ChatGPT slipped below half the AI assistant market and the trend matters more than the milestone

For the first time since OpenAI launched its chatbot in late 2022, ChatGPT no longer accounts for the majority of the AI assistant market. According to Sensor Tower’s State of AI 2026 report, published on June 16, 2026, ChatGPT’s share of global AI assistant users fell to 46.4% by the end of May 2026, the first reading the firm has ever recorded below the 50% line. Google’s Gemini sat at 27.7% and Anthropic’s Claude at 10.3%, with Grok, Perplexity, DeepSeek and Meta AI each holding under 5%.

The headline number and what Sensor Tower actually measured

The first thing worth fixing in place is what that 46.4% is a share of. Sensor Tower measures the market with a metric it calls True Audience, which counts unique users across mobile apps, mobile web and desktop web, deduplicated so that one person using ChatGPT on a phone and a laptop is not counted twice. The headline figure covers 25 markets rather than the entire planet, and it is one research firm’s estimate built on a consumer panel and modelling, not an audited census agreed across the industry. That matters, because almost every other tracker measuring “AI market share” reports different numbers, and the differences are not errors. They come from measuring different things.

The second thing worth fixing in place is the timing. The May figure is where the share settled, not the moment the line was crossed. Sensor Tower dates the actual drop below 50% to March 2026. By the time the report landed in June, ChatGPT had spent roughly a quarter of the year as a sub-majority product. The May number is the tail of a movement that started earlier, which is why the trajectory carries more information than the milestone.

That trajectory is the part of the story that does not fit in a headline. Sensor Tower’s True Audience series, as reported alongside the launch, shows ChatGPT holding about 65.3% in December 2024, falling to 52.8% by December 2025, and reaching 46.4% by the end of May 2026. Some outlets cited an even higher starting point — around 81% in early 2024 — using slightly different framing, but the direction is identical across every version. ChatGPT has shed close to nineteen percentage points of relative share in about seventeen months, and the slope has been remarkably steady. A single bad month can be noise. A steady decline across more than a year, through multiple product launches and price changes, is a structural signal.

None of this means ChatGPT is shrinking. The same report that recorded the share drop also recorded ChatGPT crossing more than 1.1 billion monthly active users in May 2026, making it the fastest mobile app in history to reach a billion users. Both facts are true at once, and holding them together is the whole point of the analysis that follows. A product can grow its user base at a record pace and still command a smaller slice of the market, if the market grows faster than the product does. That is exactly what happened.

For anyone whose work depends on these platforms — marketers, founders, product teams, investors, and the millions of professionals who now treat an assistant as part of their daily toolkit — the practical message is not “ChatGPT is in trouble.” It is that the single-app era of generative AI has ended, and the assumptions built during that era no longer hold. Strategies that assumed one assistant reached almost everyone, that distribution would follow model quality, and that user growth was the only metric that mattered were all reasonable in 2023. In 2026 they are wrong in ways that cost money. The rest of this piece works through what the numbers say, why the share moved, who absorbed it, and what a fragmenting market changes for the people who have to operate inside it.

A milestone and a warning sign in the same month

It is rare for a company to set an all-time record and trigger a wave of “is the leader slipping” coverage in the same thirty days. OpenAI managed both in May 2026. ChatGPT reached one billion monthly active users faster than any application ever built — faster than TikTok, YouTube, Instagram, Gmail, WhatsApp and Facebook, each of which took five to eight years to get there. ChatGPT did it in roughly three years from launch. By any historical standard, that is the fastest consumer adoption curve on record.

The reason the milestone and the warning sign arrived together is that they measure different axes. The billion-user figure measures absolute scale. The 46.4% figure measures relative position. When a category is expanding quickly, those two numbers can move in opposite directions, and the gap between them is precisely where the interesting story lives.

There is also a measurement wrinkle worth naming, because it shapes how the headline travels. OpenAI itself usually reports weekly active users, and it last publicly cited around 900 million weekly active users in February 2026. Sensor Tower’s billion-plus figure is monthly active users across apps. Weekly and monthly counts are not interchangeable, and conflating them produces confusion that benefits no one. A weekly figure is a stricter test of habit; a monthly figure is a broader test of reach. ChatGPT looks enormous on both, but the specific number quoted depends entirely on which window and which surface the analyst chose.

The deeper point is about narrative gravity. When a product becomes the default for a billion people, every decision it makes — a pricing change, a new feature, a government contract, an advertising test — stops being a product decision and becomes a public event. Scale converts ordinary business choices into political, commercial and personal flashpoints. A startup with two million users can sign a controversial partnership and absorb the reaction quietly. A platform with a billion users cannot. The very dominance that produced the milestone is what makes the share number fragile, because dominance at that scale invites scrutiny that smaller rivals never attract.

This is why treating the sub-50% reading as a simple “decline” misreads it. ChatGPT did not lose users in May; it almost certainly gained them. What it lost was the illusion that it was the market. For two years, “AI assistant” and “ChatGPT” were close enough to synonymous that most people, most brands and most strategies treated them as one. The May data formalizes a separation that had been building underneath the surface for a year: the category and the product are now clearly distinct, and the category is bigger than any single product inside it.

Reading a share decline that is not a user decline

The arithmetic here is the single most misunderstood part of the story, so it is worth doing slowly. Market share is a ratio, not a count. A product’s share can fall in three different ways: it can lose users while the market holds steady, it can hold users while the market grows, or it can gain users while the market grows even faster. Only the first of those is a genuine decline. ChatGPT’s situation is the third, which is a very different thing.

The overall AI assistant market expanded sharply over the measurement window. By various estimates the category grew on the order of 22% between late 2025 and early 2026, and the broader generative AI app economy is growing faster still. When the denominator grows faster than your numerator, your share falls even as your business expands. ChatGPT added users at a record pace and still saw its slice shrink, because Gemini and Claude added users at an even faster relative pace from a smaller base.

This distinction is not pedantic. It changes the strategic reading completely. If ChatGPT were losing users in absolute terms, the correct interpretation would be product fatigue, churn, or a superior competitor pulling people away one by one. That is a defensive crisis. What the data actually shows is a category maturing into a multi-player structure while the incumbent keeps growing. That is a competitive normalization, which is what almost always happens to a first mover once a market becomes large and valuable enough to attract serious, well-funded rivals.

History offers the pattern. Search had multiple credible engines before Google consolidated it. Smartphones had a brief period of near-monopoly platforms before settling into a duopoly. Browsers, social networks and cloud platforms all passed through a phase where one early leader held a commanding share that then eroded — not because the leader collapsed, but because the prize grew large enough that capable competitors invested heavily to take a piece. A dominant share in a small market often becomes a large share in a big market, and a large share is no longer a monopoly. That is the shift the 46.4% number records.

For operators, the actionable consequence is straightforward and uncomfortable. Any plan that quietly assumed “if we show up in ChatGPT, we reach everyone” is now demonstrably incomplete. A meaningful and growing share of AI-mediated activity — more than half of users, by Sensor Tower’s count — happens somewhere other than ChatGPT. The platforms absorbing that activity have different strengths, different user populations, and different ways of surfacing and recommending content. Treating them as interchangeable, or ignoring them because they are smaller, leaves reach on the table. The fragmentation is the opportunity and the risk at the same time.

Inside the true audience method and why measurement defines the story

Every market-share claim about AI assistants is really a claim about a methodology, and the methodologies disagree violently. Understanding why is not a side note. It is the only way to read any of these numbers without being misled, and it is the reason a single confident-sounding percentage should always raise a question rather than settle one.

Sensor Tower’s True Audience metric counts unique users across mobile apps, mobile web and desktop web, deduplicated and aggregated from a global first-party consumer panel combined with app-store data and modelling. Its strength is that it tries to capture a real person wherever they touch a product, rather than counting app installs or raw web sessions. Its limits are equally real: it covers a defined set of 25 markets, it leans on panel data and statistical modelling rather than direct server logs, and it does not see usage that happens through embedded surfaces or developer APIs. It is a serious, well-constructed estimate. It is not a meter reading from inside each company.

Now compare the alternatives, because the contrasts are stark. Web-traffic trackers such as Similarweb measure visits to a product’s website and exclude mobile apps, embedded experiences and API calls entirely. On that basis, one April 2026 ranking put ChatGPT at roughly 54.7% of worldwide AI chatbot web visits among the seven largest assistants, with Gemini second and the rest splitting a much smaller remainder. A DNS-based view, like Cloudflare’s 1.1.1.1 ranking, measures something different again — which services devices resolve most often — and produced a stable top order of ChatGPT, then Claude, then Perplexity for months. Search-data firms like Semrush estimate yet another picture from their own crawl and clickstream models, and have at various points shown ChatGPT anywhere from the low 60s to the high 70s in percentage terms.

These are not contradictions to be resolved by finding the “real” number. They are different instruments pointed at different parts of the same animal. A US-only ranking inflates US-concentrated products like Claude and understates Asia-heavy products like DeepSeek. A worldwide ranking does the reverse. A web-visits ranking ignores the billions of app sessions where most consumer assistant use actually happens. A panel-based unique-user ranking smooths over heavy-user intensity. Each is internally valid and externally incomparable.

This is why the most important discipline for anyone citing these figures is to state the source, the metric and the window every single time, and to treat any report that contradicts itself — one headline share with a different figure buried in its own trend data — as an estimate rather than a measurement. A ranking is only as trustworthy as the methodology a reader can inspect. The corrective for noisy numbers is disclosure, not a more confident black box.

For the specific claim at the center of this piece, the honest framing is this: by Sensor Tower’s deduplicated, multi-surface, 25-market True Audience method, ChatGPT fell below 50% in March 2026 and sat at 46.4% at the end of May. Independent equivalents for that exact split do not exist, so it should be read as a well-sourced snapshot rather than a settled fact. The direction of travel, though, is corroborated everywhere. Whatever instrument you pick, ChatGPT’s relative share is falling and the gap to Gemini and Claude is narrowing. The disagreement is about the precise altitude, not the descent.

The reason this measurement literacy matters beyond academic tidiness is that real budgets are now allocated against these numbers. A marketer deciding how much to invest in being visible inside Gemini versus ChatGPT, an investor weighing OpenAI against Anthropic, a product leader choosing which assistant to optimize an integration for — all of them are reading share figures and making bets. Reading the methodology is the difference between a bet and a guess.

The competitive scoreboard behind the 46.4 percent

Behind the single headline figure sits a fuller scoreboard, and the gaps between the players are as informative as their ranks. By Sensor Tower’s State of AI 2026 count, the order at the end of May 2026 was ChatGPT at 46.4%, Gemini at 27.7%, and Claude at 10.3%, with Grok, Perplexity, DeepSeek and Meta AI each below 5%. In raw reach, the same report put ChatGPT above 1.1 billion monthly active users, Gemini at about 662 million, and Claude at roughly 245 million.

AI assistant standing as of late May 2026 (Sensor Tower State of AI 2026)

These are Sensor Tower’s deduplicated estimates across 25 markets; user counts mix monthly active figures, and the smaller players’ shares are reported as bands rather than exact points.

A few things stand out. The first is the size of the step down from ChatGPT to Gemini and then to Claude. This is not three roughly equal players; it is a clear leader, a strong and fast-growing challenger, and a much smaller but strategically important third. Gemini’s 662 million is now within hailing distance of being a genuine peer to ChatGPT on reach, where a year earlier it was not. Claude’s 245 million is a fraction of either, but as later sections show, raw user count is the metric on which Claude looks weakest and the metric that matters least to Anthropic’s actual business.

The second is the cliff after the top three. Once you pass Claude, every remaining assistant sits under 5% by this measure. That is a two-tier market: three platforms that together account for the overwhelming majority of consumer assistant activity, and a long tail of products that are either regionally concentrated, niche by design, or embedded inside larger apps. The top three platforms command the lion’s share of time spent in the category, with Sensor Tower’s own figures attributing close to 90% of time spent across AI assistant apps in early 2026 to a small handful of leaders.

The third, and least obvious, is that ranks differ from momentum. ChatGPT leads on share and scale, but on the growth metrics that hint at where the market is heading, Claude and Gemini are the ones moving. Claude’s year-over-year user growth has been reported at roughly 640%, against about 62% for ChatGPT — different bases, but a clear signal of trajectory. Gemini’s share has roughly quadrupled in twelve months on several trackers. The scoreboard ranks the present; the growth rates sketch the next two years. Reading only the ranks, and ignoring the slopes, is how incumbents get surprised.

Gemini’s rise is mostly a distribution victory

The cleanest way to understand Gemini’s climb to 27.7% is to stop thinking about it as a chatbot and start thinking about it as a distribution layer. Google is not winning users by convincing people to download and choose Gemini; it is placing Gemini in front of people who never made a choice at all. That difference is the entire story, and it is the part of the market that pure model quality cannot touch.

Consider the surfaces Google controls. Gemini is built into Android, which runs on billions of active devices. It is woven into Google Search through AI Overviews, which the company says reach on the order of two billion monthly users. It sits inside Chrome, Gmail, Docs and the broader Workspace suite that hundreds of millions of people open for work every day. It powers Circle to Search across hundreds of millions of Android phones. Each of these is a place where a user encounters Gemini-generated output without ever opening a standalone app or forming an opinion about which assistant is best. For a standalone product like ChatGPT, every user is acquired. For Gemini, a large share of users are simply present.

The numbers reflect that structural advantage. Google’s Gemini app climbed from around 350 million monthly active users in early 2025 to roughly 750 million by the end of 2025, and Sensor Tower’s count put it near 662 million by late May 2026 on its deduplicated basis — the exact figure varies by method, but the curve is steep on all of them. On web-traffic trackers, Gemini’s share of generative AI chatbot traffic roughly quintupled in a year, from the high single digits to over 25%. That is among the fastest share expansions the category has ever seen, and it was achieved primarily through placement rather than persuasion.

Geography sharpens the picture. Gemini is strongest precisely where Android dominates: it leads ChatGPT in India, and it has built heavy adoption across Southeast Asia, where mobile-first users live inside the Google ecosystem. Its relative weak spot is Western Europe and the UK, though that gap has been narrowing as Workspace integrations deepen. The pattern is consistent with the distribution thesis — Gemini does best where Google’s pipes are widest, not where its model is provably superior.

Two developments could widen the moat further. The first is the Apple partnership: a multi-year deal, reported to be worth on the order of a billion dollars a year to Google, under which a more capable Siri and Apple’s next-generation foundation models would be built on Gemini. If it ships on schedule, it could put Gemini-powered intelligence in front of more than two billion Apple devices, which would be the single largest AI distribution event since ChatGPT launched. It is also non-exclusive — Apple’s existing ChatGPT integration is unchanged for now — so the deal expands Gemini’s reach without immediately evicting OpenAI from the iPhone. The second is pricing. Google trimmed its cheapest paid tier to around five dollars a month and pushed lower-cost plans aggressively, while reportedly cutting its premium tier sharply, using its balance sheet to make paid Gemini cheap in a way standalone rivals struggle to match.

There is a real business underneath the reach, too. Google has reported on the order of eight million paid Gemini Enterprise seats across thousands of companies, with paid seats growing at a brisk quarterly clip, and millions of developers building on its generative models. It also cut Gemini’s serving costs dramatically over the course of 2025, which makes scaling the free tier far less painful. Distribution gets users in the door; falling unit costs and enterprise seats are what turn that traffic into a durable business.

The caution worth attaching is that reach is not the same as preference. Many people who “use Gemini” are encountering it passively inside Search or Android, and passive exposure is a weaker signal than someone who actively opens an app to ask it something. Google’s challenge is converting ambient presence into genuine habit and loyalty. But even with that caveat, the lesson for the market is hard to dodge: in a fragmenting AI market, owning the surfaces where people already are beats owning the best model nobody is placed to reach.

Claude inverts the share story on revenue

If the headline metric is user share, Claude looks like a distant third. On almost every metric that pays for the business, Claude looks like something else entirely — and that inversion is the most important nuance the Sensor Tower report surfaces. Anthropic built a smaller audience that is worth far more per person, and in 2026 that strategy started showing up in the numbers that investors actually price.

Start with revenue per user. Sensor Tower found that Claude’s average revenue per user on mobile in the United States climbed from under $0.50 in September 2025 to $2.76 by May 2026. Reporting that drew on the fuller dataset put that figure at roughly 1.5 times ChatGPT’s $1.74 on the same basis. A product with a fraction of ChatGPT’s users is extracting more money from each one. The same pattern shows up in conversion: about 13% of Claude’s users pay for a subscription, the highest conversion rate of any major assistant, against around 8% for ChatGPT. Claude also charges more at the entry point — its paid tier starts around twenty dollars a month, where ChatGPT’s cheapest paid plan is far lower — and still converts better, which tells you its users are there for work they will pay for rather than casual curiosity.

The deeper inversion is structural, and it is a deliberate choice rather than an accident. OpenAI built ChatGPT for everyone: free tier, viral growth, the broadest possible base, monetization figured out later. Anthropic pointed Claude at enterprises — the CIOs, legal teams and engineering orgs, not the students and casual users. A consumer might pay twenty dollars a month; an enterprise might pay a million dollars a year for a custom deployment. Anthropic has reported serving more than 300,000 business customers, with the large majority of its revenue — by various accounts around 80% or more — coming from business rather than consumers. It has said its number of large accounts grew nearly sevenfold year over year, that hundreds and then over a thousand customers now pay more than a million dollars annually, and that a striking share of the largest companies on earth are paying customers, including most of the Fortune 100 and the bulk of the Fortune 10.

That mix produced a genuine milestone for the category. Multiple reports, citing Reuters and others, indicate Anthropic’s annualized revenue run-rate reached roughly $30 billion by April 2026, up from about $1 billion at the start of 2025 — a pace of organic scaling that analysts struggled to find any precedent for in any industry. A large part of that engine is Claude Code, the company’s coding and agent product, which reportedly crossed a $1 billion run-rate within months of launch and roughly $2.5 billion by early 2026. By several accounts, Anthropic’s enterprise spending share overtook OpenAI’s for the first time in the spring of 2026, ending a long lead.

Retention is the last piece of the inversion, and it is closing in the direction that should worry OpenAI most. ChatGPT still retains new sign-ups better — reported around 86% against Claude’s roughly 73.7% — but the gap is narrowing, and on the revenue and conversion axes Claude is already ahead. Engagement intensity is rising too, with Claude’s time-per-user climbing meaningfully over the year.

The honest caveats: several of the sharpest enterprise figures come from company disclosures and secondary reporting rather than audited filings, and Anthropic itself has flagged that some recent profitability partly reflects favorable, possibly temporary, compute terms. The growth is real but young. Still, the strategic lesson is clear and it reframes the whole share debate. User share answers “who is biggest.” Revenue per user, conversion and enterprise depth answer “who is building the most durable business” — and on those, the third-place product is arguably leading.

The smaller assistants and the markets they actually own

The four names sitting under 5% are easy to dismiss as also-rans, and that would be a mistake. None of them is trying to be ChatGPT, and several of them dominate a corner of the market that the leaders do not serve well. Sub-5% global share can coexist with category leadership in a specific use case, a specific region, or a specific audience — and in a fragmenting market, those corners are exactly where defensible positions are built.

Grok, from Elon Musk’s xAI, is the clearest example of a product built on a distribution shortcut and a sharp audience skew. It rode the X platform’s existing user base from a near-standing start to tens of millions of monthly users — one widely cited internal figure put it around 64 million monthly active users in late 2025. Its demographic tilt is unusual: Sensor Tower found Grok users were roughly four times more likely than the average assistant user to be active crypto traders, the widest demographic skew measured for any AI assistant. Grok’s 2026 has also been turbulent. A deepfake scandal late in 2025 led to country-level scrutiny and a period where Apple privately threatened to pull the app, a dispute that eased after xAI added guardrails. Its US mobile share even ticked down month over month in early 2026 after a long run of growth. Grok shows both the power and the fragility of growth built on a single platform and a provocative brand.

Perplexity occupies a different niche by design. It is not a general assistant but an answer engine — a search product that reads sources and synthesizes a cited response instead of returning ten blue links. That focus keeps its user base smaller, around 34 million monthly active users, but unusually high-intent: a large share of its users are senior professionals and high-income knowledge workers. Perplexity’s revenue and valuation have climbed steeply, reportedly past $450 million in annualized revenue at a valuation north of $20 billion, backed by investors including SoftBank and Jeff Bezos. Notably, Perplexity abandoned advertising in early 2026 to go subscription-only, a bet that its professional audience values clean, trustworthy answers more than it values a free, ad-supported experience — the opposite of OpenAI’s direction.

DeepSeek is the geography story. The China-based assistant built a large footprint — by various counts somewhere between 60 and 125 million monthly users — concentrated heavily in Asian markets, and it is known for releasing open models that developers can run themselves. On worldwide web-traffic measures it punches above its global-share ranking, around 4% of chatbot web visits, but only about 1% in the United States, which reflects how regionally concentrated it is. Its infrastructure overlaps with ByteDance’s crawling pipeline, and ByteDance’s own assistant, Doubao, cracked the top ten globally in 2026. For anyone with an audience in Asia, the leaders that matter are not always the ones that dominate Western rankings.

Meta AI is the embedded player. It does not live in a destination app people deliberately open so much as inside Facebook, WhatsApp and Instagram, where billions already spend time. That makes its usage hard to measure cleanly — some analysts exclude it precisely because in-app AI use is hard to separate from ordinary app behavior — but it gives Meta enormous latent reach to activate. Microsoft’s Copilot, meanwhile, leans on Microsoft 365’s deep enterprise penetration; its standalone consumer web app is small, but its workplace footprint through Office is substantial.

The throughline across all five is that the long tail is not weak; it is specialized. Grok owns real-time and a particular subculture, Perplexity owns cited research, DeepSeek owns parts of Asia and the open-model crowd, Meta AI owns ambient social reach, and Copilot owns the Office desktop. A strategy that only accounts for ChatGPT, Gemini and Claude will systematically miss audiences that these products serve better — which, in a market defined by fragmentation, is an increasingly expensive blind spot.

The Pentagon deal that turned brand trust into a usage metric

The single most revealing finding in the Sensor Tower report is not a share number. It is the discovery that a values-based controversy produced a measurable, dated spike in user behavior — proof that trust is now a competitive variable you can see in the data, not a soft branding concept. The event that demonstrated it was OpenAI’s defense contract, and the episode is worth reconstructing because it reset assumptions about what drives switching.

In late February 2026, the U.S. Department of Defense — referred to in some reporting as the Department of War — pressed AI companies to agree that their models could be used for “any lawful purpose,” which would have included applications Anthropic had drawn hard lines against. Anthropic’s CEO Dario Amodei said publicly that the company “cannot in good conscience accede” to terms that, in its reading, would permit mass domestic surveillance and fully autonomous weapons. The response was swift: the department designated Anthropic a supply-chain risk to national security and barred contractors from doing business with it, and the administration directed federal agencies to stop using Anthropic’s technology, with a transition period. It was a designation that, by Anthropic’s account, had never before been applied publicly to an American company.

Days later, OpenAI announced its own classified deployment agreement with the same department. OpenAI framed it as more tightly guarded than prior arrangements, citing red lines against mass domestic surveillance, autonomous weapons control and high-stakes automated decisions. The optics were difficult regardless of the wording: the Pentagon had just rejected a rival’s stricter terms, and OpenAI stepped into the gap. The consumer reaction was immediate and, crucially, quantifiable.

Sensor Tower detected a sharp surge in ChatGPT uninstalls in the United States — reported in the range of roughly 200% to nearly 300% above the normal baseline depending on the source and window — alongside a jump in negative reviews. Claude was the direct beneficiary: its US downloads spiked, and for a stretch it out-downloaded ChatGPT for several consecutive days and briefly became the number-one app in the country. A grassroots campaign to abandon ChatGPT circulated online. This was not a slow erosion; it was a dated, visible event where a contract decision moved install and uninstall curves within days.

OpenAI moved quickly to contain it. CEO Sam Altman acknowledged the rollout “looked opportunistic and sloppy” and said the company “shouldn’t have rushed.” Within days OpenAI amended the contract to explicitly prohibit intentional domestic surveillance of US persons and to extend that ban to commercially acquired personal data, and said intelligence agencies were excluded. The fallout did not stop at metrics: a senior OpenAI leader, its head of robotics and consumer hardware, resigned over the decision, writing that surveillance without judicial oversight and lethal autonomous weapons without human authorization were lines that deserved more deliberation than they received.

The reason this matters for the share story is that it severs a long-held assumption. For years the working theory was that users choose assistants on capability — whichever model is smartest, fastest, most useful wins. The Pentagon episode showed something else: a non-trivial slice of users will uninstall a more popular, more capable product over a values disagreement, and they will do it fast enough to register in the data. Features did not change that week. Trust did. And trust moved behavior.

There is nuance worth keeping. The uninstall spike was a US phenomenon concentrated in a politically engaged segment, not a global collapse, and ChatGPT’s overall trajectory kept climbing. The point is not that one contract reversed OpenAI’s lead. It is that the report identified brand trust and values alignment as genuine, measurable drivers of switching — which means AI labs can no longer treat ethical positioning as pure public relations. It is now an input to the growth curve.

Values alignment as a measurable competitive variable

The Pentagon episode is worth pulling out from the narrative because it points to a broader, durable shift in how this market works. For the first time, an AI company’s stated principles produced a visible movement in its usage curve — and once something becomes measurable, it becomes a lever that competitors and customers will deliberately pull.

The mechanism is simple but new in scale. A generative assistant is unusually intimate compared with most software. People type their health worries, their work problems, their half-formed ideas and their private questions into it. That intimacy raises the salience of who they are trusting. A spreadsheet program signing a defense contract would barely register with most users; an assistant people confide in doing the same thing reads very differently. The closer a product sits to a user’s inner life, the more its values become part of the product.

This changes the calculus for every lab in three concrete ways. First, positioning is now a retention feature, not just a marketing message. Anthropic’s safety-first identity, which once looked like a slower, more cautious path, converted directly into downloads and goodwill during the controversy. The company has leaned into that posture deliberately, launching a transparency hub to document model reports and voluntary commitments. Trust became a moat the moment it could be shown to move users.

Second, large, visible partnerships now carry reputational beta. A government or defense deal, a data-sharing arrangement, a content licensing fight — each is no longer a contained business decision but a potential trigger for a segment of users to leave. The bigger the platform, the larger the exposed surface. OpenAI’s billion users mean any controversial choice is amplified across a billion relationships, which is why Altman’s quick, public walk-back was a rational response to a quantified risk rather than mere contrition.

Third, the effect is asymmetric and segment-specific. The users most likely to switch over values are not a random cross-section; they skew toward engaged, higher-information, often higher-spending professionals — exactly the cohort whose subscriptions and enterprise budgets matter disproportionately. Losing 2% of casual free users hurts far less than losing 2% of the engaged base that converts to paid and influences workplace adoption. A values event that drives away the high-intent segment can dent revenue and enterprise momentum out of proportion to its raw user count.

The cautionary reading is that values-driven switching can be loud without being permanent. Some users who uninstalled in protest drifted back once the immediate news cycle passed and the contract was amended. Outrage is a spike; habit is a trend. A single controversy rarely rewrites a market on its own. But the existence of the effect changes behavior preemptively — labs now weigh reputational consequences into product and partnership decisions because they have seen the data move. That is the lasting outcome: not that one deal cost OpenAI its lead, but that the entire industry now treats values alignment as a real input to the growth model rather than a nicety.

From a growth land grab to a monetization race

The deeper structural message in the State of AI 2026 report is that the generative AI market is changing phase. The land-grab era — get as many users as possible, as fast as possible, and worry about money later — is giving way to a phase where revenue efficiency, conversion and retention matter as much as raw acquisition. The data points all push in the same direction.

Spending is the headline. Sensor Tower expects consumers to spend more than $4.2 billion on AI apps in the first half of 2026, up from about $1.83 billion in the same period a year earlier, with in-app purchase revenue alone on track to surpass $4 billion, a rise of around 36% over the prior six months. That is a market learning to charge. At the same time, the report notes that both download growth and spending growth have been decelerating in percentage terms even as absolute numbers climb. A maturing market grows in dollars while its growth rate cools — exactly the signature of a category moving from frontier to infrastructure.

The shift toward premium, utility-driven subscriptions is the throughline. Average revenue per user has been rising industry-wide, and the products pulling ahead are the ones converting users into payers rather than merely accumulating signups. This is why Claude’s 13% conversion and rising ARPU drew so much attention: in a monetization phase, the ability to turn attention into recurring revenue is the scarce skill, and it is a different skill from going viral. OpenAI is enormous on attention; the open question its IPO will test is how efficiently it converts that into durable margin.

There is a hard economic reason the phase is shifting now, and it is cost. Running frontier models at a billion-user scale is brutally expensive. Inference — the cost of actually answering each query — runs into the billions of dollars a year for the largest players. As long as capital was cheap and growth was the only scoreboard, labs could subsidize free usage indefinitely. As the cost of serving free users bites and public-market scrutiny looms, “monetize or subsidize” stops being a choice that can be deferred. That pressure is visible in OpenAI reportedly weighing price cuts to defend share, in Google using its balance sheet to make Gemini cheap, and in the rush to add advertising and commerce revenue alongside subscriptions.

The competitive implication is that the metrics analysts and investors weight are being reweighted in real time. A year ago, the dominant question was “how many users do you have.” Increasingly the questions are “what is your revenue per user,” “what is your conversion rate,” “how much does it cost you to serve a query,” and “how sticky is your enterprise base.” Scale still matters enormously — a billion users is a billion users — but scale alone no longer settles the argument about which business is healthiest. That reframing is arguably more consequential for the industry’s future than the specific 46.4% figure, because it changes what every player optimizes for next.

The advertising experiment running inside ChatGPT

The clearest sign that the monetization phase has arrived is that ChatGPT now shows ads. OpenAI began testing advertising inside the product in February 2026, and by May, Sensor Tower estimated that an average of around 17% of daily users were being served ads. The company has scaled both the volume of ads and the share of users who see them gradually, which is the behavior of a company running a careful pilot rather than flipping a switch.

The categories tell you what kind of ad business this is becoming. Software and shopping are the largest advertiser categories inside ChatGPT so far, followed by media and entertainment and food and dining. That mix points squarely at commerce and high-intent decisions — people asking the assistant what to buy, which tool to use, where to eat — rather than brand awareness. It is closer to search advertising and retail media than to the display advertising that funds social feeds, which is a meaningful distinction for how lucrative and how defensible it could become.

The playbook is borrowed openly. OpenAI’s monetization leadership has described testing sponsored placements that are kept “separate and clearly distinct” from the assistant’s organic answers, and the structure echoes the way Amazon built a multi-billion-dollar advertising business starting from simple sponsored product listings more than a decade ago. Retail media networks are positioning themselves as the bridge into AI advertising: Target, for instance, ran a pilot placing contextual ads inside ChatGPT through its retail media network, triggered by keywords in user prompts and labeled as ads, promoting both its own products and brands that pay through its network. That is retail media migrating from the open web into the conversation.

For OpenAI, advertising solves a specific problem. Subscriptions alone may not cover the enormous cost of serving a billion users, many of them free. Ads let the company monetize the free tier — the vast majority of its base — without forcing those users to pay. An ad-supported free tier plus paid subscriptions plus enterprise contracts is a three-legged revenue model, and it is the same shape that made Google’s economics work at scale. The risk OpenAI is managing is trust: an assistant people confide in is a fragile place to insert commercial messages, and getting the line between helpful recommendation and paid placement wrong could erode the very intimacy that makes the product valuable.

The strategic contrast across the market is sharp and worth holding onto. OpenAI is moving toward ads. Google already runs the largest advertising business on earth and can fold AI into it. Anthropic has stayed out of consumer advertising entirely, monetizing through enterprise and high-conversion subscriptions, and Perplexity actively abandoned ads to go subscription-only. Three different philosophies are being tested simultaneously: monetize free users with ads, monetize through ecosystem and existing ad infrastructure, or refuse ads and charge for value. Which one wins is not yet decided, and it may be that different models suit different audiences — ads for the broad consumer base, clean subscriptions for the professional segment that will pay to avoid them.

For marketers, the immediate takeaway is that a new high-intent advertising surface is opening inside the assistant where a growing share of product research now happens. It is early, the inventory is limited, and the rules are still being written. But the direction is unmistakable: the place people increasingly ask “what should I buy” is becoming a place where the answer can be, in part, paid for. Anyone selling a product will eventually have to decide how to show up there.

AI assistants becoming the new front door to shopping

Underneath the share and revenue numbers sits a quieter transformation that may matter more to the broader economy than any of them: AI assistants are turning into the place where product discovery begins. The Sensor Tower report and a wave of related data show shopping moving into the chat window, and the implications ripple out to every retailer and brand.

The scale of shopping intent inside these tools is already large. Research involving OpenAI and a Harvard economist found that roughly 2% of all ChatGPT queries involve shopping — on the order of 50 million shopping-related queries per day — against a backdrop of well over two billion prompts flowing through ChatGPT daily. Two percent sounds small until you multiply it by the base. Surveys reinforce the behavior: by various counts, close to 60% of US consumers have used generative AI for help with online shopping, and a meaningful and growing share of consumers — higher among Gen Z — say they are comfortable letting AI agents shop on their behalf.

Retailers can already see it in their referral logs. ChatGPT grew from roughly one in five of Walmart’s referral clicks in mid-2025 to around 36% by late 2025 on some measures, and it drives a double-digit share of referrals to Target, Etsy and eBay, with Target reporting that traffic from ChatGPT to its site grew around 40% month over month. These referral clicks are still a small fraction of total site visits, so the absolute volumes remain modest compared with established search and paid channels. But the growth rate is steep, and the intent quality is high — a person asking an assistant for a recommendation is deeper into a buying decision than someone idly browsing.

The infrastructure is being built to capture that intent directly. OpenAI has rolled out an Instant Checkout capability and struck partnerships that let people buy from Walmart, Etsy and Shopify merchants without leaving the chat, and announced plans to bring Target’s storefront inside ChatGPT. Agentic features — assistants that can complete multi-step tasks like booking or purchasing on a user’s behalf — are moving from demo to early product. The shopping journey is collapsing from “search, click through to a site, browse, check out” into “ask, get a recommendation, buy in the conversation.” That compression is the structural change, and it threatens the businesses built on the old, longer journey.

There is hard evidence the assisted model converts. Amazon, with its own shopping assistant, reported that shoppers who use it are markedly more likely to complete a purchase — by some measures converting at nearly twice the rate of non-users, with assisted sessions sustaining far higher conversion than unassisted ones. When an assistant narrows the options and answers the buyer’s real question, it removes friction that kills sales. That is why every major retailer is now forced to have an AI strategy, whether that means opening their catalog to outside assistants or building their own.

The fragmentation theme reappears here in a way that bites for marketers. Because discovery is splitting across ChatGPT, Gemini, Perplexity and retailer-owned assistants — each pulling from different sources and surfacing products differently — a brand can no longer optimize for one storefront or one search engine and expect to be found. The shopper might start in ChatGPT, which leans on third-party reviews and editorial sources; or in Gemini, wired into Google’s shopping and search graph; or inside a retailer’s own assistant. Showing up requires presence across all of them, structured so each can understand and recommend you. The front door to shopping did not just move; it multiplied.

Amazon’s walled garden and the open rail split

The shopping shift has forced the clearest strategic fork in the retail world, and it is worth understanding because it previews how every data-rich platform may respond to AI agents. Amazon has chosen to wall itself off; Walmart, Target, Etsy and others have chosen to open up. The industry has started calling these the closed rail and the open rail, and the split is not a detail — it is a bet on two different futures.

Amazon’s move was defensive and deliberate. Beginning in late 2025, the company updated its robots.txt and site code to block major AI crawlers — including OpenAI’s ChatGPT-User and search bot, as well as crawlers from Google, Meta, Anthropic and Perplexity — from scraping its product data. The practical effect is that ChatGPT cannot reliably surface real-time Amazon product links, which nudges users toward rival retailers. After the block, referral traffic from ChatGPT to Amazon reportedly fell sharply, dropping by roughly 18% to under 3% on a month-over-month basis. By one analyst’s estimate, Amazon’s absence effectively pulled hundreds of millions of product listings off the “agentic shopping shelf.”

The logic is money. Amazon’s marketplace is the world’s largest store of e-commerce data and the backbone of an advertising business worth around $56 billion a year, built entirely on shoppers browsing Amazon’s own site. Letting outside assistants surface its products and complete purchases elsewhere would risk bypassing the storefront, undermining both traffic and that enormous ad business. So Amazon is doubling down on its own assistant instead — embedding it in search, populating its interface with ads, and reporting that shoppers who use it are far more likely to convert and that it expects the tool to drive many billions in annual sales. Amazon’s wager is that it can keep shoppers inside its ecosystem by making its own AI good enough that they never need an outside one.

Walmart, Target and the others made the opposite calculation. Rather than fight to keep shoppers from starting in ChatGPT, they decided to be findable there — opening their catalogs, integrating with Instant Checkout, and treating outside assistants as a new discovery channel to be courted rather than blocked. Walmart, with its hundreds of millions of SKUs, now fills more of the surface area inside AI chat results precisely because Amazon stepped back. The reasoning is that if the customer’s journey is going to start in an assistant, the retailer wants its products in the answer, even at the cost of ceding some control over the experience.

Both strategies can work, and they imply genuinely different worlds. In the closed-rail future, every major retailer builds its own walled assistant and keeps the agent inside its own gates — a world that favors Amazon, with the data and scale to make its assistant compelling. In the open-rail future, general-purpose assistants become shopping routers that send traffic to whichever retailers make their catalogs legible and easy to transact with — a world that gives Walmart, Target and category specialists a cleaner shot at intercepting demand that once flowed automatically to Amazon.

For brands and sellers, the split creates a concrete, immediate problem. A product that is invisible to outside assistants because it only lives on a walled platform misses the audience that starts its search in ChatGPT or Gemini — tens of millions of high-intent shoppers a day. Conversely, a product optimized only for outside assistants may underperform inside the retailer-owned tools where a large share of buyers still begin. The defensible position is presence on both rails, with product data structured so any assistant — open or closed — can understand, trust and recommend it. The retailers have drawn their battle lines; the brands now have to live across them.

The engagement gap that share numbers hide

User share and user counts both miss something that the Sensor Tower data captures and that arguably matters more for the health of these products: how much time people actually spend with them. Engagement is where the leaders’ dominance looks even more entrenched than the share figures suggest, and where the real moat may sit.

The category-level number is striking. Sensor Tower projects that time spent on generative AI apps will more than double year over year, from about 17.2 billion hours in the first half of 2025 to roughly 36 billion hours in the first half of 2026. That is not a market checking in occasionally; it is a market doing serious, sustained work. Total sessions are projected to reach into the hundreds of billions, orders of magnitude above where the category sat two years earlier. People are not just trying these tools; they are living in them for research, writing, coding, analysis and decision-making.

Concentration at the top is the part that complicates the fragmentation story. By Sensor Tower’s count, a small handful of leaders — ChatGPT, DeepSeek and Gemini — accounted for close to 90% of all time spent across AI assistant apps in early 2026. That is worth stating carefully, because it is often paraphrased imprecisely: the time-spent concentration names DeepSeek among the top three, reflecting how much sustained usage its large Asian base generates, even though the three biggest products by global users are ChatGPT, Gemini and Claude. The lesson is that the market is fragmenting at the margins while remaining intensely concentrated at the core. More products are taking share, but the bulk of human attention still pools into a few.

Per-user intensity sharpens the picture further. ChatGPT users spend on the order of 215 minutes per month with the product, by Sensor Tower’s measure — well ahead of the others, even as its relative share slips. Gemini and Claude both saw their per-user time climb meaningfully over the year, with Claude’s engagement intensity rising notably as its users do more substantial work inside it. High and rising time-per-user is a stickiness signal that raw share misses entirely. A product can lose relative share while deepening the engagement of the users it keeps, which is a far stronger position than a shrinking, disengaging base.

The web mirrors the app pattern. Generative AI websites generated tens of billions of visits and tens of billions of hours of usage in early 2026, and crucially, time spent grew faster than visit counts — visits up around 28% year over year, time up around 41% on Sensor Tower’s web data. When time grows faster than visits, it means people are not just dropping by; they are settling in to work. That is the difference between a novelty and a habit, and habits are what defend market positions over the long run.

The practical reading for operators is that engagement depth changes how to think about reaching users. A platform where people spend hundreds of minutes a month is a platform where your brand, product or content has many chances to surface — and where being absent means being absent from a large and growing share of a person’s working day. Share tells you how many people use a product. Engagement tells you how much of their attention it commands. For anyone trying to be present in that attention, the second number is the one that should drive the budget.

Two IPOs racing toward the same narrow window

The market-share data landed in the middle of an extraordinary financial event, and the two stories are inseparable. OpenAI and Anthropic are both racing toward public listings in the same window, which means every share, revenue and retention figure is now being read by investors trying to price two of the largest technology offerings ever attempted. The numbers stopped being trivia the moment the S-1s started moving.

OpenAI confirmed on June 8, 2026 that it had confidentially filed a draft S-1 registration statement with the SEC — the first formal step toward going public. The company announced it itself, unusually, in a blog post that read, in effect, that it expected the filing to leak so it was simply getting ahead of it. OpenAI stressed that timing was undecided and that an actual listing “may be a while,” because some things are easier to do as a private company. Reporting has pointed to a possible window of roughly September to November 2026, with underwriters including Goldman Sachs and Morgan Stanley, and reportedly JPMorgan and Citigroup. OpenAI was last valued at around $852 billion in a March 2026 round, and some analysts expect an IPO valuation above $1 trillion, which would put it among the largest public offerings in history.

Anthropic moved first. It filed confidentially on roughly June 1, 2026, at a reported valuation near $965 billion, days ahead of OpenAI — turning the rivalry, which began over model capabilities, into a race to Wall Street. Both followed SpaceX, which ran its own roadshow and went public around June 12, 2026 at a valuation near $1.75 trillion, reaching roughly $2.1 trillion after its first day of trading. Taken together, analysts described an AI-adjacent IPO pipeline worth on the order of $3.6 trillion arriving in a single cluster — a concentration of mega-listings without real precedent.

The competitive subtext is that each company’s IPO narrative leans on a different half of the very data this report surfaced. OpenAI’s pitch is scale and consumer dominance: a billion-plus users, the fastest adoption curve ever, a brand that defined the category. Anthropic’s pitch is efficiency and enterprise depth: higher revenue per user, the best conversion rate, a commanding position among the largest companies, and a faster path to profitability. The 46.4% share figure cuts against OpenAI’s “we are the market” story; Claude’s revenue-per-user lead cuts in Anthropic’s favor. Public investors will effectively be asked to price the two competing theories of what makes an AI business valuable — reach or efficiency — at almost the same moment.

The timing also raises the stakes on everything else in the report. A values controversy that dents trust, a price war that compresses margins, an advertising experiment that either works or alienates users, a retention gap that widens or closes — each of these now feeds directly into how the market values two companies preparing to sell shares. Metrics that were once internal dashboards are about to become quarterly obligations. Going public means the share trajectory, the conversion rate and the cost of serving a query stop being strategy and start being disclosure, scrutinized every ninety days.

The honest caveat is that none of this is locked. Confidential filings are options, not commitments; both companies have stressed that timing could slip, and market conditions, regulatory review and the sheer size of the offerings could push dates. But the direction is set. The AI race has entered its public-markets phase, and the fragmenting share map is now part of the prospectus. For anyone watching the industry, that means the next year of competition will play out under a level of financial transparency the sector has never faced.

The economics underneath the trillion dollar valuations

Behind the trillion-dollar headline valuations sit two very different financial profiles, and the contrast explains why the market-share debate has become a referendum on business models. One company has the users and the losses; the other has the margins and a fraction of the reach. The numbers, drawn from reporting around the filings, make the divergence concrete.

OpenAI is a growth-at-scale story carrying heavy losses. The company has been reported as generating on the order of $2 billion in revenue per month in early 2026, with first-quarter revenue around $5.7 billion and a full-year target near $30 billion. But it is far from profitable. Reporting indicated a full-year 2026 loss forecast in the range of $14 billion, an adjusted operating margin deeply negative in the first quarter — by one account around -122%, meaning roughly $1.22 lost for every dollar of revenue — and inference costs alone in the range of $14 billion. The company reportedly does not expect positive cash flow until the end of the decade. Its revenue is growing several times faster than Alphabet’s or Meta’s did at comparable stages, and enterprise is now over 40% of revenue and climbing toward parity with consumer. The bull case is that scale plus enterprise growth eventually overwhelms the cost curve; the bear case is that the cost curve does not bend fast enough.

Anthropic is the inverse: smaller, leaner, and closer to profit. Its annualized revenue run-rate reportedly reached roughly $30 billion by April 2026, and reporting around its filing suggested a second quarter with revenue near $10.9 billion and operating income around $559 million — which would mark its first profitable quarter, though the company itself cautioned that the result partly reflected a favorable, possibly temporary, discount on a major compute deal. With a fraction of OpenAI’s users, Anthropic generates comparable revenue because it monetizes business customers who pay orders of magnitude more than consumers. A company with a few percent of consumer share producing enterprise revenue rivaling the category leader is the clearest possible illustration of where business value is actually concentrated.

Two AI leaders, two financial profiles (reported figures, 2026)

These figures come from secondary reporting around the companies’ confidential filings and funding rounds, not audited public statements; profitability claims in particular carry company-flagged caveats and should be read as directional.

The investor question underneath the table is whether the market rewards the model that has the users or the model that has the margins. OpenAI’s stake is ownership of the consumer relationship at unprecedented scale; Anthropic’s is capital efficiency and a customer base that does not churn casually. Both are plausible paths to a durable, valuable company, and they are not mutually exclusive — the consumer market is real and the enterprise market is real. What the share data and the financials together establish is that the once-dominant assumption — that OpenAI’s lead was insurmountable because it was biggest — has been tested against revenue and found incomplete. Biggest and healthiest are no longer obviously the same company, and the public markets are about to render a verdict on which one matters more.

A price war forming at both ends of the market

Falling share and rising costs have produced the predictable next move: prices are coming down, at both the consumer entry point and the enterprise tier. A market that once competed almost entirely on capability is starting to compete on cost, which is a classic signal that a category is maturing and that the incumbents feel the pressure.

The consumer signals are concrete. Google trimmed its cheapest paid Gemini plan to around five dollars a month and pushed lower-cost tiers around eight dollars, while reportedly cutting its premium tier sharply — using its enormous balance sheet and its falling serving costs to make paid Gemini cheap in a way standalone competitors cannot easily match. ChatGPT’s lowest paid plan sits well below Claude’s entry price. And reporting indicated that OpenAI was weighing significant price cuts to defend share against Anthropic, particularly on the enterprise side, where Anthropic’s commercial momentum has been strongest. When the largest player in a category starts contemplating price cuts, it is responding to competitive pressure, not generosity.

The enterprise front is where the war is fiercest, because that is where the money is. Anthropic’s gains among large companies — the high conversion, the seven-figure accounts, the enterprise spending crossover — have put real pressure on OpenAI’s corporate pricing. The battle for high-value business customers is just beginning, and it is being fought partly on capability and integration, but increasingly on price and total cost of ownership. Open-weight models add a third source of pressure: Meta and others offer models that companies can run themselves, which sets a floor under what the commercial labs can charge for commodity workloads.

Then there is the wildcard that could reset the consumer economics entirely: on-device AI. Apple has been preparing a new Siri that runs some AI directly on the device, with no subscription required. If a meaningful slice of everyday assistant tasks — quick questions, simple drafting, basic summarization — can be handled on the phone for free, it undercuts the premise that users must pay a monthly fee to a cloud assistant for routine help. That does not threaten the heavy, complex work that justifies a subscription, but it could commoditize the easy 80% of queries and force the cloud assistants to justify their price on the hard 20%.

The strategic consequence is a squeeze on the middle. Free or near-free options are improving from below — ad-supported tiers, cheap Gemini plans, on-device assistants — while the genuinely differentiated, high-value work migrates to premium and enterprise tiers above. The uncomfortable position is the mid-priced consumer subscription that is neither free nor clearly worth its cost for a casual user. Products will increasingly need to be either cheap and ad-supported at the bottom or unmistakably valuable at the top, with less room to charge a moderate fee for moderate value in between.

For users, the price war is mostly good news in the near term: more capability for less money, more free options, and competitive pressure keeping subscriptions honest. For the companies, it compresses the margins they need to fund staggering compute bills, which is exactly why monetization through ads, commerce and enterprise has become so urgent. Cheaper prices for users and a desperate hunt for revenue are two sides of the same coin — and both are direct consequences of a market where no single player can dictate terms anymore.

Regional fault lines in adoption and spending

The global headline share hides sharp regional differences, and those differences matter enormously for anyone deciding where to invest, which assistant to optimize for, or which market to enter. Adoption, spending and platform preference all vary by region in ways the single 46.4% figure flattens completely.

The most surprising regional data point is that Asia, long the engine of download growth, recorded a contraction. Sensor Tower noted Asia’s first download decline — around 3.3% in the first quarter of 2026 — driven by dips in China and India. That does not mean Asia is losing interest; the region still leads the world in total downloads. It means the early, explosive download phase is cooling in the markets that drove it first, a maturation signal rather than a retreat. The deceleration in those markets is part of why global download growth rates have softened even as absolute usage climbs.

Spending tells the opposite geographic story. North America and Europe generate higher in-app spending than Asia, even though Asia leads on downloads. This split is one of the most actionable facts in the whole report: the markets with the most users are not the markets with the most revenue per user. A company chasing raw user growth would prioritize Asia; a company chasing monetization would prioritize North America and Europe. The right answer depends entirely on which phase of the business you are optimizing, and the monetization phase the whole industry has entered tilts the calculus toward the higher-spending Western markets.

Platform preference fractures regionally too. Gemini leads ChatGPT in India and is strong across Southeast Asia, where Android’s dominance hands Google a structural advantage and where Google has invested in multilingual support. DeepSeek’s footprint is heavily Asian, concentrated in Chinese, Indian and Indonesian markets. ChatGPT remains strongest in Western markets and in the cohort of engaged professionals, while Claude’s user base skews heavily American. There is no single global assistant; there is a different leader, or a different competitive balance, in different parts of the world — which means a strategy built on the assumption that one product reaches everyone everywhere is wrong on two axes at once, by product and by geography.

Emerging markets are where the next phase of growth concentrates, and they reward the players with distribution. Regions across the Middle East, Africa, Latin America and South and Southeast Asia have shown rapid growth, much of it mobile-first, which favors assistants embedded in the devices and apps people already use. Distribution-led players like Gemini, and regionally tuned products like DeepSeek and Doubao, are positioned to capture growth in markets where standalone-app adoption faces more friction — a data plan, a download, a sign-up — than passive exposure inside an existing platform.

For operators, the regional reading converts directly into priorities. A brand or product with a Western, high-income audience should weight ChatGPT and the premium tiers, where spending and engagement concentrate. One targeting India or Southeast Asia cannot ignore Gemini’s lead and the local players. One operating in China faces an entirely separate ecosystem. The fragmentation is not only across products; it is across borders — and treating “AI users” as one undifferentiated global audience is a mistake that the regional data makes plain.

The web is fragmenting alongside the apps

The assistant market is not fragmenting in isolation. It is part of a broader restructuring of how people find information online, and the web-traffic data tells a parallel story to the app data. Generative AI has gone from an experiment to a core web behavior, and as it has, the single front door to the internet has multiplied into several.

The growth figures are emphatic. AI assistant traffic was the fastest-growing web category of the past year, surging on the order of 86% globally, and ChatGPT alone added more than 60 billion visits year over year to become one of the most-visited websites in the world. That is a website-scale phenomenon, not a niche tool. But the same data shows the growth spreading: Gemini, Claude, DeepSeek and others all expanded their web footprints, so even as ChatGPT grew enormously, its share of AI web traffic eroded — the same pattern visible in the app market, driven by the same dynamic of a category growing faster than its leader.

Two structural shifts in web behavior frame all of this. First, mobile crossed a threshold: for the first time, mobile accounted for more than half of all global web visits in early 2026, driven by mobile-first markets like India and Indonesia alongside rising mobile use in mature markets. Second, despite that, desktop still represented the large majority of total web time spent — over 70% — because longer, more serious work, including a lot of AI-assisted research and writing, still happens on a larger screen. The web is becoming mobile for discovery and desktop for depth, and assistants bridge the two: cross-platform AI usage more than doubled as people moved between phone and laptop within the same task.

The discovery layer is where the consequences land hardest. Search and social remain the dominant gateways to the internet, but AI assistants have emerged as a fast-growing third gateway, sitting alongside them rather than replacing them outright. People increasingly start a question in an assistant, get a synthesized answer, and may never visit the underlying websites at all. Sensor Tower’s web data showed travel-booking traffic from generative AI climbing sharply, retail referrals growing across categories, and time-on-AI-sites growing faster than visits — all signs that AI is becoming a genuine intermediary between people and the rest of the web.

This is the deeper reason the share number matters beyond OpenAI’s balance sheet. If AI assistants are becoming a primary way people discover information and products, then the fragmentation of that layer fragments discovery itself. When one assistant reached almost everyone, being visible there approximated being visible to the AI-mediated world. With more than half of users now spread across Gemini, Claude and a long tail, no single assistant is a complete channel. The discovery surface has the same shape as the app market: concentrated at the core, fracturing at the edges, and impossible to cover from a single position.

The practical implication for anyone who publishes, sells or markets online is that the open web’s traffic is increasingly mediated by systems that summarize rather than link. A growing share of queries resolve inside an AI answer without a click, which changes what “being found” means. The next sections turn to exactly that: what fragmentation changes for ordinary users, and what it changes for the marketers and businesses whose visibility now depends on showing up inside answers across a splintered set of platforms.

The practical effect of fragmentation on everyday users

Step back from the market structure and ask what the end of single-app dominance means for the person actually typing questions into these tools. For users, fragmentation is mostly good — more choice, lower prices, faster improvement — but it also introduces new friction that did not exist when one assistant did everything.

The most immediate change is that the single-tool habit no longer makes sense. For two years, “use AI” effectively meant “use ChatGPT,” and for many people it still does out of inertia. But the products have specialized enough that the best results increasingly come from matching the tool to the task rather than funneling everything through one. People who work seriously with these tools have largely stopped asking “which assistant is best” and started asking “which assistant is best for this,” because the honest answer is that it depends on what you are doing.

A rough division of labor has emerged from how the products are built and tuned. Claude has a strong reputation for writing, careful reasoning and long, complex documents, and is favored in work and regulated settings. ChatGPT is the broad generalist with the deepest ecosystem of third-party integrations and the widest feature set. Gemini’s advantages are its massive context window, its tight integration with Google’s tools, and its reach inside Search and Android. Perplexity is the tool of choice when you want a cited, sourced answer for research. None of these is a hard rule, and the models leapfrog each other constantly, but the pattern is stable enough to be useful: pick the tool whose strengths match the job.

The friction is real, though. Using several assistants means several subscriptions, several interfaces, several sets of quirks to learn, and context that does not follow you from one tool to another. A conversation, a set of preferences, or a project built up inside one assistant does not transfer to another. Switching has a cost that single-platform users never paid, and for casual users that cost may not be worth it — many will sensibly stick with one good-enough tool. The fragmentation benefits power users more than casual ones, because the payoff from picking the right tool scales with how much you depend on it.

Privacy and trust enter the user’s calculation in a way they did not before, and the Pentagon episode showed why. When an assistant becomes a place people share genuinely personal information, who owns it, what it does with the data, and what it stands for become part of the decision — not just how clever its answers are. Users now have a meaningful choice about which company to trust with their questions, and a fragmented market gives them somewhere to go if a provider does something they object to. That is a healthier dynamic than a monopoly where the only option is to accept or abstain.

The bottom-line shift for users is one of posture. The era of passively defaulting to one assistant is ending, and a little deliberateness now pays off — choosing tools by task, paying attention to who you are trusting, and not assuming the most famous product is the best one for what you need. The market did the fragmenting; the practical upside is that users have more and better options than at any point since the category began. The cost is that getting the most from them now takes a bit more thought.

Reach split across apps, web, embedding, and APIs

One reason these market-share debates never quite resolve is that “using an AI assistant” no longer means one thing. A person can touch the same underlying model through a standalone app, a website, a feature embedded in another product, or a developer-built application running on an API — and most trackers see only some of those surfaces. Understanding this split is essential to reading any usage claim honestly, because the surface an analyst measures determines the number they report.

Consider the surfaces in turn. There is the dedicated app or website, which is what most consumer rankings measure and where a lot of deliberate assistant use happens. There is embedding, where the model is woven into a product the user already opens — Gemini inside Google Search and Workspace, Copilot inside Office, Meta AI inside WhatsApp — and where the user may not even perceive a separate “assistant” at all. There is the API layer, where developers build the model into their own software, so the end user is interacting with some other company’s product that happens to run on Claude or GPT underneath. Each surface generates real usage, and they are measured by completely different instruments, if they are measured at all.

This is why the same product can look dominant or modest depending on the lens. Claude’s consumer web traffic is a fraction of ChatGPT’s, which makes it look like a minor player on web-visit rankings. But a large share of Claude’s actual usage happens through enterprise deployments and API integrations that never touch claude.ai — embedded in software, coding tools and internal company systems. A web-visits tracker cannot see any of that. The same is true in reverse for embedded products: Gemini’s reach through AI Overviews and Android dwarfs its standalone app, but a tracker counting only the Gemini app misses the ambient usage that is arguably its biggest advantage.

The metric mismatch compounds the confusion. OpenAI tends to report weekly active users; Sensor Tower reports monthly active users; web trackers report visits; DNS data reports resolution frequency. A weekly figure and a monthly figure are not the same test, and a visit is not the same as a unique user, and none of them captures API or embedded usage. When a headline says one product has “900 million” and another has “1.1 billion,” those numbers may be measuring different windows on different surfaces and may not be directly comparable at all. The honest move is always to ask: weekly or monthly, which surfaces, whose panel, which markets.

For developers and businesses building on these models, the surface split has a direct strategic consequence. The model your customers interact with may not be the one whose logo they see — and the leading model for embedded, API-driven work is not necessarily the leader in consumer popularity. The fact that Claude is third in consumer share but a heavyweight in enterprise and developer integrations means the “right” platform to build on depends on where your usage lives, not on which app tops the consumer charts. The consumer scoreboard and the builder’s scoreboard are different documents, and conflating them leads to bad bets.

The takeaway is a discipline as much as a fact: treat every reach claim as a claim about a specific surface measured a specific way. The 46.4% figure is a deduplicated, multi-surface, 25-market estimate of consumer assistant use. It is a good number for what it measures. It does not capture the embedded and API usage where some of the most valuable activity now happens — which is one more reason the simple ranking tells only part of the story.

Practical guidance for choosing and combining assistants

Most coverage of the share shift stops at the horse race. The more useful question for anyone who actually works with these tools is operational: given a fragmented market, how should a person or a team decide which assistants to use, and how to combine them without drowning in subscriptions and context-switching? The data points to a few defensible principles.

Start by matching the tool to the dominant task, not the other way around. If your work is writing-heavy, reasoning-heavy or document-heavy, the products tuned for careful, long-form work earn their place. If you need a generalist with the broadest set of features and integrations, the category leader is the safe default. If your work lives inside Google’s tools, the assistant embedded in those tools removes friction that a standalone product cannot. If you need sourced, cited research you can verify, the answer-engine products are built for exactly that. The point is to map your highest-frequency, highest-stakes tasks to the products whose strengths align, rather than forcing everything through whichever app you opened first.

Be deliberate about how many tools you actually run. For most individuals, one strong primary assistant plus one specialist for a recurring need is a sensible ceiling. A common, workable pattern is a capable generalist for the bulk of daily work and a second tool for the one thing the generalist does worst — sourced research, or a specific integration, or work that benefits from a very large context window. Running four assistants for the sake of completeness usually costs more in subscriptions and mental overhead than it returns. The fragmentation rewards selective use, not maximal use.

For teams, the calculus is different and more consequential. A team should standardize enough to keep workflows coherent, while leaving room for specialists where the work demands it. Picking a primary enterprise platform matters for security, data handling, billing and training, and the enterprise-grade products differ meaningfully on compliance, data retention and administrative controls — which is part of why Claude has done well in regulated industries and why Copilot wins where Microsoft 365 is entrenched. The right enterprise choice is rarely the most famous consumer product; it is the one whose security posture, integration story and support match the organization’s actual requirements.

Treat capability rankings as perishable. The models leapfrog each other constantly, so any “best model” judgment has a short shelf life, and building a rigid workflow around one provider’s current lead is fragile. The more durable approach is to build processes that are reasonably portable — clear prompts, documented workflows, exportable outputs — so that switching or adding a tool later is cheap. Lock-in to a single provider’s proprietary features is a real cost in a market moving this fast.

Mind the things that do not transfer. Conversation history, saved preferences, custom instructions and project context generally do not move between assistants, so spreading work across many tools fragments your own accumulated context. If continuity matters to how you work, that is an argument for consolidating more than the raw capability rankings would suggest. The convenience of one assistant that knows your history can outweigh another’s marginal quality edge on a given task.

Finally, factor in trust and data handling as first-class criteria, not afterthoughts. Who you are willing to share sensitive information with is a legitimate input — for personal privacy, and for businesses, for compliance and liability. The Pentagon episode showed that values and data practices now move users; for an organization handling regulated or confidential data, those practices are not optional considerations but core requirements. The honest summary is that there is no single right answer for everyone. The fragmented market means the best setup is the one matched to your specific tasks, your tolerance for managing multiple tools, and your requirements around trust and data — and revisiting that setup periodically, because the ground keeps moving.

A multi-platform market rewrites discovery and marketing

For marketers, founders and anyone whose livelihood depends on being found, the 46.4% figure is not a tech-industry curiosity. It is a direct instruction to change strategy. The single most important consequence of ChatGPT losing its majority is that optimizing for one AI platform no longer reaches the AI-using audience — and the discipline built to address that, generative engine optimization, has shifted from optional to essential.

The mechanics of discovery have genuinely changed, and the change is structural rather than cosmetic. Traditional search optimization was about earning a high rank among ten blue links, where the user still chose which to click. AI-mediated discovery is about earning a place among the two to seven sources a model cites or synthesizes in a single answer — and often the user consumes that answer without clicking anything at all. When an assistant names your brand, product or statistic in its response, it delivers an implicit endorsement no organic listing ever could; when it does not, you are invisible to that query regardless of how well you rank on Google. The prize and the penalty are both larger and more binary than in classic search.

Fragmentation makes this exponentially harder, which is the whole point. When one assistant dominated, being cited there approximated being visible to the AI-using world. Now the audience is split across ChatGPT, Gemini, Claude, Perplexity and others, each of which retrieves and cites from different sources and weights them differently. ChatGPT leans heavily on certain reference and community sources; Gemini is wired into Google’s search and shopping graph; Perplexity foregrounds sources it can cite cleanly. Optimizing for the citation behavior of one does not guarantee visibility in the others. A brand that “won” inside ChatGPT can be absent from the Gemini-powered answers a growing share of users now see, and vice versa. Single-platform optimization is a partial strategy in a market this fragmented, and the gaps are where competitors get cited instead of you.

The evidence base behind effective practice is converging, and it rewards substance over tricks. Academic work that coined the term — research out of Princeton and collaborators — found that content optimized with cited sources, concrete statistics and direct quotations measurably improved the rate at which AI engines surfaced it, by meaningful margins. Industry data consistently shows AI systems favoring authoritative third-party sources and earned media over self-promotional brand content, and that community and reference platforms — sources like Reddit, established encyclopedic references and professional networks — are among the most-cited domains across the major assistants. The practical reading is that AI visibility is earned through genuine authority, structured clearly, more than through keyword manipulation — which, for credible operators, is a more level playing field than classic SEO ever was, because expertise and trust count for more than domain age or backlink volume.

There is a measurement gap that compounds the urgency. A large majority of brands still do not track their visibility inside AI answers at all — by some counts well over 80% — even as a rising share of marketers plan to create content specifically for AI citation and a growing share of decision-makers have allocated budget to AI search optimization. The teams that can see their AI visibility, across multiple platforms, have an advantage over the many who are flying blind. Zero-click behavior is rising sharply, especially inside AI-mode experiences, which means impressions, mentions and brand authority increasingly matter more than raw click counts — a hard adjustment for teams trained to optimize clicks above all.

The strategic conclusion writes itself from the data. Discovery has moved from “rank on Google” to “be present, cited and recommended across a fragmented set of AI surfaces as well as traditional search.” Some practitioners call the broader idea search-everywhere optimization; the AI-specific core of it is being referenced by the assistants where your audience now asks questions. The brands that build this discipline now — across platforms, grounded in real authority, measured properly — will be the ones AI systems cite in the years ahead, because citation authority, like the domain authority that preceded it, compounds over time. The window where most competitors have not started is the opportunity.

Building a GEO plan that survives fragmentation

If a multi-platform market makes single-platform optimization insufficient, the obvious next question is what a durable plan actually looks like. A generative engine optimization plan that survives fragmentation is built on authority, structure, breadth and measurement — the same foundations as good search strategy, reweighted for a world where models, not link lists, mediate discovery. The practical components are concrete enough to act on.

Begin with authority and original substance, because that is what the models reward. AI systems disproportionately cite content that is authoritative, original and verifiable — proprietary data, genuine expertise, clear first-hand experience. A useful test is whether your content offers something no one else has: a dataset, a benchmark, a framework drawn from real work. Lookalike content that restates what a dozen other pages already say gives an assistant no reason to cite you over them. Original research and expert commentary attract citations precisely because they are not interchangeable. This is the part of the strategy that cannot be faked, and it is the part that compounds.

Structure content so machines can extract it. Models pull discrete, answer-shaped passages from across the web, which means content organized to answer specific questions directly — clear claims, defined terms, concrete numbers, well-formed sections — outperforms keyword-dense prose for AI retrieval. Question-and-answer formats, statistics stated plainly, clear definitions, and use-case-specific sections that match how people actually ask conversational questions all help an assistant find and lift the relevant piece. The shift is from writing for a ranking algorithm to writing for an extraction system, and the two reward different structures.

Build presence beyond your own website, because the assistants pull from far more than brand-owned pages. The most-cited sources across the major assistants include community platforms, professional networks and authoritative third-party references, not just company sites. That means a credible footprint on the platforms models trust — genuine participation in relevant communities, earned coverage, a presence on professional and reference sites — feeds AI visibility in a way a polished homepage alone does not. Earned media and third-party validation matter more in AI discovery than in classic search, because models weight outside corroboration over self-description.

Treat freshness as a ranking factor. AI retrieval systems weight recent content for time-sensitive queries, so cornerstone material needs regular updating — current data, fresh examples, visible “last updated” signals, and explicit notes on what changed. A guide that goes stale loses ground to a newer one on the same topic. For a market moving as fast as AI assistants, where the numbers and the leaders change quarterly, this is not optional maintenance; it is core to staying citable.

Optimize across platforms, not for one. Because each assistant retrieves and cites differently, a resilient plan accounts for the distinct behavior of the major surfaces — the way one leans on community sources, another on its search graph, another on cleanly citable references — rather than tuning for a single one. This does not mean producing entirely separate content for each; it means building authoritative, well-structured, widely corroborated content that travels well across them, and understanding where each surface sources its answers. The goal is presence across the fragmented set, not dominance of one corner of it.

Finally, measure, because most competitors do not. AI visibility can and should be tracked — how often your brand appears in AI answers, your share of voice against competitors across platforms, the sentiment of those mentions, and the referral traffic AI sources actually send — and tools to do this have matured. Standard web analytics can be configured to surface AI referral traffic, and dedicated platforms track citations and share of voice across the major assistants. Teams that track only traditional rankings and clicks miss their AI performance entirely; adding AI visibility tracking is a small effort that turns a blind spot into a managed channel. A plan you cannot measure is a hope, not a strategy — and in a fragmenting market, the operators who can see where they stand across platforms will out-execute the majority who cannot. The throughline across all six components is that none of them is a trick. They are the disciplines of genuine authority, made legible to machines and spread across the surfaces where audiences now ask their questions.

Business impact across sectors as assistants mediate decisions

The fragmentation of the assistant market lands differently in different industries, because each sector uses these tools for different work and faces different stakes when discovery and decision-making move into the chat window. Treating “the AI market” as one undifferentiated thing obscures that the practical impact is sector-specific — and the sectors that understand their own exposure will adapt faster than those that read only the headline.

In retail and consumer goods, the impact is the most immediate and the most existential. As shown earlier, assistants are becoming a front door to product discovery, with tens of millions of shopping queries a day and rising referral traffic to retailers that make themselves findable. The open-rail versus closed-rail split forces every retailer to choose between opening their catalog to outside assistants and building their own. For brands, the stakes are visibility: a product an assistant does not surface is a product a growing segment of shoppers never sees, regardless of its search ranking or shelf position. Structured product data, presence across multiple assistants, and earned third-party reviews that models trust become competitive necessities, not nice-to-haves.

In financial services, the priorities are accuracy, compliance and data control, which favors the enterprise-grade, safety-positioned products. Banks and insurers handle regulated, sensitive data and cannot tolerate the confident errors that language models still produce, so they gravitate toward platforms with strong compliance postures, clear data handling and reliable performance on complex reasoning — part of why Claude has done well in regulated settings. The fragmentation gives financial firms genuine choice about which provider’s risk profile and data practices fit their obligations, and that choice is now a procurement decision with legal weight, not a matter of which tool is most popular.

In healthcare, the tension is between enormous potential and acute caution. Assistants can help with documentation, research synthesis and patient-facing information, but the cost of a hallucinated fact is unacceptable, and privacy rules are strict. Healthcare organizations need providers that can meet data-protection requirements and whose outputs can be governed and audited. The values and trust dimension is sharpest here, because the data is among the most sensitive a person has — which makes the provider’s posture on privacy and reliability a primary, not secondary, consideration.

In education, assistants have already reshaped how students research, write and study, and the sector is split between embracing them as learning tools and managing their misuse. The distribution-led products reach students passively through the platforms they already use, while the question of which assistant a student adopts has long-term implications, since early habits create lasting preference. For institutions, the practical issues are access, equity and academic integrity — and the fragmentation means students arrive already using different tools, which complicates any single-platform policy.

In media and publishing, the impact is double-edged and frankly threatening. AI assistants increasingly answer questions by synthesizing content without sending a click back to the source, which pressures the traffic-based business models that fund journalism and publishing. When an assistant summarizes an article instead of linking to it, the publisher gets the cost of producing the content and little of the traffic, a squeeze that intensifies as more discovery moves into AI answers. The strategic responses — licensing content to AI companies, optimizing to be cited, or restricting crawler access — each carry trade-offs, and the industry has not settled on one.

In B2B software, the assistants are both a channel and a capability. Companies build AI into their own products via APIs, choosing platforms on capability, cost and reliability rather than consumer fame — which is why the consumer-share leader is not automatically the platform builders choose. At the same time, B2B firms must market themselves into AI discovery, where their buyers increasingly research vendors. The clearest cross-sector pattern is that the consumer popularity ranking is the wrong guide for most business decisions: the right platform for building, for compliance, or for being discovered depends on the sector’s specific work, and the fragmentation means each industry should evaluate the field against its own requirements rather than defaulting to whichever assistant has the most users.

The regulatory and trust questions the data raises

The same data that shows a fragmenting, monetizing, increasingly powerful assistant market also surfaces a set of regulatory and trust questions that are no longer theoretical. As assistants mediate more commerce, more decisions and more personal information, the policy stakes rise — and the report’s findings touch several live fault lines that regulators, companies and users will have to navigate.

The most charged is the question of AI in defense and surveillance, which the Pentagon episode dragged into the open. The episode exposed a genuine policy divide: how far commercial AI should be deployed in classified and military contexts, what guardrails are enforceable, and who decides. The disagreement between a lab that drew hard lines and a government that pressed for broad use is not settled by one amended contract — it is the opening round of a long argument about the role of frontier AI in national security, with real consequences for which companies governments will work with and on what terms. The fact that a values stance moved consumer behavior adds a market dimension to what was a purely political question.

Advertising disclosure is the next pressing issue, and it arrived with OpenAI’s ad tests. When an assistant people trust for honest answers begins inserting paid placements, the line between a genuine recommendation and a sponsored one becomes a consumer-protection question. OpenAI has emphasized keeping ads “separate and clearly distinct,” but the principle that users should know when an answer is influenced by payment is exactly the kind of thing regulators tend to formalize once the practice scales. As retail media flows into AI answers, expect scrutiny of how clearly paid content is labeled and how it is separated from organic responses.

Privacy and data handling sit underneath everything. Assistants collect unusually intimate data — the questions people would not ask another person — and what providers do with that data, how long they keep it, and whether it trains future models are questions with direct regulatory exposure, particularly under strict regimes like Europe’s. The fragmentation gives users some power to choose providers on data practices, but it also means personal data is now spread across multiple companies with different policies, which complicates both individual privacy and regulatory oversight. Data-handling commitments have become a competitive feature precisely because they are also a compliance requirement.

Protections for younger and vulnerable users are an area of rising attention. As assistants become daily tools for students and teenagers, questions about age-appropriate design, safeguards and the effects of intimate AI interaction on developing users are moving up the agenda, and providers face pressure to demonstrate responsible handling of minors and vulnerable people. This is a domain where the gap between capability and governance is wide, and where missteps carry outsized reputational and regulatory risk.

The market-structure questions round out the list. With two of the largest players preparing to go public at trillion-scale valuations, and with distribution power concentrated in a few hands — Google’s reach across Android and Search, the role of app-store gatekeepers, the dependence on a handful of cloud providers — competition authorities are watching how AI distribution intersects with existing platform power. The Apple-Google arrangement, the bundling of assistants into dominant operating systems and search products, and the crawler-access fights over who can index whose data all raise antitrust and market-access questions that regulators in multiple jurisdictions have signaled interest in. The fragmentation at the consumer level coexists with serious concentration at the infrastructure level, and that tension — many assistants, but few who control the pipes and the compute — is likely to define the regulatory conversation as much as any single product’s share.

Risks, limits, and what the numbers cannot settle

A responsible reading of the State of AI 2026 report has to be as clear about what the data cannot tell us as about what it can. The headline figures are useful and well-sourced, but they rest on estimates, single-vendor methodology and a young market, and several of the most-quoted claims carry caveats that the headlines drop. Treating the numbers with appropriate humility is not hedging; it is the only way to use them well.

The first limit is that almost every figure traces to one research firm. The 46.4% share, the user counts, the revenue-per-user numbers and the engagement data are Sensor Tower estimates built on a consumer panel and modelling. They are a serious, internally consistent measurement, but they are not an audited industry census, and independent equivalents for the exact splits do not exist. The corroboration across other trackers is directional — everyone agrees ChatGPT’s relative share is falling — but the precise altitude is one firm’s reading. Anyone making a large decision on a specific percentage should remember it is an estimate with a methodology, not a fact carved in stone.

The second limit is comparability. As earlier sections showed, weekly and monthly figures, web visits and unique users, app and embedded and API usage are all different measurements, and they are routinely mixed in coverage. A number that looks authoritative may be quietly incomparable to the number next to it. The whole category suffers from a measurement fog that makes confident cross-product claims harder than they appear.

The third limit is that the standout financial claims are partly secondary and partly caveated. The most dramatic figures — enterprise win rates, revenue crossovers, profitability milestones — often come from company disclosures and reporting around private filings rather than audited statements, and the companies themselves have flagged that some results reflect temporary conditions. Anthropic’s reported first profitable quarter, for instance, came with an explicit caveat that it partly reflected a favorable, possibly temporary, compute discount. The growth is real, but it is young and not yet seasoned by the discipline of public reporting.

Then there are the persistent limits of the technology itself, which no market-share figure captures. Language models still hallucinate — they produce confident, fluent statements that are simply wrong — and as assistants take on higher-stakes roles in shopping, finance, healthcare and decision support, the cost of those errors rises. A market that is growing in users and revenue is not necessarily growing in reliability at the same pace, and the gap between capability and trustworthiness is one of the most important unmeasured variables in the whole story.

Sustainability is the open economic question. OpenAI is loss-making at scale, with enormous compute costs and positive cash flow reportedly years away. The price war compresses margins. The advertising experiment may or may not work without eroding trust. The whole industry is spending heavily against a future that is assumed but not guaranteed, and the IPOs will test whether public investors share that assumption. A market can grow impressively for years and still face a reckoning if the unit economics never resolve.

Finally, churn and the durability of switching are genuinely uncertain. The Pentagon-driven uninstall spike showed that users will switch over values — but some drifted back once the news passed. Loud, news-driven movement is not the same as permanent migration, and a market where users hold several assistants and move fluidly between them may be more volatile than any snapshot suggests. The honest conclusion is that the report captures a real and important moment — the end of single-app dominance — with reasonable accuracy on the direction, real uncertainty on the precise magnitudes, and genuine open questions about whether the trends it identifies will hold or reverse.

Capability is converging, so the fight moves to distribution and habit

One of the least-discussed forces behind the share shift is that the models themselves are becoming harder to tell apart. As the leading assistants converge on capability, the competition is moving from “which model is smartest” to “which assistant is most embedded in your life” — and that shift structurally favors distribution over raw quality. It is a major reason Gemini’s placement-driven rise worked and why ChatGPT’s quality lead no longer translates into an unassailable share.

For the first three years of the category, model quality was the primary battleground. A noticeably better model could win users on merit, and being first with a capability leap mattered enormously. That era is fading. On most everyday tasks, the top assistants now produce broadly comparable results, and while they still leapfrog each other on specific benchmarks and the frontier keeps advancing, the gaps that matter for typical use have narrowed. The labs themselves have shifted emphasis toward reliability and factual grounding rather than only raw capability, a sign that the easy quality gains have been largely captured and the differences that remain are increasingly marginal for ordinary users.

When products are roughly comparable on the core job, the deciding factors move elsewhere — to distribution, integration, habit and trust. This is exactly the dynamic that has played out in other maturing technology categories: once browsers, smartphones or cloud platforms reached rough parity on the fundamentals, competition shifted to ecosystem, defaults and lock-in rather than feature checklists. The assistant market is following the same path, and the 46.4% figure is partly a record of that transition. Google’s advantage is not that Gemini is provably smarter; it is that Gemini is everywhere a user already is. That advantage is unavailable to a standalone product no matter how good its model, which is why distribution is eating into a quality lead.

Habit is the quieter half of this. A product that someone opens by reflex every day has a moat that a marginally better competitor cannot easily breach, because switching means giving up accumulated context, learned workflows and simple familiarity. ChatGPT’s enormous engagement — hundreds of minutes per user per month — is a habit moat, and it is part of why its absolute usage keeps climbing even as its relative share slips. Gemini’s strategy is to build habit through ubiquity, surfacing in tools people already use until the assistant becomes part of the furniture. Claude’s is to embed in workflows deeply enough that the work itself becomes hard to move. All three are competing for reflexive, repeated use, not for a one-time judgment of who is best.

This reframing has a direct consequence for how to read the rest of the market. If capability is converging, then the share map will increasingly be drawn by who controls distribution and who builds the stickiest habits, not by who ships the next benchmark-topping model. That favors players with existing reach — Google’s ecosystem, Apple’s device base, Microsoft’s enterprise footprint — and it means a standalone product, even a category-defining one like ChatGPT, has to defend its position through brand, ecosystem and habit rather than through model superiority alone. The fight has moved, and the products that understand the new battleground will fare better than those still racing only on raw intelligence.

The shift from assistants that answer to assistants that act

The market data captures usage and revenue, but it sits on top of a deeper product transition that will shape the next phase of competition: assistants are evolving from tools that answer questions into tools that take actions on a user’s behalf. This agentic shift — assistants that book, buy, schedule and complete multi-step tasks — changes what the products are for, and it raises the stakes on every theme this piece has covered.

The early signs are already in the data. Instant checkout and in-chat purchasing turn the assistant from an advisor into a transaction surface — the user does not just ask what to buy, the assistant completes the purchase. Agentic features that can carry out tasks across the web, automated buying tools that act when conditions are met, and browser-using agents that navigate sites on a user’s behalf are all moving from demonstration to early product. The shopping sections of this analysis are really early agentic commerce: the assistant is beginning to do the shopping, not just inform it.

This matters because action is far more valuable, and far riskier, than answering. An assistant that completes tasks captures a much deeper slice of a user’s life and a much larger share of economic activity than one that merely responds — it sits at the point of transaction, not just discovery. That is why the commercial stakes are escalating: whoever owns the agent that buys, books and arranges things owns a position in the economy that a question-answering chatbot never could. It is also why the distribution and trust themes intensify. A user might try several assistants for answers, but they will grant the power to spend their money and act on their behalf only to one they deeply trust — which raises the bar on reliability, security and values all at once.

The agentic shift sharpens the open-rail versus closed-rail divide too. If assistants are going to complete purchases, the question of which retailers and services an agent can transact with becomes decisive. A retailer that has blocked outside agents is invisible not just to AI-mediated discovery but to AI-mediated buying, while one that has opened up can be the default an agent transacts through. The crawler-access fights and the storefront strategies described earlier are skirmishes in a larger war over which platforms agents will be allowed to act within — and that war determines where the money flows when the assistant, not the human, clicks buy.

It also magnifies every risk this analysis has flagged. An assistant that hallucinates an answer is a problem; an assistant that hallucinates an action — buys the wrong thing, books the wrong date, shares the wrong data — is a far larger one. The reliability gap that is tolerable for informational use becomes unacceptable when real money and real consequences are on the line. The regulatory questions around disclosure, data and consumer protection all intensify when assistants act autonomously, because the user is delegating judgment, not just seeking information. The technology’s tendency toward confident error collides directly with the demand for trustworthy action.

For everyone operating around these platforms, the agentic shift is the direction to watch. The market-share map today is drawn by who people ask; tomorrow’s may be drawn by who people authorize to act. Reach matters now; the ability to be the trusted agent that completes tasks will matter more. The fragmentation, the trust dynamics, the distribution advantages and the commerce battles all point toward a future where the decisive question is not which assistant gives the best answer, but which one a user — and a business — will trust to do something on their behalf. That is the phase the current data is the prelude to.

OpenAI’s choices for defending the position it still holds

Slipping below half the market is not the same as losing it, and the more useful question is what OpenAI can actually do from here. The company still holds the largest single position in the market by a wide margin, the deepest engagement, and the strongest brand — but each of its obvious defensive moves carries a cost, and the ones that protect share fastest are often the ones that strain the business hardest. Reading the available options clarifies why the next year is genuinely uncertain rather than a simple continuation of the slide.

The first option is distribution through partnership, and the clearest example is the agreement to power a rebuilt Siri on hundreds of millions of Apple devices. Borrowing someone else’s reach is the fastest way to counter Gemini’s placement advantage, because it puts the company’s technology in front of users who never open a standalone app. The catch is that the arrangement is non-exclusive and rented rather than owned: the same surface can carry a competitor tomorrow, and the device maker, not OpenAI, controls the relationship with the user. It blunts the distribution gap without closing it, and it leaves the company dependent on a partner whose interests only partly align with its own.

The second option is monetization through advertising and commerce, which the company has already begun. Turning a billion users into advertising and transaction revenue is the most direct path to funding the enormous cost of staying competitive, and the shopping and ad experiments are early attempts to build that engine. The risk is the one this analysis has returned to repeatedly: the trust that makes an assistant valuable is exactly what advertising can erode, and a product that feels commercially compromised invites the values-driven switching that already cost the company users once. The faster OpenAI monetizes, the more it tests the loyalty it depends on.

The third option is depth in the enterprise and developer markets, where the company competes hardest with Anthropic. Winning the high-revenue professional segment matters more to the business than defending raw consumer share, because those users pay more, churn less and anchor long-term contracts. But this is the arena where a rival has been growing fastest and, by several reported measures, pulling ahead on the metrics that matter most — revenue per user, enterprise win rates, developer mindshare. Competing here means out-executing a focused competitor on its strongest ground, and the early data suggests that fight is far from won.

The fourth option is to keep pushing the capability frontier and hope quality reasserts itself as the deciding factor. If the next model generation opens a clear, sustained quality gap, the convergence dynamic that favors distribution could partly reverse and OpenAI’s research strength would again translate into share. The difficulty is that capability has been converging precisely because every major lab can now produce broadly comparable results, and betting the strategy on a durable quality lead runs against the trend of the past two years. It is the option most aligned with the company’s identity and least supported by the recent direction of the market.

The fifth and least comfortable option is price. Cutting prices to defend share is always available, but in a market where a well-funded rival can match any cut and a cheaper competitor already undercuts everyone, a price war compresses margins without guaranteeing loyalty. For a company already losing money at scale, sustained discounting is the move that most directly threatens the path to profitability, which is why it reads as a lever of last resort rather than a strategy. The realistic path is some blend of the first three — rented distribution, careful monetization and enterprise depth — executed without breaking the trust the brand still commands. None of these is a clean win, and the combination that defends share most aggressively is also the one that strains the business most, which is the real reason the company’s trajectory from here is open rather than settled.

Scenarios for how the share map could move by 2027

Forecasting a market this young is guesswork dressed as analysis, but laying out a few coherent paths is more honest than a single confident prediction. The trends in the report point in several directions at once, and which one dominates depends on variables that are genuinely undecided — distribution deals, monetization outcomes, the durability of trust, and whether capability stays converged. Rather than one forecast, the useful exercise is to sketch the distinct futures the current data makes plausible and the signals that would tell you which is unfolding.

The first scenario is continued fragmentation toward a stable top three. In this path, ChatGPT’s share keeps drifting down toward the low forties while Gemini and Claude consolidate the rest, and the market settles into a recognizable shape: one large generalist, one distribution-powered ubiquitous assistant, and one professional and developer specialist, with the long tail of smaller products serving niches. This is the straightforward extrapolation of what the report shows, and the signal that it is happening would be the slide continuing at a gentle, undramatic pace while absolute usage of all three keeps climbing. It is arguably the most likely path because it requires nothing to change — only the existing forces to keep operating.

The second scenario is distribution-driven consolidation, where Gemini’s placement advantage compounds. If the device and platform integrations keep pushing the assistant in front of users who never chose it, Gemini’s share could rise faster than a simple trend implies, pulling closer to ChatGPT and turning the race into a genuine two-horse contest at the top. The signal here would be Gemini’s user count climbing on the back of defaults and embedded surfaces rather than active preference, and ChatGPT’s slide steepening rather than flattening. This is the path that the convergence-favors-distribution logic points toward, and it is the one Google’s strategy is explicitly built to produce.

The third scenario is a values-and-trust shock that reshuffles the order suddenly. The Pentagon episode showed that a single high-profile decision can move millions of users in days. Another trust event — a privacy scandal, a damaging advertising backlash, a safety failure, or a sharp differentiation on values — could produce a discontinuous jump that no gradual trend would predict. The signal would be a sudden spike in installs or uninstalls tied to news rather than product, as already happened once. This scenario is inherently unpredictable in timing but increasingly plausible in a market where users have shown they will move on principle and where the products are close enough on quality that values can be the deciding factor.

The fourth scenario is a capability breakout that partly reverses fragmentation. If one lab ships a model generation that opens a clear, sustained quality gap on tasks people care about, the convergence dynamic could weaken and quality could reassert itself as a driver of share, concentrating users around the leader again. The signal would be a measurable, durable performance lead translating into share gains that distribution alone cannot explain. This runs against the recent trend, which is precisely why it would be significant if it appeared — and why its absence so far is itself informative about where the technology stands.

The fifth scenario is the agentic reordering, where the action layer matters more than the answer layer. If assistants that complete tasks rather than only answer questions become the center of the market, the leaders in agentic commerce and trusted action could pull ahead regardless of their standing as chatbots. The signal would be share and revenue increasingly tracking transaction volume and task completion rather than queries and time spent. This is the most structurally different future, and the one the shopping and agentic sections of this analysis suggest is already beginning to form beneath the current numbers.

None of these is mutually exclusive, and the real future is likely a blend. The single most useful habit for anyone tracking the market is to watch which signal is firing — gradual drift, distribution surge, trust shock, capability gap, or agentic shift — rather than fixating on the next headline percentage. The share number is an outcome; these forces are the causes, and reading the causes is what turns a snapshot into foresight.

The open questions that will decide the next phase

The most valuable thing a market report can do is sharpen the questions that matter, and the State of AI 2026 data does exactly that even where it cannot answer them. The 46.4% figure closes the book on single-app dominance, but it opens a set of questions whose answers will determine the shape of the market far more than any one quarter’s share movement. Holding these questions clearly is more useful than pretending the data resolves them, because the next phase will be decided in the gaps the numbers leave open.

The first open question is whether trust can be monetized without being spent. Every assistant is racing to turn usage into revenue through advertising, commerce and higher-priced tiers, and none has proven that this can be done at scale without eroding the confidence that makes the product worth using. The advertising experiment now underway is the live test, and its outcome — whether users accept commercial influence in their assistant or punish it by leaving — will shape the business model of the entire category. It is the question on which the most money rides and the one with the least precedent to guide it.

The second open question is whether distribution beats quality permanently or only while capability stays converged. If the models remain hard to tell apart, the player with the best distribution wins, and that favors the platform giants over any standalone product. But if a genuine quality gap reopens, the logic could flip. Which of these holds is not yet decided, and it determines whether the future belongs to whoever owns the most surfaces or to whoever builds the best model. The answer will be written in whether the next capability leaps prove decisive or merely incremental for ordinary users.

The third open question is how durable values-driven switching turns out to be. The market has shown that users will move over trust and principle, but it has not shown whether that movement sticks or fades once the news cycle passes. Some who switched during the Pentagon episode drifted back. If values-based loyalty proves durable, differentiation on trust becomes a lasting strategy; if it proves transient, it becomes a temporary disruption that the largest player can ride out. The distinction matters enormously for how labs position themselves, and the data so far is genuinely ambiguous.

The fourth open question is whether the economics ever resolve. The industry is spending against an assumed future, and the public markets are about to weigh in on whether that assumption holds. Whether the loss-making leader can reach profitability, whether the price war stabilizes or spirals, and whether the revenue engines being built can carry the enormous cost of staying competitive are all unsettled. The IPOs will deliver a verdict of sorts, but the underlying question — whether this market can pay for itself — remains open and consequential for every participant.

The fifth open question is what the shift to action does to everything else. As assistants move from answering to acting, the demands on reliability, trust, distribution and regulation all intensify at once, and no one yet knows how the products or the rules will handle that weight. The agentic transition could be the development that reorders the market more than any share movement to date, or it could arrive slowly enough that the current leaders adapt. Which it is — fast and disruptive, or gradual and absorbable — is among the most important unknowns the report points toward.

What ties these together is that the headline number is the least interesting part of the story. The 46.4% figure is a marker that the single-assistant era is over; the questions of trust, distribution, values, economics and action are where the next era will actually be decided. A reader who walks away tracking those five questions, rather than waiting for the next percentage, will understand the market far better than the headlines alone allow. The data has done its job not by settling the future but by showing precisely where the future is still open.

The three-year arc from one dominant app to a crowded field

It is worth stepping back to register how fast this happened, because the speed is part of what the 46.4% figure means. A market that one product effectively defined and dominated has, in roughly three years, become a contested field with three serious players and a long tail of specialists — a pace of competitive erosion that is unusual even by technology standards. Seeing the arc compressed makes the present moment legible in a way the single number cannot.

When the first mass-market assistant launched in late 2022, it had the category almost to itself. For a stretch, being the assistant and being the category were nearly the same thing, and the early share figures — by some accounts well above eighty percent — reflected a market with no real alternative rather than a market that had chosen one option over others. That dominance was real but fragile, because it rested on a head start rather than on any structural lock-in, and head starts in software are temporary by nature.

The first challenge came from the company with the most reach. A major search and platform company moved its own assistant into the surfaces it already controlled — its search engine, its mobile operating system, its productivity tools — and converted distribution into users at a speed no standalone product could match. The lesson of that rise was that in a market where the underlying technology was becoming widely reproducible, owning the places people already are beat owning the best standalone app. That single dynamic explains a large share of the fragmentation the report now records.

The second challenge came from focus rather than reach. A competitor concentrating on professional, developer and enterprise use built a smaller but far more valuable position, growing its user base by several hundred percent year over year and, by various reported measures, pulling ahead on the economics that matter most. The lesson there was that a generalist leader is vulnerable to specialists who win the highest-value segments even while trailing badly on raw user count. Depth, not just breadth, turned out to be a viable path to a durable position.

Around these three, a long tail formed: a real-time social assistant, an answer engine, open-weight models with regional strength, embedded assistants inside messaging and productivity suites, and new entrants from large technology companies that had been slow to start. None threatened the leaders for the top spots, but together they absorbed enough of a fast-growing market to make the category genuinely plural. The market did not so much dethrone its leader as outgrow the era in which a single leader was the whole story.

What makes the arc striking is that the leader did almost nothing wrong in the conventional sense. Its usage kept growing, its engagement stayed dominant, its brand remained the strongest — and its relative share fell anyway, because the market expanded and diversified faster than any one product could hold. That is the real shape of what happened: not a collapse, but the natural maturing of a category from a single-product novelty into a competitive market. The 46.4% figure is the moment that maturing became impossible to overlook, and the three-year arc behind it is the clearest evidence that the next phase will be defined by competition, not dominance.

Questions readers are asking about the AI assistant market shift

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

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

ChatGPT’s market share slips below 50% for the first time TechCrunch’s report on the Sensor Tower findings, covering the 46.4% figure and the rise of Gemini and Claude in the AI assistant market.

Sensor Tower — State of AI 2026 report The primary research report behind this analysis, detailing market share, user counts, time spent and spending across leading generative AI assistants.

State of AI 2026: key findings and analysis Sensor Tower’s own summary of the report’s headline conclusions, including the True Audience methodology and engagement trends.

Global time spent on generative AI apps projected to more than double year over year The official press release announcing the State of AI 2026 report, with figures on usage growth and the concentration of time spent among the top assistants.

ChatGPT loses ground to Gemini and Claude as share falls below 50% Fast Company’s coverage of the market shift, framing the decline in relative share against continued growth in absolute users.

Claude and ChatGPT revenue per user, per Sensor Tower The Next Web’s report on Anthropic’s reported revenue-per-user crossover and the monetization gap between assistants.

Claude beats ChatGPT in revenue per user as the AI market shifts strategy Analytics Insight’s analysis of Claude’s enterprise-led economics and higher paid-conversion rates relative to ChatGPT.

ChatGPT user base tops 1.1 billion but its market share just dropped Memeburn’s summary of the apparent paradox of record user numbers alongside a falling share of a fast-growing market.

Digital 2026: one billion people using AI DataReportal’s overview of global AI adoption, providing context for the scale and growth of the overall assistant market.

Sensor Tower — State of Web 2026 Companion research on web usage, including the surge in AI assistant web traffic and the shift in how people reach assistants across platforms.

OpenAI’s agreement with the Department of War OpenAI’s own statement on its government agreement, the document at the center of the trust controversy discussed in this analysis.

OpenAI amends Pentagon deal with surveillance limits CNBC’s report on the amended contract and Sam Altman’s response to criticism of the original terms.

OpenAI alters deal with Pentagon as critics sound alarm over surveillance NBC News coverage of the backlash to the government deal and the concerns raised about domestic surveillance.

Inside OpenAI’s Pentagon deal and the fallout Built In’s explainer on the agreement, the competitive context with Anthropic, and the user reaction that followed.

ChatGPT is now 20% of Walmart’s referral traffic while Amazon wards off AI shopping agents Modern Retail’s reporting on the growing role of assistants in retail referral traffic and the divide between open and closed commerce platforms.

Amazon quietly blocks more of OpenAI’s ChatGPT web crawlers from accessing its site Modern Retail’s account of Amazon restricting assistant crawlers and the effect on referral traffic to its product pages.

Amazon shuts out ChatGPT bots in the agentic shopping race eMarketer’s analysis of Amazon’s crawler blocks and the broader contest over AI-mediated shopping.

Target pilots ads within ChatGPT Retail Brew’s report on retail-media advertising arriving inside ChatGPT, an early example of the assistant’s commercial monetization.

State of AI 2026: AI is becoming the new front door to shopping The Neuron’s explainer on the report, focusing on the shift of product discovery and purchasing toward AI assistants.

What is generative engine optimization (GEO)? Search Engine Land’s primer on GEO, the discipline of making content discoverable and citable inside AI assistants.

Plan for GEO in 2026: evolve your search strategy Search Engine Land’s practical guidance on adapting discovery strategy for a market where attention is fragmenting across assistants and search.

Sensor Tower State of AI 2026 coverage Yahoo Finance’s syndicated report on the State of AI 2026 findings, summarizing the headline market-share and usage figures.