Google AI Overview statistics

What share of AI Overview citations come from top-10 pages?

Both figures come from Ahrefs' top-10 citation analysis, with the year-over-year framing picked up by Search Engine Journal. For how they sit against the rest of the field, see the AI SEO statistics index.

Two snapshots of a population, not a tracked cohort

This distinction decides what the pair of numbers can be used for, and it is almost never stated.

Each measurement samples a population of AI Overview citations at one moment and asks what share of them happened to come from top-10 pages. No individual page was followed from July 2025 to March 2026. Nothing tracked a cohort. The comparison therefore says the composition of the cited set changed. It does not say that pages lost citations, that top-10 pages were displaced, or that any specific page's odds moved in either direction.

Read as composition, the finding is strong and useful: most cited pages now sit outside the classic top 10 at the moment of measurement, so a page outside it has a real path to citation. Read as a cohort study — "top-10 pages lost half their citations" — it says something the design cannot support. That misreading is the single most common thing done with this statistic.

What an AI Overview actually is

An AI Overview is a generated summary appearing above classic organic results for a qualifying query. It synthesises an answer from multiple retrieved sources and attaches citations to some or all of the claims within it. It is a distinct product from AI Mode, a separate conversational search experience, and from a classic featured snippet, which extracts one passage verbatim rather than generating new text.

The last row breaks statistics most often. A user who chose a conversational surface behaves differently from one served an answer they did not ask for, so click-through figures from one cannot be applied to the other — and frequently are.

The source-level difference between the two Google surfaces is measurable in principle, and the AI Mode versus AI Overviews source delta is the study registered to measure it. Until it runs, treat any figure pooling the two as describing neither. Google's documentation on AI features in Search and its AI optimisation guidance remain the only first-party descriptions in existence.

One historical note affects older figures: AI Overviews emerged from Google's earlier Search Generative Experience (SGE) trial. Figures reported under the SGE label describe an earlier, more limited product and are not directly comparable to current measurements.

Why the overlap likely fell, and why we say "likely"

Two mechanisms are proposed in industry discussion; neither has been directly tested. Hypothesis Google may have broadened the source pool AI Overviews draw from, independent of the ranking machinery behind the ten blue links, as the product matured past its early reliance on the same signals. Hypothesis Or AI Overviews may draw increasingly on a retrieval layer resembling the fan-out mechanism documented for AI Mode and described by Google via Search Engine Journal. Such a layer would surface sources a purely ranking-based system would not have selected.

A comparable weakening appears in Ahrefs' cross-engine overlap work, which found only a small minority of AI-cited URLs ranking in Google's top ten at all. Whether the engines agree with each other is a separate question, taken up in the cross-platform citation concordance study and in the most-cited-domains comparison.

Confounds in the overlap trend

The decline is the best-sourced claim here. That does not make it free of alternative explanations, and four deserve stating.

The query sets may not be identical. Two studies a year apart rarely sample exactly the same queries. If the second skews toward query types where overlap was always lower, part of the decline is composition rather than change.

The trigger population moved underneath the measurement. If overviews began appearing on a wider range of queries during the year, the newly included queries could carry different overlap. The average would then fall without behaviour on any original query changing.

"Top 10" is itself unstable, since ranked results are personalised and volatile. And the method may have evolved. One organisation running a comparable method a year later is the strongest available design; it is not a pre-registered, frozen one.

How an AI Overview is assembled

A trigger decision determines whether this query gets a generated answer at all; it is a variable and it moves, so a query that triggered last week may not today. Retrieval gathers candidate documents — this is the stage the overlap statistic lives in, and the fact that overlap is not near-total tells you retrieval is doing something beyond classic ranking. Generation writes the answer, and not everything retrieved gets used. Citation attachment links some claims, and it is the only stage a site owner can observe.

What do AI Overviews cite?

AI Overviews draw on a broader source mix than classic organic results, with reported skew toward multimodal and video content. That is consistent with Google's access to YouTube and the Knowledge Graph as retrieval inputs other engines lack. The specific percentage breakdown circulating in vendor coverage relies on a non-public corpus, so it is graded Partial here rather than restated as a precise figure.

Profound's citation-pattern work and Discovered Labs' platform-by-platform breakdown both describe Google's mix as the most multimodal of the major surfaces, and both rest on corpora nobody outside the vendor can inspect. The direction is more trustworthy than the decimal; the most-cited-domains page carries the per-engine numbers.

None of this transfers to other engines. The retrieval corpus does not carry over to an engine renting a different index. The ranking relationship has no denominator where there are no public ranked results. Trigger rate is meaningless on a product that generates an answer for every query. And citation density differs enough between interfaces that cross-engine citation counts compare interface design rather than visibility. Bing is the clearest illustration — Seer's finding that most SearchGPT citations matched Bing's top results has no Google analogue at all, and the Bing and Copilot guide covers that surface directly.

Click-through impact: why no single number here

Multiple reports describe reduced click-through to organic results when an AI Overview is present, with figures ranging widely by study and definition. The three most quoted — Ahrefs' click-reduction measurement, Semrush's AI Overviews study and Similarweb's referral tracking — do not agree with each other, and that disagreement is the finding. The disambiguated treatment is at AI search CTR statistics; what never arrives at all is zero-click search statistics, and what happens to the traffic that does is AI search conversion benchmarks.

Trigger rate varies the same way. Informational and how-to queries trigger overviews far more often than transactional or highly localised ones, per widely consistent but largely vendor-sourced Partial reporting. Category effects are large enough that local queries, shopping queries and YMYL topics are treated separately here; a blended trigger rate across them tells you almost nothing about any one.

Full statistics table

Common misreadings of 76 to 38

"Ranking no longer matters." It matters less as a predictor than it did. 38% is not zero, and a relationship weakening is not a relationship ending.

"AI Overviews now prefer smaller sites." Nothing in the figure says anything about site size. It says cited pages are frequently not in the classic top 10. Different claims.

"Google changed the algorithm in March." Two measurements a year apart cannot locate a change in time. The intermediate shape is unknown, which is why the sparkline above is labelled illustrative.

"So I should stop doing SEO." Ranking still delivers organic clicks independently of citation. A weaker link between two outcomes is not an argument for abandoning either.

Measuring your own exposure, step by step

No aggregate figure describes your queries. A first-party picture is buildable in an afternoon. Choose twenty to fifty queries that actually drive value, plus five you do not care about as controls. Fix the conditions — same location, same language, no signed-in session — and write them down, because they are part of the measurement.

Record three things per query: did an overview appear, were you cited, and what is your classic ranking. The third column makes the other two interpretable. Repeat at least three times, spread out, and record the variation rather than averaging it away. Segment by query type before concluding anything. When your numbers differ from this page, that gap is usually a fact about your sample, not a discovery about Google.

Those five are where first-party tracking usually goes wrong. The one that costs most is the missing control arm: without queries you are not optimising for, you cannot separate your change from the platform's. The measurement standard sets out the minimum any visibility measurement must disclose to be worth repeating. The technical GEO audit covers the prior question — whether crawlers can fetch and render your pages at all.

The numbers that do not exist

A statistics page is more honest when it names its own gaps. Four are missing here.

Trigger rate by query category, independently sampled. Directional reporting exists; a disclosed-methodology breakdown does not.

Citation rate by content type. Widely asserted, thinly evidenced. The schema study and the author E-E-A-T study test two of the more commonly asserted structural factors, and Ahrefs found no clear schema effect on its own corpus — the sort of result that rarely survives the journey into a tactics listicle.

Time from publication to first citation. One of the more useful numbers in the field, with no public dataset reporting it. The crawl-to-citation latency study is our attempt, and it reports into the AI Citation Index.

The value of an uncited mention. Brand names appear inside generated answers without links, and nobody has published a credible estimate of what that is worth.

Who these numbers apply to, and who they do not

These figures describe queries, not businesses. The denominator is a sample of search queries and the citations attached to them — not a sample of companies, industries or content types.

They apply most directly to informational publishing, where value comes from answering questions. If your traffic is mostly branded or navigational, this page describes a risk you may barely carry. They apply less to transactional and local businesses, where a generated summary is a poorer fit and overviews trigger less often; reading a general trigger figure as your own exposure will overstate the problem. They also apply differently by market and language, since most widely circulated measurement in this field is English-language and heavily United States weighted.

Above all, they do not apply to any single page. A population-level composition figure sets a prior; your own measurement sets the estimate. Use this page to know what questions to ask, and your own instrumented queries to answer them. When you report upward, report exposure rather than projected doom. "This share of our valuable queries currently returns a generated answer" is measurable and defensible. A forecast traffic loss is not.

Verification status

The overlap trend, ~76% to ~38%, is the strongest-sourced claim on this page: same organisation, comparable method, disclosed sample sizes both times, graded Traceable . It would move to a new version on a third overlap measurement, an independently sampled trigger-rate study, a measured relationship between citation and downstream visits, or a structural change to the product. Everything else here carries the grade it earned rather than the grade a cleaner-looking bullet would suggest. Who is behind this site, and how it is funded, is stated on the about page.