AI Mode

AI Mode, defined

AI Mode is a conversational search surface inside Google Search: rather than returning a ranked list of results, it returns a written answer assembled from several searches Google runs on the user's behalf, and the thread stays open for follow-up questions. Users reach it from a tab inside Search; it does not replace the ordinary results page. It is a different product from AI Overviews, which is a summary block on that ordinary page, and from the Gemini app, which sits outside Search altogether.

It arrived as a Search Labs experiment on 5 March 2025, US only and subscriber only, reached general US availability on 20 May 2025, then India in July and the United Kingdom later the same month. By August 2026 Google's help pages give availability as more than 195 countries and territories and more than 80 languages, with France — the last major withheld market — added on 22 July 2026. The engine is generative AI, meaning a model that composes an answer rather than retrieving one, and Gemini 3.5 Flash has been the stated global default since 19 May 2026.

At the same event Google said AI Mode had passed one billion monthly users. That is a first-party figure with no published definition of a monthly user and no external audit, and it should be read as product scale and nothing else.

Fan-out moves the decision to a query you never see

Query fan-out — decomposing one question into several sub-queries, running them in parallel against Search and other data sources, and synthesizing the results — is the mechanism, and Google names it in both its consumer help page and its documentation for site owners. Robby Stein, a Google vice president of product for Search, described it in July 2025 as the system thinking up a set of related questions and going off to Google them, and in October 2025 as appending perhaps dozens of queries and searching in the background.

Everything that makes AI Mode difficult follows from that. The page that gets cited need not rank for the question the user typed; it may rank for a sub-question the user never saw and the publisher cannot see either. Google exposes the fan-out set nowhere — not in Search Console, not in Analytics, not in any paid product. Tools that display your fan-out queries are generating plausible sub-queries with a language model. That can be a useful exercise in finding gaps in coverage. It is not observation of Google, and reporting it to a client as Google data is a misrepresentation.

The surface is also not purely a web-retrieval product. Stein described integrating Google's real-time systems into it — finance calls, flight data, film information, and a shopping catalog he put at fifty billion products refreshed roughly two billion times an hour. Where an answer is assembled from Google's own feeds, no external page is cited for that part of it, and no amount of publishing will change that. Google's help page adds a detail that reverses the usual assumption in the other direction: AI Mode will show a set of web links where there is not high enough confidence in the quality of an AI response. Links can be a fallback for a weak answer, not only a citation of a confident one.

What two vendors measured, and what the numbers do to common practice

SE Ranking parsed 10,000 US keywords on 20 June 2025 and published on 29 August 2025. Exact-URL overlap between AI Mode citations and the top 10 organic results was 14%, roughly two and a half URLs per query; at domain level it was 21.9%. Responses carried 12.6 links on average across 9,734 triggered answers and 122,617 citations — a figure that predates Gemini 3 and should be treated as history rather than as a current expectation.

The same study did something more useful than measuring overlap. It parsed the identical keyword set three times on the same day, and found that only 9.2% of cited URLs appeared in all three answer sets, with 18.5% appearing in any two. That single number should govern how anybody reports on this surface. A visibility audit built on one pull is one draw from a noisy distribution, and week-on-week citation tracking at the sample sizes most teams can afford is measuring variance and calling it a trend. Repeated sampling is not a refinement here; it is the difference between a measurement and an anecdote. The full study publishes its method.

Ahrefs compared the two Google surfaces directly across 540,000 query pairs of September 2025 data, publishing on 15 December 2025. Citation overlap between AI Mode and AI Overviews was 13.7%, and 16.3% among the top three citations, even though the answers themselves were 86% semantically similar. AI Mode cited something 97% of the time against 89% for AI Overviews, and carried about 3.3 entities per response against about 1.3. SE Ranking arrived at 10.7% independently. Two vendors within about three percentage points, plus Google's own refusal to promise the surfaces behave alike, is as close to settled as this subject gets — and it means AI Overviews research cannot be reused as AI Mode research, which is exactly how most of it is being used.

Google says there is nothing special to do. Read the sentence precisely.

Google's guidance for site owners states that there are no additional requirements to appear in AI Overviews or AI Mode and no special optimizations necessary. Gary Illyes told a Search Central event in July 2025 that normal SEO practices are sufficient and that GEO, LLMO and the rest are not needed — GEO being generative engine optimization, the practice of trying to influence whether AI systems cite you. Danny Sullivan had made the same argument 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."

Read narrowly, all of that is true and useful. There is no submission channel, no markup, no file and no separate index to register with. Google's generative-AI optimization guide of 15 May 2026 makes the point under a heading it titles Mythbusting generative AI search: what you don't need to do, where it states that structured data is not required for generative AI search, and that llms.txt files will neither harm nor help a site's visibility or rankings in Google Search, as Google Search ignores them. Note the precision of that final clause. It is a statement about what the system does, not an admission that nobody has checked.

Read as a claim about outcomes, the same sentence does not survive contact with Google's own fan-out documentation or with a 14% overlap measurement. A page can be indexed, snippet-eligible and ranking well for its target query and still sit outside the citation pool, because the citation pool was assembled from questions its owner never saw. The honest formulation is that classic search work is necessary and insufficient, and that the insufficient part is not addressable with any lever Google has published. That is an uncomfortable thing to put in a proposal, which is precisely why so few do.

Impressions, but no clicks, no queries and no separation

Organic CTR cannot be calculated for AI Mode at all, and it is worth being exact about why, because the reason is not the one usually given. Since 3 June 2026 the Generative AI performance report has shown impressions only, for AI Mode and AI Overviews merged, broken down by Pages, Countries, Devices and Dates, with no query dimension, no clicks, no CTR and no average position. Its data begins on 18 May 2026 and it reached a subset of property owners first.

Meanwhile, and separately, AI Mode clicks and impressions have been folded into the main Performance report under the Web search type since before that report existed. Brodie Clark's experiment, published in June 2025 and updated on 8 June 2026, recorded them there and logged an average position of 1.1 on a small sample. Both facts are true simultaneously. The clicks are not withheld; they are unlabeled, mixed into ordinary Search totals, and there is no filter that separates them.

Two further constraints are easy to miss. Impressions in the generative surfaces are counted only when the link is scrolled or expanded into view, so a citation nobody reaches down to is not counted at all and the report can only under-report. And since 27 May 2026 a searcher's own Preferred Sources selections are reflected in AI Mode responses, which means part of the citation set is chosen by the user rather than by Google's assessment of your page — and that two people running the same query are not guaranteed the same sources.

Claims about AI Mode that do not survive checking

• That AI Mode is now the default Google Search experience. Google denied it on the record twice in June 2026, including denying plans to make it the default in Chrome. What is real is a test, reported on 28 August 2026, of moving searchers from an AI Overview into an AI Mode conversation without a "Show more" click. The belief spread through conference framing and headline compression, not through any product announcement.
• That AI Mode and AI Overviews merged in May 2026. Google's own post about its 2026 Search announcements does not say so, and a contemporaneous analysis published on 21 May 2026 noted specifically that Google declared no merger and did not make AI Mode the default.
• That blocking Google-Extended keeps you out of AI Mode. It does not. That token is scoped to training and grounding for Gemini Apps and Vertex AI, and Google states it does not affect inclusion in Google Search. The control that governs AI Mode is the search generative AI control in Search Console, live since 17 June 2026.
• That FAQ markup earns citations here. The +112% figure attached to this claim comes from a study of 1,000 Gemini queries run between November 2024 and March 2025 — a different product, a superseded model generation, and correlational by its own authors' statement. It is quoted about AI Mode by people who have not opened it.
• That Google's billion monthly users and Similarweb's visit counts describe the same thing. Similarweb measured AI Mode visits rising from 126 million in June 2025 to 279 million in May 2026. Different unit, different method, no relationship. Putting the two in one sentence produces a comparison that means nothing.

What would change the picture on the citation question is a replication of the overlap work by a party with no tool to sell, and a study with repeated sampling built into the design. Neither exists as of 29 August 2026, and the surface comparison that comes closest says plainly which parts of it are correlation.

Frequently asked questions