works now, scaled up. The slogan’s kernel of truth lives here: the fundamentals that survived featured snippets and knowledge panels survive AI answers too. What has always failed to survive is the gaming, and each step has discarded more of it.

Advertising arrives in AI answers and changes the incentives

A development that will shape GEO more than any tactic is the arrival of advertising inside AI answers, because it changes what the platforms are optimizing for and therefore what visibility will cost. The economic model underneath AI search is being decided now, and it splits the platforms in ways that matter for strategy.

The split became visible in a single week in February 2026. OpenAI launched advertising inside ChatGPT for the first time on the ninth, and eight days later Perplexity confirmed it was abandoning advertising for good, publicly arguing that paid placements would undermine the trust that drives its Pro subscriptions. Two of the most prominent AI search products chose opposite business models in the same week. Google, for its part, has been testing ads within AI Mode, and ads now appear in a sizable share of AI results, with reports of paid placements in around a quarter of AI results at materially higher cost-per-click than the same queries without an AI answer.

The implications for visibility are direct. On a platform that monetizes through advertising, organic citation competes for space with paid placement, and the incentive for the platform is to answer in place and sell the surrounding attention, which intensifies the zero-click effect. On a platform that monetizes through subscription, like Perplexity’s chosen path, the incentive is to preserve trust and citation quality, which favors clean sourcing and may keep organic citation worth more. A brand’s GEO strategy has to account for which model each target platform runs, because the organic opportunity is shaped by how the platform makes money.

The paid-and-organic interaction produces one counterintuitive finding worth holding. Seer Interactive found that brands cited organically inside an AI Overview earned not only more organic clicks but around 91 percent more paid clicks on the same queries, and that appearing in the AI answer lifted paid performance by around 40 percent. Organic citation and paid performance reinforce each other: being recommended by the AI engine primes the buyer to click the brand’s ad and to carry the brand into later searches. Organic AI visibility and paid AI visibility are becoming complementary rather than substitutes, which means a brand cannot simply buy its way in and ignore citation, nor earn citation and ignore the paid surface.

The cost trajectory is the part that should concern budget holders. With organic clicks compressing and cost-per-click rising on AI-answer queries, the blended cost per lead is climbing whether a brand adjusts its bids or not. The queries that trigger AI answers are becoming both more expensive to win on the paid side and harder to earn traffic from on the organic side, squeezing from both directions. This is why the high conversion rate of AI-referred traffic matters so much: it is part of what justifies competing for these queries despite the rising cost, because the traffic that survives is worth more.

The strategic reading is that advertising’s arrival makes organic citation worth more, not less, because it is the visibility a brand does not have to keep paying for and because it lifts the paid performance a brand does pay for. As the platforms monetize, the brands with strong earned citation presence will have an advantage over those forced to buy every impression, which is the same advantage that organic search visibility conferred over paid search for two decades. The medium changed; the logic did not.

Regulation, privacy, and the data underneath AI answers

GEO does not sit outside the law, and the regulatory and data questions surrounding AI search are moving fast enough to reshape what is possible, which makes them part of any serious strategy rather than a footnote.

The training-data disputes are the most immediate. AI answers are built partly from content the platforms trained on and partly from content they retrieve, and both have become contested. Reddit sued Perplexity over scraping in October 2025, after which Perplexity’s Reddit citations reportedly dropped sharply and YouTube partially filled the gap, a concrete example of a legal dispute directly altering which sources an AI engine cites. Publishers have been renegotiating licensing terms with AI platforms, and the outcomes will determine which sources remain in the citation pools and on what terms. A brand’s visibility can shift not because it changed anything but because the platform’s access to a source changed, which is a risk outside the brand’s control and worth monitoring.

Crawler behavior sits at the intersection of technical practice and policy. AI crawlers now make up a large share of all bot traffic, with one analysis putting AI crawlers at around 22 percent of bot traffic and roughly half of that volume fetching content for model training rather than for retrieval that would generate a citation or a visit. A site owner faces a genuine tension: blocking training crawlers protects content from uncompensated ingestion but may reduce the brand’s presence in the training data that drives entity strength, particularly for a platform like ChatGPT that relies heavily on parametric knowledge. The robots directives that govern this are the real control surface, respected by the major AI crawlers, and the decision about what to allow is now a strategic one with visibility consequences on both sides.

The formal regulatory picture is emerging unevenly. The EU AI Act shapes how AI systems, including search-adjacent ones, can operate in Europe, and the rollout of features like AI Mode into European markets has been paced against compliance with it, with broader European availability expected to follow the Act’s requirements. Some jurisdictions are discussing answer-engine and AI-citation standards, and proposals for AEO-related regulation have been floated in the EU, while markets like the UK did not yet have AI-specific citation standards as of early 2026. The regulatory environment will affect where and how AI search operates and what transparency the platforms must provide about their sourcing, which bears directly on how measurable and gameable citation is.

Privacy enters through the data AI answers use and the way they are personalized. AI systems can tailor answers to a user, which raises the same personalization and data-handling questions that dogged personalized search, and the platforms’ handling of user queries and the content they ingest is subject to the general tightening of privacy law. For a brand, the practical privacy exposure is mostly indirect, in how its content and data are ingested and represented, but the direction of regulation toward more transparency and more control over data ingestion will shape the field.

The honest summary is that the legal and data foundations of AI search are unsettled, and a brand’s visibility is partly hostage to disputes and rules it does not control. A source can drop out of a citation pool because of a lawsuit, a licensing breakdown, or a policy change, and a brand can gain or lose training-data presence based on crawler decisions and legal outcomes. This is a genuine limit on how much any optimization strategy can guarantee, and it argues, again, for building broad, resilient authority across many sources rather than depending on any single platform or citation channel that a regulatory or legal shift could close.

A working answer to whether good SEO is good GEO

The claim that good SEO is good GEO is neither true nor false. It is a floor mistaken for a building, and the honest answer separates the part that holds from the part that does not.

The part that holds is substantial. Good SEO fundamentals are the prerequisite for AI visibility, because every generative engine sits on a retrieval layer that is a search layer, and a site that fails at crawlability, authority, clarity, or trust fails everywhere. Google is right that there is no GEO without SEO fundamentals. A business with strong SEO starts the AI race ahead, sometimes far ahead, and in trust-heavy sectors like health and finance, where the engines lean hard on established authoritative sources, the overlap between ranking well and being cited remains high enough that the slogan is nearly true. The technical, content-quality, authority, structure, and intent work of good SEO transfers cleanly and keeps its worth.

The part that does not hold is equally substantial and getting larger. The overlap between exact-query rankings and AI citations roughly halved between mid-2025 and early 2026 and continues to weaken, so a top ranking no longer guarantees a citation. Query fan-out moved the target from a position to a topic, rewarding breadth over single-page optimization. The engines agree so little with each other, cross-engine citation overlap near 11 percent, that no single strategy wins them all, and ChatGPT in particular rewards entity strength and brand mentions over rankings so completely that good SEO barely predicts good ChatGPT visibility. Off-site brand mentions beat backlinks roughly three to one, YouTube presence correlates more strongly than any on-site signal, and the gaming layer of old SEO fails or backfires. The economics changed too: the click that justified SEO is disappearing, and the prize is now being the cited source in a zero-click answer.

The synthesis a practitioner can act on is this. Good SEO is necessary for good GEO and no longer sufficient for it. It buys eligibility, not citations. The work that converts the SEO floor into AI visibility is largely off-site and brand-level, building genuine topical authority, earning wide and consistent mentions, establishing a clear entity, maintaining content for freshness, structuring it for extraction, and doing all of this per platform because the platforms diverge. This work is not a betrayal of good SEO; it is its natural extension, the part good SEO always gestured at through E-E-A-T and authority but never made central.

The businesses that will do well are the ones that keep the fundamentals, add the off-site and platform-specific work, measure citation rigorously per engine, and treat the slogan as a starting intuition rather than a strategy. The businesses that will struggle are the ones that hear “good SEO is good GEO,” conclude their existing work is enough, and wait for their rankings to carry them into AI answers that increasingly cite someone else. The comforting version of the slogan is the dangerous one. The useful version is narrower and harder: good SEO gets you to the door, and getting through it is a different, larger job than the one that got you there.

Questions people ask about SEO and GEO overlap

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

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

GEO: Generative Engine Optimization (Aggarwal et al., KDD 2024) The peer-reviewed paper that named generative engine optimization, introducing the GEO-bench benchmark and showing that source citations, statistics, and quotations could lift a site’s visibility in generative answers by up to 40 percent while keyword stuffing performed below baseline.

GEO: Generative Engine Optimization (arXiv preprint) The open-access preprint of the Princeton-led GEO study, useful for the full method, the nine optimization strategies tested, and the per-strategy visibility results underpinning most later GEO practice.

How often do top-ranking pages get cited in AI answers (Ahrefs) Ahrefs research from mid-2025 finding a high overlap between top-ranking pages and AI Overview citations at that point, one of the datasets that fueled the “good SEO is good GEO” reading before the overlap began to fall.

The overlap between AI search engines and Google (Ahrefs) An analysis across roughly 15,000 prompts measuring how little the major AI engines agree with Google and with each other, reporting overall overlap around 12 percent and Perplexity as the most search-aligned engine near 28 percent.

Google AI Overview citations from top-ranking pages drop sharply (Search Engine Journal) Coverage of the sharp early-2026 decline in how often AI Overviews cite the highest-ranking organic pages, a key data point for the collapsing exact-match overlap between rankings and citations.

AI Overview citations drop, Ahrefs finds (DesignRush) Reporting on the Ahrefs finding that the share of AI Overview citations coming from top-ranking pages fell to around 38 percent, documenting the downward trajectory of ranking-to-citation overlap.

Rank overlap after 16 months of AI Overviews (BrightEdge) BrightEdge tracking data showing how the overlap between organic rankings and AI Overview citations shifted over sixteen months, offering a longitudinal counterpoint to single-snapshot studies.

AIO and organic rankings overlap research (seoClarity) A study reporting very high overlap, above 90 percent, when AI Overview citations are compared against the top 20 organic results, illustrating how a wider ranking window produces far higher overlap figures than exact-position comparisons.

Google ranking and AI citations study (Originality.AI) An independent analysis of how often AI citations come from top-ranking pages, reporting overlap figures in the roughly 48 to 52 percent range and adding methodological nuance to the wider debate.

Overlap between top Google rankings and AI-cited sources has collapsed (PR Newswire, 5W/Brandlight) Research announcement reporting that the overlap between top Google rankings and AI-cited sources fell from about 70 percent to under 20 percent, the headline figure for the divergence thesis.

Brand mentions and AI Overview visibility correlation (Ahrefs) The Ahrefs correlation study across tens of thousands of brands finding that brand mentions correlated with AI visibility far more strongly than backlinks, the basis for the roughly three-to-one mentions-over-links finding.

What correlates with AI brand visibility (Ahrefs) A broader Ahrefs correlation analysis identifying YouTube mentions as the single strongest signal associated with AI brand visibility, ahead of classic on-page factors.

Ahrefs study on AI brand visibility signals (Business Wire) The press release summarizing the Ahrefs correlation findings on the off-site and brand-level signals most associated with AI visibility, useful as a primary-source confirmation of the headline numbers.

Why SEO fundamentals still matter in 2026: what Google said in Zurich (iBrandLabs) Coverage of Google’s public position, articulated at a Zurich event, that there is no generative engine optimization without solid SEO fundamentals underneath it, the clearest statement of the necessary-but-not-sufficient argument.

llms.txt in 2026: the full guide (Limy) A detailed guide documenting how llms.txt works, why major AI crawlers largely ignore it, and why the file has produced no measurable search or citation benefit despite heavy promotion.

Google confirms llms.txt has no SEO value (Digital Applied) Reporting on Google’s confirmation that llms.txt carries no ranking or citation value, alongside audit findings that the file goes unread by the crawlers it was meant to guide.

In 2026, less than one-third of Google searches send a click (SparkToro) SparkToro clickstream analysis quantifying the zero-click reality of 2026, with under a third of Google searches producing an organic click, the economic backdrop for the shift from clicks to citations.

Google AI Mode zero-click rate and 100M users (Nobori) An analysis of Google AI Mode adoption and its very high zero-click rate, documenting how the conversational search surface answers most queries without sending traffic onward.

Why organic traffic dropped in Q1 2026 (Digital Strategy Force) A practitioner analysis connecting stable rankings to falling organic traffic through the growth of AI answers, useful for the mechanism behind the ranking-holds-traffic-falls pattern.

How ChatGPT, Google AI Overviews, and Perplexity source information in 2026 (Leapd) A comparison of how the three major engines retrieve and cite, documenting the divergence in sourcing behavior that makes cross-engine strategy necessary.

AI engines citation comparison (Whitehat SEO) A comparative study of citation behavior across AI engines, including the low cross-engine overlap near 11 percent that underpins the argument that no single strategy wins every platform.

ChatGPT vs Perplexity for AI visibility in 2026 (QuickSEO) A head-to-head analysis of ChatGPT and Perplexity citation, traffic, and conversion behavior, supporting the finding that Perplexity is search-aligned while ChatGPT leans on training data and trusted sources.

GEO vs SEO (Semrush) Semrush’s practitioner overview of how generative engine optimization differs from and overlaps with classic SEO, useful for definitions and the shared-fundamentals framing.

What is generative engine optimization (Search Engine Land) A reference explainer defining GEO, its history, and its relationship to SEO, providing the baseline vocabulary used throughout the analysis.

Google AI Mode and SEO in 2026 (Wecoode) An analysis of Google AI Mode and the query fan-out mechanism, explaining how one user query becomes many machine sub-queries and why that ends one-keyword-one-page thinking.