Understanding AI search

What AI search actually names

AI search is the class of products that answer a question in generated prose, assembled at the moment of asking from pages retrieved off the live web, with links to some of those pages attached. Three parts carry the weight. The text is written by generative AI — a language model producing sentences — rather than selected from a page; the sources are fetched in response to the question rather than recalled from training; and the attribution, where it exists, is a display decision made after the prose is written, not a ranking published in advance.

The label covers surfaces that differ far more than a shared name suggests: Google AI Overviews, generated summaries on an ordinary results page, present on 43% of US searches by July 2026 against 15% a year earlier (Similarweb via TechCrunch); Google AI Mode, a separate conversational surface Google stated twice in 2026 is not the default; ChatGPT search, retrieval inside a chat assistant; and Perplexity, Microsoft Copilot and the smaller assistants.

Ahrefs compared AI Overviews with AI Mode across 540,000 query pairs on 15 December 2025 and found 13.7% citation overlap, despite the answers being 86% semantically similar: two Google products, the same questions, near-identical prose, almost entirely different sources. Any sentence beginning “AI search does X” without naming a surface and a date describes an average of systems that disagree with each other.

Three dates, and the wrong one everybody uses

Most arguments about what AI search did to the web are between people using different starting points without noticing. 10 May 2023 is Search Generative Experience, an opt-in Search Labs experiment in US English. 14 May 2024 is AI Overviews, shipped to all US users with no opt-in — the correct boundary for any US traffic study. 5 March 2025 is the AI Mode announcement, with US general availability on 20 May 2025.

Google was not first, either: Microsoft began rolling out Bing Chat on 7 February 2023, and Perplexity had been citing sources by default since 7 December 2022. Geography breaks the boundary again. AI Overviews reached 100-plus countries by October 2024 and France only on 22 July 2026, so a European before-and-after study anchored to May 2024 measures a period in which nothing happened in its market.

Then there is the part with no announcement at all. AI Mode moved to a Gemini 3 Flash default on 17 December 2025 and to Gemini 3.5 Flash on 19 May 2026; AI Overviews moved to Gemini 3 globally on 27 January 2026. None was presented as a change to citation behavior and every one plausibly was, so a measurement window straddling a model swap measures two systems and reports the difference as a trend.

Four decisions stand between a question and an answer

A generated answer is four acts, not one: the question is decomposed, documents are retrieved against the pieces, some subset enters the model's context window and is used, and a citation set is chosen and displayed. Google documents the first step and only that one: AI Overviews and AI Mode “may use a 'query fan-out' technique — issuing multiple related searches across subtopics and data sources — to develop a response.” Query fan-out is the practice of splitting one question into several machine-generated sub-queries before anything is retrieved. The candidate pool is therefore assembled per sub-query, so a page that could never rank for the question the user typed can be pulled in for a sub-question it does answer — and no operator exposes the sub-questions.

Steps three and four are separate, and the gap has a name: Seer Interactive's March 2026 work called it a ghost citation, content retrieved and relied on while the brand is never named. Only the entry condition is documented — a page “must be indexed and eligible to be shown in Google Search” — which collapses a great deal of AI-specific advice back into ordinary indexation work. That retrieval works on passages rather than whole documents remains an inference, supported by Semrush's July 2025 finding that AI Mode overlapped Google's top ten at roughly 54% of domains but only 35% of URLs.

Where the documentation stops

Google publishes the most and it is still thin. Its AI features and your website page, last updated 10 December 2025, states that “there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary,” that “there's also no special schema.org structured data that you need to add,” and that “you don't need to create new machine readable files, AI text files, or markup.” That is not guidance. It is a denial, and it is the only operator statement bearing directly on whether this field's core commercial claims are true.

OpenAI publishes a disclaimer — ChatGPT “ranks search results using multiple factors” and “placement is not guaranteed” — without saying what the factors are. Perplexity publishes two sentences, gathering information “from authoritative sources” without defining authoritative or naming its index; the product whose value proposition is attribution documents its selection least. Microsoft is the outlier, rewriting its Bing Webmaster Guidelines on 26 February 2026 to cover Copilot and grounding, with two instructions no other operator has given: state facts directly rather than implying them, and keep entity names clear and consistent.

None of them publishes how sources are selected, how many enter the context window, or any click data, so there is no first-party benchmark against which a third-party study can be checked. That silence is a finding about the documentation, not evidence about the system, and reading it as either confirmation or denial is the commonest analytical error here.

What actually happened to clicks

The evidence on the commercial question is better than the evidence on the mechanism. The strongest independent study is Pew Research Center's, 22 July 2025: 900 US adults who consented to browsing collection, 68,879 Google searches in March 2025, 12,593 producing an AI summary. Users clicked a traditional result on 8% of visits with a summary against 15% without, and clicked a link inside the summary on 1% of visits. Pew has nothing to sell in this market. It is observational, though, and summary-triggering queries skew informational, so some of the gap is query mix.

The randomized evidence closes that gap. Agarwal (Indian School of Business) and Sen (Carnegie Mellon) posted a field experiment to SSRN on 3 April 2026: a browser extension randomly hid the AI Overview for 1,065 US desktop Chrome users across 68,089 searches. Outbound organic clicks fell 39.8% and zero-click searches rose 34.5%. The finding that makes it credible cuts against its likely audience — no difference in satisfaction or perceived information quality. Users were not worse off. Publishers were.

Against that, Google's position: Liz Reid, 6 August 2025, that total organic click volume has been “relatively stable year-over-year.” The post carries no figures, no baseline and no method, and defines a quality click only as one “where users don't quickly click back” — a definition under which cutting low-value clicks registers as improvement. Google holds the only data that could settle it, which makes the statement a dated position rather than a finding, and equally not something the studies have disproved.

A ratio is not a traffic number

A ratio is one quantity divided by another and moves when either half moves. Clickthrough rate is clicks divided by impressions. Almost every alarming number in this subject is a clickthrough rate, and almost every retelling converts it into a claim about traffic on the way past.

Seer Interactive's own data is the cleanest illustration. Its November 2025 study reported organic clickthrough on AI Overview queries falling from 1.76% to 0.61%, a 61% decline, and that is the figure that traveled. Queries without an AI Overview fell 41% in the same period, and that is the figure that did not. Then Seer's fourth-quarter data, reported April 2026 across 5.47 million queries: impressions rose from 15.8 million to 33.1 million while clicks moved from 398,798 to 400,271, so clickthrough halved from 2.52% to 1.21% with no clicks lost. A clickthrough collapse against flat clicks is an arithmetic event, not a traffic event.

The counting rules compound it. Google confirmed in August 2024 that an AI Overview impression registers only when the link is scrolled or expanded into view, while a classic organic impression is counted when the result set is generated. Two headline figures also collide: Ahrefs measured a 34.5% reduction in position-one clickthrough in April 2025, Agarwal and Sen a 34.5% rise in zero-click searches. Different quantities, same digits, cited interchangeably.

The instrument moves while you are reading it

Classic rank tracking rested on an assumption nobody had to state: query the same thing twice, get the same thing back. That assumption is gone, and its absence has been measured. SE Ranking parsed the same 10,000 keywords in AI Mode three times on one day in June 2025; 9.2% of cited URLs appeared in all three runs. Ahrefs, across 540,000 query pairs in December 2025, found 45% of AI Overview citations change between generations of the same answer.

Part of the variation is now deliberate. Since 27 May 2026 a user's Preferred Sources selections are reflected inside AI Overviews and AI Mode, making part of the citation set user-configured: two people are not meant to see the same answer. Any study reporting what “the” AI Overview cites is reporting what one scraper in one location saw once.

Composition shifts faster than reporting cycles, too. Semrush's weekly series caught ChatGPT changing its source mix within a fortnight in mid-September 2025: Reddit fell from about 60% of responses to roughly 10% and Wikipedia from about 55% to under 20%, while AI Mode and Perplexity did not move. Wikipedia remains among the most cited domains anywhere — Pew found Wikipedia, YouTube and Reddit supplying 15% of AI summary sources — but a domain-share table has a shelf life of weeks. The three-run test is the closest thing this field has to an error bar: SE Ranking's AI Mode research.

How to read the next claim you see

None of this requires new tooling, only a reading habit applied to every number in the subject, including the ones above.

• Demand a date. Where the default model behind AI Overviews changed in January 2026, an undated statement about behavior is a rumor, not a claim.
• Demand a surface. A Gemini finding restated as an AI Overviews finding is a measurement of a different system.
• Look for the control group. Ahrefs set 150,000 non-AI-Overview keywords against 150,000 treatment keywords and subtracted the natural decline before attributing anything. Most studies do not.
• Ask whether it has been superseded. Ahrefs put top-ten pages at 76.10% of AI Overview citations in July 2025 and about 38% in March 2026 — same publisher, same metric, seven months apart, with a parser change between that stops the two being strictly comparable.
• Ask who paid. Not disqualifying, disclosing: ask what result would have been commercially inconvenient, and whether the study could have produced it.

The habit is not skepticism for its own sake. This subject has a structural evidence problem: proprietary systems, non-deterministic outputs, no operator changelog for behavior, and a literature written largely by companies selling tools to their own readers. The result is rarely fraud. It is harder to spot — honest measurements of narrow things, restated broadly by parties who benefit from the broad version.

Frequently asked questions