Published
What query fan-out is
Query fan-out is the practice of an AI search surface taking one question from a user, breaking it into several related sub-queries — also called synthetic queries or query variants — running those searches at the same time against one or more indexes, and composing a single answer from the combined results. The user types one thing; the system searches for many things. The term is Google's own, and it is one of very few mechanical details Google has volunteered about how its AI answers are assembled.
The consequence for a publisher is direct and it is the reason the subject matters. If the pool of candidate pages is assembled per sub-query rather than per typed question, then ranking for what the user asked stops being the gate. A page can enter an answer by matching a question the user never asked and will never see.
It is worth separating fan-out from its classical antecedent. Query expansion, a long-standing information-retrieval technique, reformulates one query by adding terms to improve recall. Fan-out issues several distinct queries in parallel and merges what comes back. They share an ancestor and they are not synonyms, and treating them as the same thing obscures the part that changed.
Google's own words, in three escalating versions
Google has described the technique three times, each version slightly fuller than the last, and the wording is consistent enough to be worth reading closely.
On 5 March 2025, announcing AI Mode — the conversational search experience Google launched in Labs and extended to all US users that May — Robby Stein wrote that it "uses a 'query fan-out' technique, issuing multiple related searches concurrently across subtopics and multiple data sources and then brings those results together to provide an easy-to-understand response." On 20 May 2025, Elizabeth Reid restated it at I/O: AI Mode works by "breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf." On 10 December 2025 the term entered Google's developer documentation, where 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."
Three elements are constant: decomposition into subtopics, concurrent execution, recombination. Two things are absent from all three. Google has never published how many sub-queries a question produces, how they are generated, whether they are ranked against one another, or how the merged results are scored and de-duplicated. And the documentation offers no optimization advice at all — the same page says there are no additional requirements to appear in these features and no special optimizations necessary. Note also the hedge in the documentation: may use. Whether AI Overviews fans out on every query or only some is not established by any source found as of 29 August 2026.
The same operation elsewhere, under other names
Fan-out is not a Google invention so much as a Google label. OpenAI documents the same operation without the term: ChatGPT search "rewrites your query into one or more targeted queries that it sends those providers." OpenAI does not say how many, and does not name the technique.
The most concrete public evidence that fan-out is real and countable sits in a developer product rather than a consumer one. The Gemini API's grounding with Google Search returns, in its response, the search queries the model actually executed — the sub-queries themselves, visible. Google's billing documentation for the same feature states that if the model decides to execute multiple search queries to answer a single prompt, that counts as two billable uses. A company does not build a billing rule around behavior that does not happen.
That is useful and it has a limit worth stating. The Gemini API is a developer surface with its own configuration; its behavior is evidence that Google's models decompose queries, not proof of what consumer AI Mode does on any given search. Nobody has published a bridge between the two.
Where the circulating numbers came from
Google has given exactly two figures, and both are routinely detached from their subjects. Announcing Deep Search on 20 May 2025, Elizabeth Reid said it "uses the same query fan-out technique but taken to the next level. It can issue hundreds of searches." Deep Search is a distinct, slower mode. Every article claiming Google fires hundreds of searches for an ordinary query is that quotation with its subject removed. The second figure is Dounia Berrada, a Search senior engineering director, saying on 5 March 2026 that "AI Mode is basically doing a dozen searches for you in the time it takes to do one." That is first-party, which is why it is worth repeating, and it is an interview aside rather than a measurement: no method, no distribution, no query mix.
Everything more precise than those two statements traces to somewhere other than Google. The claim that fan-out generates eight specific types of query variant — related, implicit, comparative, recent, personalized and so on — is a practitioner reading of patent language, most prominently iPullRank's, repeated in agency writing through 2025 and 2026 until it acquired the texture of documentation. So is the claim that fan-out iterates roughly twenty times. Google has published neither taxonomy nor count. Both are unsupported as descriptions of live behavior, and the eight-variant taxonomy has become the most confidently repeated unsourced claim in this subject, largely because it is specific, memorable and fits on a slide.
Patents describe rights, not products
The practitioner model of AI Mode — synthetic queries, a custom corpus pulled per query, passage-level scoring, personalization through a user embedding — comes from a reconstruction built on six Google filings, published by Mike King on 2 June 2025. The two doing most of the work are US12158907, "Thematic Search," filed December 2024, which describes a search engine generating narrower sub-themes and obtaining another set of results for each; and WO2024064249A1, on prompt-based query generation for diverse retrieval.
The reconstruction is a genuinely useful hypothesis and a poor citation, and the distinction is the whole point of this section. A patent describes what a company has secured the right to build. It is typically filed years before anything ships, usually covers several alternative implementations, and carries no commitment that any of them is running. Roger Montti put the caveat plainly in his own coverage on 29 May 2025: Google generally does not confirm that the inventions in a patent or research paper are in use.
A page that asserts Google generates eight query variant types and links a patent has substituted a claim about a legal filing for a claim about a product. That substitution is the characteristic error in this subject, and it is expensive because it is invisible — the citation looks authoritative, and the reader has no way to tell that the source cannot support the sentence.
What it changes about which of your pages gets pulled in
If candidates are gathered per sub-query, the relationship between organic rank and AI citation loosens, and the data is consistent with that. Ahrefs, across 863,000 keyword SERPs and four million AI Overview URLs, reported on 2 March 2026 that 37.9% of AI Overview citations came from URLs in the first ten organic results, roughly 31% from positions 11 to 100, and roughly 31% from URLs ranking nowhere in the top 100. Ahrefs attributes the fall from about 76% in its own July 2025 run to fan-out expansion.
That attribution is the authors' inference and should be read as one. Competing explanations are not excluded: index changes, a shift in the query mix sampled, a change in Google's citation-display policy, or a change in the parser Ahrefs uses to read the results. Ahrefs also sells tools whose value proposition is affected by the finding, which does not make the measurement wrong but does make the framing a commercial choice. And BrightEdge's series over an overlapping period runs the other way, from 32.3% overlap in May 2024 to 54.5% in September 2025, which no fan-out story explains. The direction of travel is contested even where the mechanism is not.
What survives all of that is modest and still useful. Fan-out changes which page on a site is likely to be pulled into an answer, favoring the page that cleanly answers a narrow sub-question over the page that ranks for the broad head term. It does not remove the requirement to be indexed and to rank somewhere for something. The claim that fan-out makes keyword research obsolete is contradicted by Google's own documentation, which says the same foundational practices apply.
You cannot see the sub-queries
This is the operationally decisive fact and it is the one most often glossed over. No platform exposes the sub-queries it ran on a consumer surface, to anyone, as of 29 August 2026.
Google Search Console does not show them, and the confusion about this has a specific source. Search Console documentation states that if a user asks a follow-up question within AI Mode, they are essentially performing a new query, and its data is counted as such — which describes user follow-ups, not machine-generated sub-queries. Suganthan Mohanadasan's analysis of sixteen months of one property's data, published 13 August 2026, found exactly that: 1,127 query strings across 20,300 impressions, consisting of conversational fragments a person typed, with 57.7% of impressions carrying no query string at all.
Every fan-out tool on the market therefore generates a prediction of what the sub-queries might be, usually by prompting a language model, and none observes Google's. That may still be worth doing — it produces the questions a reasonable system would plausibly ask, which is a decent content-planning prompt. It is topic ideation. Reporting it to anyone as the queries Google ran is a misrepresentation of what the tool did.
The one published test of optimizing for fan-out
Exactly one public experiment on fan-out optimization was found. Semrush ran it and published the result on 26 September 2025: four blog articles updated over one week to address ten to twenty predicted fan-out queries each, then tracked for a month.
Citations rose from 2 to 5. That is the number that traveled, usually as "+150% citations." The three numbers that did not travel moved the other way over the same period: share of voice fell from 23.4% to 20.0%, brand visibility from 13.6% to 10.6%, and brand mentions from 18 to 10. Four articles is not a sample, the authors say so themselves, and they note that ChatGPT cut citations across all brands during the test window, which confounds the whole exercise.
The honest reading is that the only published test of this tactic had a sample of four, ran for a month, was confounded by a platform-wide change, and produced results pointing in both directions. It is not evidence that optimizing for predicted sub-queries works, and it is not evidence that it fails. It is evidence that nobody has run the study yet. Meanwhile the common implementation — generating thin blocks of question-and-answer content for every predicted sub-query — has no support in Google's documentation, and Bing's February 2026 guidelines explicitly target large-scale content generated without oversight, quality control or editorial review.
Frequently asked questions
What does query fan-out mean?
It is Google's term for splitting one user question into several related searches, running them at once across subtopics and data sources, and building a single answer from the combined results. Google first used the term publicly on 5 March 2025 when announcing AI Mode, and added it to its Search Central documentation on 10 December 2025, where both AI Overviews and AI Mode are said to use it.
How many sub-queries does Google run for one question?
Google has never published a number. The only first-party statement about ordinary AI Mode is an engineering director's remark in March 2026 that it is "basically doing a dozen searches" in the time one takes. The frequently quoted "hundreds of searches" describes Deep Search, a separate and slower feature, not a normal query.
Are there really eight types of fan-out query?
There is no Google source for that taxonomy. It comes from practitioner readings of Google patent filings and spread through agency writing until it began to be cited as documentation. The same applies to the claim that fan-out iterates about twenty times. Patents describe what a company may build, not what it runs, and Google does not confirm which filings are in use.
Can I see the sub-queries Google used in Search Console?
No. Search Console shows follow-up questions a user typed inside AI Mode, which Google treats as new queries, not the sub-queries the system generated. An August 2026 analysis of sixteen months of one property's data found conversational fragments, with 57.7% of impressions carrying no query string at all. The Gemini API does return executed queries, but that is a developer surface, not AI Mode.
Should I write content targeting predicted fan-out queries?
The evidence does not answer that yet. The only published test, by Semrush in September 2025, covered four articles for one month and produced a citation rise alongside falls in share of voice, brand visibility and brand mentions, during a window when ChatGPT cut citations across all brands. Google's documentation gives no fan-out optimization guidance and says no special optimization is necessary.