AI Overviews Explained

By Bill Hartzer · Updated 17 September 2026

What an AI Overview is

An AI Overview is a summary written by a Gemini model that Google places above the ranked results on a Google search results page. It is a few paragraphs or a short list, with links to some of the pages it drew on, and it appears on some queries and not others without the reader choosing it.

Two things about that description do most of the work. First, an AI Overview is generated, not retrieved: no page on the web contains that text, and refreshing the same query can produce different wording. Second, it is grounded on Google's existing index — the same stored copy of the web that produces the ranked links underneath it, crawled by the same Googlebot. Google has not built a separate AI crawl to feed it. Google describes the retrieval technique as a "query fan-out": the system issues a set of related searches derived from the original query, gathers results across them, and generates over the combined material.

That is why the Overview and the links below it frequently disagree in emphasis but rarely come from different worlds. They are two presentations of the same index.

How often an Overview appears, and why every published figure disagrees

Published estimates of how often Google generates an AI Overview run from 8.1 percent of queries to 93.8 percent. That spread is not measurement failure. It is a consequence of which queries were sampled. Work by researchers at the New Jersey Institute of Technology and Rutgers University, presented at SIGIR '26 and drawing on the ORCAS log of real user queries rather than a vendor keyword set, found that Overviews were generated for 51.5 percent of real-user queries and 65.6 percent of its full benchmark of 14,212 queries — but for 89.9 percent of interrogative queries, 93.8 percent of political queries and only 8.1 percent of trending queries. A separate 2026 audit from Washington University in St. Louis, posted as an arXiv preprint and covering 55,393 queries over 40 days across 19 topical categories, put activation at 13.7 percent of queries overall and 64.7 percent of question-form queries.

The two overall figures, 51.5 percent and 13.7 percent, cannot both describe the same population, and neither is wrong. They describe different query samples measured on different dates against a feature Google adjusts continuously. The question-form figures from the two studies, 89.9 and 64.7 percent, sit much closer together than the two overall numbers do, which is the tell: the query mix is doing most of the work. The Washington University audit is a preprint and has not been through peer review, so check for a published version before treating it as settled.

The consequence for a reader is that a prevalence claim means nothing on its own. Three questions make one readable: which queries were sampled and where that list came from, whether the share is counted over distinct queries or over impressions, and on what dates the sample was collected. A figure built from a keyword tool's head terms, one built from a commercial rank-tracking panel, and one built from a log of what people actually typed will not agree, and there is no reason to expect them to.

How it arrived, with dates

The feature has been renamed once and expanded repeatedly, which is why coverage from different years describes apparently different products.

May 2023 — announced at Google I/O as the Search Generative Experience (SGE), available only as an opt-in Search Labs experiment.
May 2024 — relaunched as AI Overviews and rolled out to all users in the United States, out of Labs and on by default. This is the moment it stopped being an experiment.
August 2024 — extended to the United Kingdom, India, Japan, Indonesia, Mexico and Brazil.
28 October 2024 — extended to more than 100 countries and territories.
March 2025 — AI Mode, a separate conversational surface with follow-up questions and multimodal input, introduced as a Search Labs experiment for subscribers in the United States and widened through the year.
1 December 2025 — Gemini 3 Pro made selectable in AI Mode, with AI Mode available in nearly 120 countries and territories in English.
27 January 2026 — Gemini 3 becomes the default model behind AI Overviews globally, and the "Show more" control on an Overview begins handing the reader directly into AI Mode's chat interface on mobile.

As of August 2026, AI Mode is a surface alongside classic results, not a replacement for them. Nothing published by Google states that it has become the default search experience.

AI Overviews and AI Mode are not the same feature

These two are conflated constantly, including in trade coverage, and the distinction is simple.

An AI Overview is a block on an otherwise conventional results page. The reader typed a query, got a page of ranked links, and a generated summary is sitting on top of it. There is no conversation and no memory of the previous query.

AI Mode is a different destination. It is a chat interface with follow-up turns that carry context, closer in shape to Perplexity or ChatGPT search than to a results page. Since January 2026 the two are joined: expanding an Overview can hand the session into AI Mode.

The reason the distinction matters is that they impose different costs on the reader. An Overview can be scrolled past — the links are still there, further down. AI Mode replaces the list entirely, which means the reader has given up the ability to see what else the index held.

Can they be turned off?

Not by a setting, as of August 2026. There is no official account-level or global switch to disable AI Overviews in Google Search.

What exists is a per-search workaround. Appending &udm=14 to a Google search URL returns a plain web-links filter, and Google added a Web tab alongside Images, News and the rest in May 2024 that does the same thing through the interface. Both are filters applied to one search, not a preference that persists. Browser extensions and custom search-engine entries that pre-apply the parameter exist, and are a workaround on a workaround.

This is a genuine point of contrast rather than a complaint, because other engines have made the opposite choice and documented it:

DuckDuckGo publishes a dedicated AI-free endpoint at noai.duckduckgo.com, described as turning off AI features with no settings required.
Brave Search exposes a setting to disable its Summarizer, the AI answer it launched in March 2023.
Kagi treats AI as something the reader invokes; its Quick Answer does not fire on an ordinary query at all.
Microsoft ships Copilot Search in Bing, launched April 2025, which shows the generated answer and conventional web results on the same page rather than in place of each other, while classic Bing remains available.

Whether an engine forces a generated answer, offers a switch, or requires the reader to ask for one is a structural fact about the product, and it is the sort of thing worth checking before changing a default.

Where the source pages come from, and the crawler question

Because AI Overviews are grounded on Google's ordinary index, the material eligible to appear in one is the material Google has crawled and indexed for Search. There is no separate opt-in.

Google documents a robots.txt token called Google-Extended, which is unusual in having no user-agent of its own — it exists purely as a control. It governs whether a site's content may be used to train and ground Gemini models in Google's other products, and Google states it does not affect inclusion in Search. The conventional preview controls — nosnippet, data-nosnippet, max-snippet and noindex — are what limit the information Google shows from a page.

Google's own documentation on AI features in Search, last updated 10 December 2025, is blunt on the subject of whether anything new is required:

"There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary."
"You don't need to create new machine readable files, AI text files, or markup to appear in these features."

The index is the pool, but the Overview does not simply re-present the links below it. The Washington University in St. Louis audit found that nearly 30 percent of the domains cited in an AI Overview did not appear anywhere on the first page of results for the same query, and that 11.0 percent of the 98,020 individual claims it examined were not supported by the pages cited alongside them. Claim fidelity — whether a generated sentence is actually carried by the page linked next to it — is a narrower measurement than accuracy against the world, and it is the one a reader can check personally by opening the link. The audit is an arXiv preprint.

That is a platform statement about its own system, and a large commercial field has grown up around disputing its practical implications. Both facts are worth holding at once.

The argument about traffic

The dispute over AI Overviews is fundamentally about who receives the visit. A summary that answers the question removes the reason to click the page it was assembled from, and the pages were made by people who were, until recently, paid by those clicks.

Google's position is that Overviews send higher-quality traffic and that links remain prominent. Publisher and trade-body positions vary from cautious to litigious, and several have produced traffic measurements showing declines they attribute to the feature. Almost all of those measurements are correlational: a panel records traffic before and after, with no way to separate the effect of the Overview from everything else that moved at the same time. Google does not publish click data at the query level, so a single confident percentage for "traffic lost to AI Overviews" is usually one methodology rather than a settled fact.

The one estimate with a control group

One 2026 study escapes that problem, and the reason it does is the interesting part. Researchers at the University of Washington Foster School of Business used the fact that Google rolled AI Overviews out to countries in stages, which meant that at any given moment some language editions of Wikipedia were exposed to the feature and others were not. Across 161,382 matched article-language pairs they ran a difference-in-differences design — tracking how traffic to exposed articles moved against traffic to their matched, not-yet-exposed counterparts over the same period — and estimated that exposure to AI Overviews caused an approximately 15 percent reduction in daily traffic to English Wikipedia articles. The unexposed editions supply the counterfactual that panel studies do not have. That is what makes this a causal estimate rather than an association.

Two limits belong with the number. The paper is an arXiv preprint and has not been peer reviewed. And Wikipedia is an unusual publisher: overwhelmingly informational, non-commercial, and exactly the kind of page a short summary can substitute for. The 15 percent figure should not be carried across to news, retail or transactional sites without evidence of its own.

What the correlational evidence looks like

Ofcom, the United Kingdom communications regulator, collated the zero-click evidence in its 2025 discussion paper The Era of Answer Engines. It reports that news-related searches ending with no onward click rose from 56 percent to 69 percent following the release of AI Overviews, that session-ending rates rose from 16 percent to 26 percent when an AI summary appeared, and that only 1 percent of users clicked a link embedded inside the summary itself. Ofcom is collating third-party measurement here rather than running its own clickstream analysis — the underlying sources are Similarweb and the Pew Research Center — and what it reports is an association, not a causal estimate.

The Reuters Institute for the Study of Journalism's Digital News Report 2026 carries a comparable figure from outside the English-language market. Its Poland country page reports Google Search Console data from more than 1,400 websites, collected by the analytics vendor Senuto, showing that the introduction of AI Overviews was associated with a decline in click-through rates of approximately 19 percent year on year. That series runs only to June 2025, the data comes from a commercial vendor rather than from Google, and the Reuters Institute presents it as an association rather than a causal estimate.

Pointing the other way, Google's own internal measurement reached the public record through the United States antitrust case. The court's memorandum opinion on remedies, issued in September 2025 in United States v. Google LLC, records that United States Google Search query volume increased by 1.5 to 2 percent after the introduction of AI Overviews. More searches are not more clicks, and the public version of the opinion redacts the share of queries that trigger an Overview, so the figure settles less than it first appears to. The judgment is under appeal. But it does contradict the simplest version of the story, in which the feature shrinks search itself.

Setting also matters more than the argument usually allows. A 2026 user study from the GESIS Leibniz Institute for the Social Sciences and the University of Wolverhampton, circulated as an arXiv preprint, put 30 participants in front of AI summaries in a domain-specific academic search engine. Half preferred the summary interface to the baseline, the summaries significantly reduced self-reported mental demand, and there were no significant differences in clicks, bookmarks or query reformulations. Thirty participants in a specialist engine cannot be generalized to general web search, and the same study's own audit rated only five of ten summaries fully correct. It is still a reminder that a summary does not remove clicks automatically in every context.

What can be said without arithmetic is structural. Google's index is built from pages it did not write, its Overviews are generated from those pages, and its revenue is advertising sold against the resulting page. That arrangement was already true of the ten blue links; the difference is that a blue link was an invitation to leave and a summary is not. Google Search was separately found to be an unlawfully maintained monopoly in general search by a United States federal court in August 2024, and the remedies decision of September 2025 ordered search-index and click-and-query data sharing with qualified competitors — so the question of who gets to build on this index is live in more than one forum. Both sides of that case appealed and no appellate decision had issued as of 19 August 2026.

For a reader, the practical consequence is smaller and more useful than the industry argument suggests. An AI Overview is a starting point produced by a system that cannot tell you how confident it is. The links beneath it are still the evidence, and the well-documented errors that followed the May 2024 United States rollout are a reasonable argument for treating the summary as a hypothesis rather than a result.

Frequently asked questions

What are Google's AI Overviews?

AI Overviews are summaries written by a Gemini model and shown above the ranked links on Google's results page. They are generated rather than retrieved, so no page on the web contains that exact text, and they are grounded on Google's ordinary index rather than on a separate AI crawl. They appear on some queries automatically, without the reader asking for them.

When did AI Overviews launch?

They were announced in May 2023 as the Search Generative Experience, an opt-in Labs experiment. They relaunched as AI Overviews and rolled out to all United States users in May 2024, reached the UK, India, Japan, Indonesia, Mexico and Brazil in August 2024, and passed 100 countries on 28 October 2024. Gemini 3 became the default model behind them globally on 27 January 2026.

Can AI Overviews be turned off?

Not through a setting. As of August 2026 Google offers no account-level or global switch. Appending &udm=14 to a search URL, or using the Web tab that Google added in May 2024, returns plain links for that one search. By contrast DuckDuckGo publishes an AI-free endpoint, Brave Search has a toggle for its Summarizer, and Kagi requires the reader to invoke AI answers.

What is the difference between AI Overviews and AI Mode?

An AI Overview is a block on a normal results page, with the ranked links still beneath it. AI Mode is a separate conversational surface with follow-up turns that carry context, closer in shape to Perplexity or ChatGPT search. AI Mode launched in March 2025 and, since January 2026, expanding an Overview can hand the reader into it.

Do AI Overviews use a different index from normal Google results?

No. They are grounded on the same index Googlebot builds for Search, using what Google calls a query fan-out — issuing several related searches and generating over the combined results. There is no separate AI crawl. The Google-Extended robots.txt token governs training and grounding in Google's other Gemini products, and Google states it does not affect Search inclusion.

Are AI Overviews accurate?

They are accurate often and wrong sometimes, and the interface gives no signal of which. Errors were widely documented after the May 2024 United States rollout, and a generated summary carries no confidence estimate. The linked sources beneath an Overview are the evidence; the Overview itself is best read as a hypothesis that the reader still has to check.

Do AI Overviews reduce clicks to websites?

Most published figures are correlational and disagree. The strongest causal estimate is a 2026 arXiv preprint from the University of Washington Foster School of Business, which used Google's staggered country rollout as a natural experiment across 161,382 matched article-language pairs and found an approximately 15 percent reduction in daily traffic to English Wikipedia articles. Wikipedia is an unusual publisher, so that figure may not generalize. Ofcom reports that news searches ending with no click rose from 56 percent to 69 percent after the rollout, which is an association rather than a cause.

Does Google say anything special is needed to appear in AI Overviews?

It says the opposite. Google's documentation on AI features in Search, last updated 10 December 2025, states there are no additional requirements to appear in AI Overviews or AI Mode and no special optimizations necessary, and that new machine readable files, AI text files or markup are not needed. A commercial field disputes the practical implications of that statement.

How often do AI Overviews appear?

Published estimates run from 8.1 percent of queries to 93.8 percent, because the answer depends almost entirely on which queries were sampled. SIGIR '26 work from the New Jersey Institute of Technology and Rutgers University, using a log of real user queries, found 51.5 percent overall but 89.9 percent for interrogative queries and 8.1 percent for trending ones. A Washington University in St. Louis arXiv preprint covering 55,393 queries found 13.7 percent overall and 64.7 percent for question-form queries. Any prevalence figure should be read with its query sample and its dates attached.

AI Overviews explained