How Many People Actually Use AI Search? The Measured Data

Most AI search statistics you will read are laundered — a vendor blog cites another vendor blog citing a press release, and a number with no methodology becomes industry fact. This page is the opposite exercise: every figure below traces to a named primary study on our closed source list, with its period stamped, and the gaps where no measured data exists are labeled as gaps.

How often do AI answers appear in Google Search?

The best measured answer is 18% of searches, as of March 2025. Pew Research analyzed 68,879 real Google searches from 900 consenting U.S. adults over March 1–31, 2025, and found 18% produced an AI summary as part of the results (Pew Research Center, July 2025).

The same study measured what those summaries do to behavior. All figures below are from the one March 2025 dataset:

The trigger rate is not uniform: question-shaped, informational queries produce summaries far more often than navigational or transactional ones — which is why our Search Console approximation method filters to informational queries before looking for the AI Overview signature.

Do AI answers reduce clicks to websites?

Every study that measured it found the same direction: fewer clicks when an AI answer is present. The magnitudes differ by methodology, and the differences are worth keeping visible rather than averaging away:

Two things reconcile the spread. Panel studies like Pew measure all users on all queries; site-side studies like Ahrefs and Seer measure affected queries specifically, so they read steeper. And Seer's +35% is the number that keeps the whole picture from being purely grim: being cited inside the Overview correlated with a higher click-through rate, not a lower one (Seer Interactive). Sources: Ahrefs, Seer 2026 update. What that citation position is actually worth is a separate question — argued through in are AI citations worth anything.

How much traffic do AI platforms send back to the web?

Cloudflare measures this as a crawl-to-refer ratio: HTML pages crawled per page-visit referred, across sites on its network. The January-versus-July 2025 comparison (Cloudflare, 2025):

Read it plainly: in July 2025, Google crawled about 5 pages per visitor it sent; OpenAI crawled about 1,100; Anthropic — even after an 86.7% improvement since January — still crawled roughly 38,000 pages per referred visit, the widest imbalance among major AI platforms. Assistants answer in place; the crawl mostly does not come back as traffic. What that means for whether you should allow the bots at all is a business-model question, worked through in our block-or-allow framework for publishers.

One more measured mechanic belongs here: Kevin Indig's 2026 analysis found 24% of ChatGPT answers are generated without fetching any live page at all, and that content under 3 months old is roughly 3× more likely to be cited (Growth Memo, 2026). Some AI answers never touch the live web — a structural ceiling on how much traffic they can ever send.

Is ChatGPT replacing Google?

No measured dataset says so as of August 2026. In Pew's March 2025 panel, Google search remained the default behavior, with AI summaries appearing on 18% of searches; the summaries change clicking behavior, but the searching still happens on Google. Semrush's 17-month clickstream study of ChatGPT usage documents fast growth and shifting query behavior (Semrush), and Semrush separately projects that AI search visitors will surpass traditional search visitors by 2028 (Semrush, 2025).

Label that last one correctly: it is a projection — an extrapolated curve, not a measurement. Projections in this niche have a short shelf life, and the honest position is that search is redistributing rather than being replaced. The strategic version of that question — what actually changed and what didn't — is the subject of is SEO dead.

Which AI usage statistics should you not trust?

Any absolute user count without a methodology attached. The "X hundred million weekly users" class of figures traces to company press statements; the "Y% of consumers now start their search in AI" class traces to vendor surveys with undisclosed sampling. Neither is auditable, both get laundered between blogs until they read like measurements, and this niche's most-repeated numbers — like Perplexity's supposed "0.75 content quality score" — have no locatable primary source at all.

Our rule, which costs us some impressive-sounding paragraphs: if a figure is not in a named study we can link, or in our own logs marked [our data], it does not ship. Against interest, that rule cuts both ways — it also means we cannot tell you AI search is about to overtake everything, because no measured source says so.

What do we see on our own sites?

Our fleet data is consistent with the measured studies, at small scale. Across our 3 production builds, AI-assistant referrals remain a thin slice of sessions next to organic search, while AI crawler activity in our server logs runs far ahead of the visits those platforms send back [our data] — the same take-versus-give asymmetry Cloudflare measures across its network. The setup we use to see it is public: the GA4 channel group for AI referrals on the analytics side, and server-log review on the crawler side.

That is the honest state of the data in August 2026: real behavioral change, measured click costs, thin but non-zero referral flows, and a lot of confident numbers in circulation that no one can trace. Build strategy on the traceable ones.