ChatGPT Statistics 2026: 15 Numbers That Matter

That number is 68.01%, and it covers the four months to April 2026. Audience-research firm SparkToro measured it. The rankings that produced those searches mostly held their positions. The traffic didn’t.

Fifteen numbers below. Each one heads its own section, each is backed by more than one source, and each carries the date its source published it. Where credible measurements disagree, and on the commercial questions they disagree violently, both are here.

How many people actually use ChatGPT?

One billion people use it weekly worldwide, and 44% of U.S. adults have used it.

The billion mark was confirmed by OpenAI itself in early August 2026. The last full disclosure on the company’s own site, dated February 27, 2026, put ChatGPT past 900 million weekly users with more than 50 million paying subscribers. The disclosed chain runs roughly 100 million weekly users at the end of year one, about 350 million by year two, 700 million in July 2025, 900 million by February 2026, and a billion by August, with search usage specifically up nearly threefold year over year.

The U.S. number is the one that sizes an addressable market, and it comes from Pew Research Center rather than OpenAI: 44% of American adults said they had used ChatGPT when Pew surveyed 5,119 of them in February 2026. In 2025 that figure was 34%. Ten points in twelve months, on a probability sample rather than a panel estimate.

Worth knowing why this is a percentage and not a headcount. OpenAI has never published a U.S. user number, and it has never published a global monthly figure either; it reports weekly actives only. So when headlines had the ChatGPT app passing 1 billion monthly users back in June 2026, that was a Sensor Tower panel estimate dressed as a disclosure, measuring a different thing than the weekly milestone OpenAI confirmed two months later. The U.S. counts circulating on competing statistics pages, 77.2 million here, 67.7 million there, are the same species: third-party panel estimates that no primary source supports. A penetration rate you can trust beats a headcount you can’t.

Similarweb, measuring independently, logged 5.52 billion visits to chatgpt.com in December 2025. Visits get swapped for people so casually that a page can inflate its own headline by an order of magnitude without printing a single false word.

Is AI search mainstream yet, or still early?

Just under half of American adults now use an AI chatbot.

Pew Research Center surveyed 5,119 U.S. adults in February 2026 and found 49% using chatbots, with 24% reaching for one daily. Two years earlier the figure was 33%. Pew flags a question-wording change between those surveys, so read the sixteen-point jump as direction rather than precision. Its cleaner series is brand-level, and there the climb is unambiguous: ChatGPT ever-use among U.S. adults ran 18% in 2023, 23% in 2024, 34% in 2025, and 44% in 2026.

McKinsey & Company, one of the world’s largest management-consulting firms, arrived somewhere similar from a different direction in October 2025. About half of consumers were already using AI-powered search, and 44% of those called it their primary and preferred source of information, against 31% who still said traditional search.

That second figure is the one that changes planning. Adoption is a habit. Preference is a default, and a default is what turns a tool into an economy.

How big are the engines that aren’t ChatGPT?

Gemini reached 950 million monthly users, and Copilot ships into more than 30 million paid corporate seats.

Sundar Pichai gave the Gemini number on Alphabet’s Q2 2026 earnings call in July, with daily actives tripling over the year. At Google I/O two months earlier he put AI Overviews past 2.5 billion monthly users and AI Mode past a billion inside its first year. Microsoft put Copilot past 30 million paid seats at its July 2026 earnings, up from 20 million a quarter earlier, with customers running fifty thousand seats or more up roughly sevenfold in a year.

Claude is the sharpest illustration of why consumer share misleads in B2B. Pew has it in sixth place among U.S. adults at 6% ever-use. Menlo Ventures, surveying enterprise buyers in December 2025, has it in first place on enterprise LLM spend at 40%, against OpenAI’s 27%, a finding worth reading alongside the fact that Menlo is an Anthropic investor.

Panel measurement is not disclosure, and visits are not users. Read the column as orders of magnitude.

Do Google searches still send clicks to websites?

Fewer than one in three does.

SparkToro, working with Similarweb data, measured 68.01% of U.S. Google searches ending without a click between January and April 2026. In 2024 the same measurement returned 60.45%. Rand Fishkin is unusually forthcoming about the limits: the decade-long trend line splices three separate measurement panels, and the mobile portion covers browsers only, excluding Google’s own app, so the published figure is probably the optimistic one.

Two other instruments point the same way. Pew, tracking the real browsing of 900 U.S. adults in March 2025, found people clicking a search result on 8% of visits to pages carrying an AI summary against 15% of visits without one. Ahrefs, aggregating Search Console data across tens of thousands of sites, has Google supplying a steadily thinning share of tracked referral traffic through mid-2026.

Three organisations, three instruments, one direction. That agreement is the only reason a number this consequential is safe to build a budget on, and it’s the arithmetic underneath what a zero-click economy does to a pipeline.

How often do AI Overviews actually appear?

On about one in five searches, and question-shaped searches trigger them six times as often.

Ahrefs analysed 146 million results pages using September 2025 desktop data and found AI Overviews on 20.5% of them. The average conceals the pattern that matters: 57.9% of question queries produced one, against 9.5% of single-word queries, with anything seven words or longer landing at 46.4%.

Pew reached 18% from an entirely different instrument: watching real people browse rather than sampling a keyword database. Two methods, four months apart, landing within two and a half points of each other. Ignore the 48% and 60% figures circulating elsewhere; each traces back to a live tool readout with no stated sample.

The longer and more conversational the question, the likelier an engine answers it rather than routing it. Which is precisely the shape of a buyer’s opening question, and nothing like the shape of a navigational one.

Are AI Overviews reaching the keywords that make money?

Commercial-intent AI Overviews grew 71% in six months.

Semrush studied more than 600,000 U.S. desktop keywords between November 2025 and April 2026. Commercial-intent results pages carrying an AI Overview rose 71% over that window, with Finance up 231% and Computers & Electronics up 108%. Transactional-intent coverage actually fell 5%, so the encroachment is landing on research and comparison rather than on the moment of purchase.

The pricing signal is blunter than the prevalence one, and it is vertical-specific. In Jobs & Education, where the gap runs widest, keywords showing an AI Overview averaged $5.02 in cost per click against $1.51 without; Finance ran $4.84 against $2.14, while several low-CPC categories flipped the pattern entirely. The Overviews are concentrating where clicks cost the most. Set that against Ahrefs’ earlier baseline, which had commercial intent triggering an Overview just 4.3% of the time, and this reads as a category moving fast rather than a category that was always exposed.

Google has stopped summarising trivia. It is summarising the part of the funnel you pay for.

What do AI Overviews do to the top-ranking page?

They cut its clicks by 58%.

Ahrefs tracked position-one click-through rate on AI Overview keywords falling from 0.073 in December 2023 to 0.016 in December 2025, against a forecast counterfactual of 0.037. The damage thins with depth without ever disappearing: down 46.4% at position three, 32.6% at position five, and still 19.4% at position ten.

One methodology note, because this figure gets mangled constantly. Ahrefs’ earlier study reported 34.5% using a one-year comparison window; this one uses two years, because by December 2024 the rollout had already happened. The two are not points on a line, and anyone presenting “34.5% rising to 58%” as a progression is describing a change in method as a change in the world. Pew’s independent 8%-against-15% finding is the sounder corroboration.

Rank and traffic used to be one measurement wearing two names. They aren’t anymore, which is the whole reason a ranking report no longer answers the question a CFO is asking.

So has AI search replaced Google?

Google’s own AI Mode reached just 0.34% of U.S. searches between January and April 2026.

That figure, from SparkToro’s panel work with Similarweb, is the most useful corrective on this page. The most-hyped new search surface in the world was serving roughly one search in three hundred. Whatever is happening to your traffic in early 2026, AI Mode is not yet what caused it.

Scale and share are different questions, though, and Google’s own numbers answer the first one. Sundar Pichai put AI Overviews past 2.5 billion monthly users at I/O in May 2026 and AI Mode past a billion inside its first year, with queries more than doubling each quarter. A surface reaching a billion people and 0.34% of searches is a surface early in its rollout, not a surface that failed.

Here is what nobody can currently measure, and the honest thing to say about it. Comparing Google’s click behavior to an assistant’s referral volume cannot settle whether AI has replaced search, because the two datasets don’t share a denominator. A question a buyer takes to ChatGPT instead of Google never enters Google’s numbers at all: it doesn’t register as a zero-click search, it registers as a search that never happened. Referral-share comparisons measure the click, and the click is the thing that stopped being the scoreboard.

Google’s own position deserves quoting, because it disputes the premise of this whole page. “Total organic click volume from Google Search to websites has been relatively stable year-over-year,” wrote Liz Reid, the company’s head of Search, in August 2025. She added that “average click quality has increased and we’re actually sending slightly more quality clicks to websites than a year ago,” and dismissed the third-party studies as resting on “flawed methodologies, isolated examples, or traffic changes that occurred prior to the roll out of AI features.”

Read that paragraph again and notice what isn’t in it. No baseline. No magnitude. No definition of “relatively stable.” No data release. It is a position rather than a measurement, and it is also the only account of this from anyone holding the actual server logs.

Is AI traffic worth more than the organic traffic it replaced?

Yes, and the studies measuring it disagree by a factor of fifty.

Most statistics pages resolve this by quoting whichever multiple flatters hardest. Four credible measurements exist, and they answer four different questions.

Adobe’s figure carries the detail worth keeping. In March 2025, shoppers arriving from AI converted 38% worse than everyone else. Twelve months later, Adobe measured them converting 42% better. Same channel, opposite verdict: this cohort didn’t improve gradually, it flipped. And it kept going. Adobe’s data through May 2026, reported by Digital Commerce 360 in June, put the premium at 54%.

Then the counterweight, from the only source in that table with no corporate stake in another entry. Ahrefs examined 81,947 sites in June 2025 and found AI visitors viewing 4.0 pages against 5.2 for search visitors, and bouncing at 67.8% against 63.7%. Adobe’s retail-only sample looks considerably healthier than the open web does. And Adobe has owned Semrush since April 2026, so the first two rows are not independent confirmation of one another.

Smaller, better qualified, and measured four ways by four parties working from four different definitions of “better.” That is the honest version, and it’s the version an ROI model has to survive.

Are B2B buyers really using AI to buy?

94% used AI in their last purchase, and half of software buyers now open a chatbot before Google.

Forrester surveyed roughly 18,000 business buyers and reported in January 2026 that 94% had used AI during their buying process, up from 89% the year before. Half of B2B software buyers now reach for a chatbot before Google, according to G2’s April 2026 survey of 1,076 buyers. A year earlier it was under a third.

The buying committee is where this compounds. Forrester found a typical B2B decision drawing in 13 internal stakeholders and 9 external influencers, and buying groups roughly doubling for purchases involving generative AI, from seven people to fourteen.

Fourteen people, each running their own prompts, most of them using whichever engine their employer already pays for. The research that used to arrive at your website arrives in conversations you will never read.

Does getting cited actually change who wins the deal?

Seven buyers in ten changed vendors on a chatbot’s recommendation.

G2’s April 2026 survey is the most commercially consequential finding on this page. Sixty-nine percent of buyers switched away from their intended vendor because of what a chatbot recommended, one in three purchased from a company they had never previously heard of, and 85% said an AI answer naming a vendor raised their opinion of it.

6sense supplies the mechanism. Across roughly 4,000 buyers in November 2025, the vendor a buying group favours before contacting anyone wins 77% of the time, and the eventual winner comes from the day-one shortlist 95% of the time, up from 85%. First contact with a seller now happens 61% of the way through the journey, down from 6sense’s own 69% a year earlier. The widely-quoted “70% of the journey” statistic is a rounded version of a number its own author has since revised.

Put those together and the mechanism stops being mysterious. The shortlist gets drafted before anyone speaks to you, increasingly by an engine, and it almost never changes afterward.

What actually earns a citation?

In 19 of 24 tests, adding statistics to a page made it less likely to be cited.

For two years the field ran on a single 2023 experiment reporting that inserting statistics, quotations and citations into a page lifted its visibility in generative answers by 30 to 40%. In 2025 a team led by Haritz Puerto tested those exact rewrites properly, running 1,921 queries across product recommendation and question answering on GPT-4o-mini, Claude 3.5 Haiku, o3 and o4-mini, and published at NeurIPS. “Out of 54 cases, we uncover only three where the ranking improvements are statistically significant,” they wrote. The add-statistics rewrite significantly lowered citation rank in 19 of 24 settings, and nothing worked on Claude at all.

So the tactic is dead. The principle underneath it survived, and the distinction is worth the paragraph. In a May 2026 preprint, Rahul Vishwakarma’s team ran 252,000 controlled trials across six models, from Gemini 2.5 Flash and GPT-5.2 down to Claude 3.5 Haiku and Kimi K2, and found evidence-backed claims beating unsupported ones in every single model, at odds ratios (the multiple on a claim’s chances of being chosen) running from about two to beyond ten thousand. Only two factors cleared the highest bar in every model tested: matching the topic of the question, and being the first source retrieved rather than the second. Explicit specifics such as price and a recent timestamp pointed the same way in most models, without clearing it in all.

Read the two studies together and the instruction changes shape entirely. Bolting a statistic onto a sentence does nothing, because the model is not counting statistics. Making a claim that is genuinely substantiated does a great deal, because the model is assessing whether the claim holds. One is decoration. The other is the difference between being retrieved and being quoted.

Do rankings still get you into AI answers?

Fewer than four in ten AI Overview citations come from the top 10.

Ahrefs examined 863,000 results pages and 4 million cited URLs in March 2026. It found 37.9% of AI Overview citations appearing in the first ten blocks of the same results page, 31.2% ranking somewhere between 11 and 100, and 31.0% ranking nowhere in the top 100 at all.

The caveat here is large enough to change what the number means. Ahrefs’ 2025 study reported 76%, and the gap is not purely behavioural: the earlier analysis captured only the top three citations per Overview, the newer one captures all of them, position counting moved to a block-based system that pushes organic listings deeper, and Gemini 3 shipped in between. Three changes at once. The blue-links-only variant, 37.10%, is the closer like-for-like, and it still shows a collapse.

Pew’s browsing data fills in the shape of the citation set: 88% of AI summaries cited three or more sources, only 1% cited a single one, and the median summary ran 67 words. Pew captured up to three URLs per summary, so “three or more” is bounded by the instrument and can’t be read as an average.

Ranking still helps. It simply stopped being the whole qualification: retrieval and selection are two separate decisions an engine makes, and only the first one looks anything like SEO.

What doesn’t work?

97% of llms.txt files went unread, and schema markup moved AI Overview citations −4.6%.

Three tactics with confident advocates and no measured payoff: structured data, llms.txt, and formatting for its own sake. Ahrefs ran the cleanest test in May 2026, tracking 1,885 pages that added JSON-LD structured data against 4,000 control pages with similar prior citation levels. AI Overview citations fell 4.6%. AI Mode rose 2.4% and ChatGPT 2.2%, both indistinguishable from zero. What makes this worth reading twice is the correlational version of the same question, run by the same team across six million URLs: AI-cited pages were three times likelier to carry schema. Same question, opposite answer, and the only difference was whether anyone bothered with a control group.

On llms.txt, Ahrefs surveyed 137,210 domains in June 2026 and found 28% publishing a valid file, of which 97% received zero requests in May. The largest single category of requests came from SEO audit tools. Google’s own documentation, updated 10 July 2026, states plainly that Google Search “doesn’t use them” and that publishing one “will neither harm nor help” visibility. Jeremy Howard, who wrote the specification, has been saying as much himself: “llms.txt has nothing to do with SEO. It’s basically a pre-written CLAUDE.md/AGENTS.md file to help people using stuff like Cursor and Claude Code use your site/product.”

Formatting fared no better. In Vishwakarma’s trials, structured versus dense paragraphs returned odds ratios between 0.78 and 1.68, significant in none of the six models. “Formatting changes showed minimal return and can be deprioritized,” the authors concluded.

Structure helps a human read. It isn’t what makes a model quote.

Can you even measure any of this?

Ask an engine the same question twice and roughly a third of the sources change.

Repeated-sampling research published on arXiv in 2026 ran 200 queries across three platforms in nine separate rounds, then measured how much two identical runs actually agreed. Median overlap of cited domains came to 0.29–0.31 on Gemini, 0.33–0.40 on SearchGPT, and 0.50 on Perplexity. On Gemini, two responses to the same query returned an identical source set between 0.01% and 0.10% of the time. The paper, a single-author preprint, so read it as strong rather than settled, gives platform-specific guidance instead of one magic number: roughly 40 to 50 queries for stable estimates on Gemini, about 100 on Perplexity, and 150 or more on SearchGPT, with an explicit warning that the apparent smoothing near 120 queries on SearchGPT is the sample running dry, not the noise resolving.

Stretch the window and it worsens. Profound re-ran about 80,000 prompts per platform a month apart in 2025 and found 59.3% of the domains cited in AI Overviews were new, alongside 54.1% in ChatGPT and 40.5% in Perplexity. The answers themselves are shakier than they read, too: a study commissioned by The New York Times and published by Oumi in April 2026 found AI Overviews correct about 91% of the time under Gemini 3, but only 39% fully grounded in the sources they cited: roughly a third of individual claims unsupported by the citation sitting next to them.

What follows is unglamorous and worth stating plainly. A single manual check of whether an engine mentions your brand tells you almost nothing, which is why tracking platforms like Profound, Peec AI and Ahrefs Brand Radar exist, and why Semrush found 45% of marketing leaders unable to accurately measure their visibility in AI answers, with only 9% holding tools that work across platforms. This problem will not be solved by a dashboard next quarter, and pretending otherwise is how AEO budgets get approved and then quietly cancelled.

Cite First AI keeps a dated register behind this page and re-verifies every figure at its source each quarter, because a statistics page nobody maintains becomes a liability faster than it became an asset.

Find out what the engines currently say about your category

Somewhere this week, a buying group drafted a shortlist for your category without visiting your website to do it. On 6sense’s numbers, that shortlist is 95% likely to contain the eventual winner. If the engine didn’t name you, you were not in the room where it was written.

That is a knowable thing rather than a mysterious one. Cite First AI’s free mini audit runs your highest-value buyer prompts across the major engines and reports back what share of the answers name you, which sources got pulled instead, and where the gap runs widest. It’s the diagnostic stage of the Citation Cascade Method™, and it’s the same starting point that took InCorp to 4.5x traffic in nine months, the biggest revenue jump in their 29 years in business.

No dashboard will tell you this. Ask the engines directly.

Frequently asked questions

How many people use ChatGPT in 2026?

One billion weekly worldwide, a milestone OpenAI confirmed in early August 2026, up from 900 million in February and 700 million in July 2025. For the U.S. specifically, Pew Research Center found 44% of American adults had used ChatGPT in February 2026, against 34% a year earlier. OpenAI publishes no U.S. or monthly figures.

How many Americans use ChatGPT?

Pew Research Center surveyed 5,119 U.S. adults in February 2026 and found 44% had used ChatGPT, up from 34% in 2025, 23% in 2024 and 18% in 2023. Treat U.S. headcounts circulating elsewhere with suspicion: OpenAI has never published one, so they are panel estimates.

What percentage of Google searches end without a click?

68.01% of U.S. Google searches ended without a click between January and April 2026, SparkToro measured using Similarweb data, up from 60.45% in 2024. The mobile portion excludes Google’s own app, so the real figure is likely higher than the published one.

Do AI Overviews reduce organic traffic?

Substantially, though Google disputes it. Ahrefs measured a 58% drop in position-one click-through rate on AI Overview keywords using December 2025 data, and Pew found users clicking results on 8% of visits with an AI summary against 15% without. Google maintains total click volume is “relatively stable” but has published no figures.

Which AI search statistics matter most for B2B?

Four. Ninety-four percent of business buyers used AI in their last purchase (Forrester, January 2026); 51% of B2B software buyers now start with a chatbot rather than Google (G2, April 2026); 69% switched vendors because of a chatbot recommendation (G2); and the pre-contact favourite wins 77% of the time (6sense, November 2025).

Does adding statistics to a page make AI engines cite it?

Not on its own. Puerto et al. tested that rewrite at NeurIPS 2025 and found it significantly lowered citation rank in 19 of 24 settings. What works, across 252,000 controlled trials in a 2026 preprint, is claims that are genuinely substantiated: the model assesses whether a claim holds, not how many numbers surround it.

How reliable is checking your own AI visibility manually?

Not reliable. Two identical queries minutes apart share only 29–50% of their cited domains depending on the engine, and month over month more than half the cited domains change. Stable estimates take roughly 40 to 150 queries depending on the platform, which is why tracking platforms beat spot checks.