AI Search Statistics You Can Actually Cite (Primary Sources Only)

Why most AI search statistics cannot be trusted

Because stats pages cite each other, not the research. One roundup invents or garbles a number, five roundups quote the roundup, and by the third generation the number has no source at all. It just has momentum. Search for AI search statistics today and the top results are lists of 60, 70, even 250 data points. Many of those points link to other lists, or to nothing.

This page runs on a stricter rule. Every number below links to a named primary source, with its date, its method, and what it truly measured. Each one was re-checked against that source in July 2026. Numbers we could not trace are flagged in their own section, not laundered back into use. The set is smaller than the big roundups. That is the point.

Citation policy: every statistic on this page links to a named primary source with a date and method, re-verified July 2026. If a number cannot be traced, it is flagged as untraceable, not repeated.

Do AI summaries reduce clicks? The Pew data

The cleanest public evidence comes from Pew Research Center. It analyzed the real browsing behavior of roughly 900 US adults during March 2025, and published the results on July 22, 2025.

Read it plainly. In this panel, an AI summary cut the chance of a click roughly in half, and users rarely clicked the sources cited inside summaries. Two honest limits: the data is from one month, March 2025, and it measures one panel of US adults. It is still the strongest named-source click data in circulation, which is why it is the one worth citing.

What gets cited by AI engines? Two named studies

Two separate 2025 and 2026 studies put numbers on what AI-cited pages have in common. They are the closest thing this young field has to repeated findings, and they point the same direction: structure and original substance, not tricks.

The method notes matter more than the headlines. The Saltbox audit covered 15 domains, close to 2 million monthly organic sessions, and about 7,500 ChatGPT referral sessions. Real, but modest in scale. The SE Ranking finding is the one vendors misquote most. They cite the 65 percent as proof that schema markup earns citations. The same research found no schema type raises citation odds by itself. Both halves belong together, as the full guide to getting cited by chatgpt breaks down.

Is GEO proven? What the research paper measured

The term generative engine optimization comes from one traceable place: a paper by Aggarwal and colleagues, first posted in November 2023 and accepted to KDD 2024. Its headline number: content changes like adding citations, quotes, and statistics boosted visibility "by up to 40%" in generative engine responses.

Cite it with its context attached. The 40 percent came from a bench of test queries in a lab, not from live traffic on the open web. It is proof that how you present things changes what AI engines quote. It is not a promise that any site gains 40 percent of anything. Numbers stripped of that caveat are how the graveyard below gets new residents.

Policy facts with dates: what platforms actually changed

These are not statistics, but they are the dated facts that AI search statistics get built on, and they are the ones writers most often get wrong.

One small original dataset, offered honestly

This site is a rebuilt domain that has run an Entity-First Method for about three months. Here is its own Search Console data, with the sample size stated plainly: one site. Between April 21 and July 20, 2026, daily impressions roughly doubled, from the low thirties to a 60 to 85 range, on zero link building. And among the logged queries are prompts clearly written by AI assistants searching on behalf of their users, complete with instructions to the search engine inside the query text.

We have not seen that last observation documented elsewhere. AI agents are already visible inside ordinary Search Console reports, before most tracking tools report them. One site proves nothing at scale. But unlike an untraceable percentage, you know exactly where this came from, and you are welcome to cite it as what it is.

The graveyard: numbers we could not trace

Each claim below circulates in current stats roundups. As of July 2026, we could not follow any of them to a named primary source with a stated method. That does not prove they are false. It does mean you should not cite them until someone shows the study. If you know the primary source for one, email us and we will move it upstairs with credit.

• The share of AI citations that come from earned media, commonly given as a suspiciously clean 84 percent, with no named study attached.
• The share of total search that AI tools now handle, quoted anywhere from a quarter to more than half, with methods that are never described.
• Precise AI referral market shares quoted to one decimal place, with no measurement panel named.
• Click-loss percentages attributed to unnamed "field experiments" with no author, venue, or date.

Here is the 60-second check that filled this section. Click the link under any statistic. If it lands on another roundup, keep clicking. If you never arrive at a named group with a date and a method, the number is folklore wearing a percent sign. That check is also step one of any honest content audit.

How to cite this page

Everything here is free to cite with a link. The page says exactly when each number was last checked, and we re-check the sources quarterly. A link here will not rot into the graveyard it warns about. Suggested citation: "AI Search Statistics You Can Actually Cite, JNAbear, updated July 2026." Corrections are welcome at info@jnabear.com, and confirmed corrections get credited on this page.

Want the strategy behind the numbers instead? Start with the ai seo checklist, which sorts the work by evidence strength. Or see how topical authority mapping turns findings like these into a content plan. Both follow the same rule as this page: claims trace to sources, or they get labeled as bets.