AI Overviews

What AI Overviews is, and what it is not

AI Overviews is a feature of Google Search that places a generated summary at or near the top of the results page for some queries. The summary is written by a Gemini model working alongside Google's ranking systems and the Knowledge Graph, and it carries links out to pages that Google says support the statements it makes. Nobody opts in per query. It began as the Search Generative Experience, an opt-in Labs experiment announced in May 2023, was renamed and given to general US users in May 2024, and now runs in more than 200 countries, with France added on 22 July 2026.

Three Google products are routinely treated as one, and almost every piece of bad advice on this subject starts there. AI Overviews sits inside the ordinary results page. AI Mode is a separate conversational surface reached from a tab within Search. The Gemini app is a separate product, with its own address, its own publisher control and no Search Console reporting at all. All three run on generative AI — models that compose an answer rather than retrieve one — but their retrieval paths, controls and reporting differ, and Google declines to promise otherwise: its documentation says AI Mode and AI Overviews may use different models and techniques. A finding about one surface is not a finding about another.

The summary is corroborated, not extracted

Google's own explainer, published in July 2024, describes a customized Gemini model that works in tandem with the existing Search systems, and says the system is designed to find high-quality results from the index to corroborate what the overview says. That phrasing repays attention. The model is not lifting a paragraph out of the best-ranking page and putting quotation marks around it. It writes a statement and then attaches pages that support it.

The practical consequence is that a cited page is evidence for a claim, not necessarily the origin of it. Site owners who go looking for the sentence Google took from them often cannot find one and conclude something has gone wrong. Nothing has. Asking which of your paragraphs was quoted is the wrong question; asking which assertions your page is a credible support for is closer to the mechanism Google describes.

Sitting underneath that is query fan-out, Google's term for issuing several related searches across subtopics and data sources before an answer is written. Robby Stein, a Google vice president of product for Search, put it in plainer terms in an interview published on 13 October 2025: the system appends other queries, perhaps dozens of them, and searches in the background. Google names the technique. It publishes none of the output — not the number of sub-queries, not their wording, not how their results are weighted against each other.

Eligibility is documented. Selection is not.

There is exactly one documented requirement. To be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to appear in Google Search with a snippet. That is the whole bar, and it means the nosnippet directive removes a page from AI Overviews as a side effect — the one reliable lever a publisher has, and the one almost nobody wants to pull, because it strips the snippet from ordinary results too.

Everything past eligibility is undocumented: which eligible pages get chosen, how many links a response carries, and whether organic position affects the odds. Google has neither asserted a relationship between rank and citation nor denied one. That silence is a finding, and it should be reported as one rather than filled with inference.

Google has, however, been explicit about what publishers do not need to do. Its generative-AI optimization guide, published on 15 May 2026, contains a section headed Mythbusting generative AI search: what you don't need to do, which states that structured data is not required for generative AI search and that there is no special schema.org markup to add — while adding, in the sentence practitioners tend to drop, that structured data is still worth keeping for rich results. Danny Sullivan made the same point from a stage at WordCamp US, the WordPress community's annual US conference, on 28 August 2025: "Good SEO is good GEO, or AEO, AIO, LLM SEO, or LMNOPO." GEO — generative engine optimization, the practice of trying to influence whether AI systems cite you — is sold as bundles of markup work that Google has said in writing does not do what the bundle claims. Google's guide is short, and reading it is cheaper than buying one.

What the click data actually shows

Organic CTR — the share of people who see a result and click it — is where this subject stops being academic. Two independent studies have measured what happens to it, and neither came from a search tool vendor.

The Pew Research Center tracked the browsing of 900 US adults across 68,879 searches in March 2025 and published on 22 July 2025. Where an AI summary appeared, users clicked a traditional result on 8% of visits, against 15% where none appeared. They clicked a link inside the summary on 1% of visits, and ended the browsing session on 26% of pages carrying a summary against 16% without one. Agarwal and Sen went further with the first randomized field experiment, posted to SSRN on 3 April 2026: 1,065 US desktop Chrome users, a browser extension randomizing whether the overview was hidden, 68,089 searches. Outbound organic clicks fell 39.8%, zero-click searches rose 34.5%, and removing the overview from the top position produced an 88% relative increase in outbound clicks — with no measured difference in reported satisfaction.

Google's answer, from Liz Reid on 6 August 2025, is that total organic click volume to websites has been relatively stable year over year, with slightly more quality clicks, and that third-party analyses rest on flawed methodologies. Google published no figures. Two studies with published samples and stated methods sit on one side of this argument and a set of adjectives sits on the other, and that asymmetry is itself the most useful thing to know about it.

None of which licenses the single dramatic number that circulates in decks. Seer Interactive found organic CTR on AI Overview queries falling from 1.76% to 0.61% between June 2024 and September 2025 — and reported in the same write-up that queries without AI Overviews fell 41% year over year in that sample. When the control group falls almost as fast as the treatment group, most of the headline decline belongs to something else.

The top ten stopped being the citation pool

In July 2025 Ahrefs examined 1.9 million citations from a million AI Overviews and found 76.10% of cited pages ranking in the top 10 for the query being answered, the median position of the first citation being 2. That number went into a great many strategy documents, because it supported a comfortable conclusion: keep ranking, and citations follow.

The same authors re-ran it on 863,000 keyword SERPs and roughly four million AI Overview URLs, publishing on 2 March 2026. The top-10 share had fallen to 37.9%, with 31.2% ranking between 11 and 100 and 31.0% nowhere in the top 100. Ahrefs attributes part of that to Gemini 3, the default model for AI Overviews since 27 January 2026, leaning harder on fan-out.

Read it carefully before repeating it. This is one vendor measuring one index, it has not been independently replicated, and Ahrefs notes that it improved its citation parsing between the two runs — so some unknown share of the movement is measurement rather than behavior. The direction is well supported; the exact delta is not a clean before-and-after. The instructive part is that Ahrefs' own evergreen guide to ranking in AI Overviews, updated on 12 August 2026, still carries the superseded 76% figure. A stale number can survive inside the publication that superseded it, which is worth remembering whenever a percentage arrives without a date. The follow-up study states its sample and its caveats plainly.

What Search Console reports, and at what cost of precision

Two things are true at once, and confusing them produces nonsense. AI Overview clicks and impressions have always been folded into the main Performance report under the Web search type, undifferentiated from ordinary results. Separately, since 3 June 2026, there has been a Generative AI performance report. It carries a single metric — impressions — for the two surfaces together, sliced four ways: by page, by country, by device and by date. Queries are absent from it, and so are clicks, click-through rate and average position. Its data starts on 18 May 2026, and it reached a subset of property owners first.

So the clicks are not missing — they sit unlabeled in the Web totals. Treating the impressions-only figure as though clicks had been withheld, and reasoning about the gap, is a mistake this report will keep generating for as long as people read its name rather than its help page.

The counting rule matters more than the missing dimensions. Since August 2024, an AI Overview impression is counted only when the link is scrolled or expanded into view, while a classic organic impression is counted when the result set is generated. These are not the same event, so the two figures are not comparable and their ratio is neither a CTR nor a visibility rate. A citation behind a control nobody expands generates no impression at all, so the report can only under-count citation, never over-count it. Carolyn Shelby's description, written on 4 August 2026, is the right frame: a new diagnostic lens, not a new scoreboard.

Two opt-outs, and the one people reach for by mistake

Google-Extended is a robots.txt token scoped to training and grounding for Gemini Apps and Vertex AI. Google's crawler documentation, updated on 14 July 2026, states that it does not affect a site's inclusion in Google Search and is not used as a ranking signal there. It does not remove a page from AI Overviews. It never did.

The control that does is in Search Console settings, called the search generative AI control, live since 17 June 2026. It covers AI Overviews, AI Mode and the generative features in Discover; Google states it is not used as a ranking or inclusion signal elsewhere in Search, and that content drops out within one to two days. Sites using it forgo impressions and traffic from those features, which is the point of it.

The error persists because Google-Extended shipped in late 2023 as the first AI-associated publisher control, and was written up everywhere as the AI opt-out before AI Overviews existed under that name. The label stuck to the wrong product. Sites have blocked it believing they had left AI Overviews, remained in them throughout, and quietly left the Gemini app instead — usually the reverse of what was intended.

The same pattern — activity mistaken for outcome — applies to llms.txt, a proposed file listing a site's content for language models. Google's optimization guide says such files neither harm nor help visibility or rankings in Google Search, because Google Search ignores them. That is a statement about behavior, not an absence of investigation.

Frequently asked questions