How AI search works · Stage 2

What is query fan-out, and how does it change what gets retrieved?

This is the one stage of the pipeline an engine describes in its own words, and it quietly rewrites what a “page about a topic” has to contain to compete.

Deep dive on stage 212 min readUpdated 30 Aug 20269 sources, 7 primary and 2 vendor-described

Back to the roadmap Next: Retrieval

On this page

The short answer

Query fan-out is the step where an AI engine turns one question into several searches instead of one, and each generated search gets its own candidate set at retrieval. A page that answers four of a topic’s sub-questions is therefore eligible at four entry points and a page that answers one is eligible at one, which makes fan-out an argument about breadth, not length. The gap it is invoked to explain is large: a 4,706-query academic audit of Google AI Overviews found that on average 53% of the domains the system consults are not in the organic top 10 for that query at all.²

Key takeaways

Google is the only engine that documents the technique, and it hedges: AI Overviews and AI Mode “may use a ‘query fan-out’ technique”. Any account of how a different engine expands a question is inference.¹
Length is not the lever. A vendor dataset of 174,048 cited pages put the correlation between word count and citation position at Spearman 0.04, with 53.4% of citations going to pages under 1,000 words.⁶
Padding costs relevance. In a KDD 2026 benchmark, ten body-only strategies averaged a top-20 presence drop from 0.58 to 0.53; the worst hit 0.37.⁵
You cannot see your own fan-out queries: Search Console folds AI-surface impressions into the ordinary “Web” search type, so any list sold as yours is a reconstruction.¹

Definition

What is query fan-out?

Query fan-out is the generation of multiple search queries from a single user question, so that the engine retrieves for a topic rather than for a string. It sits between interpretation and retrieval: the intent goes in, a set of related searches comes out, and each of those searches runs its own retrieval. Google states it in one sentence in its own documentation, and the hedge in that sentence is worth keeping: AI Overviews and AI Mode “may use a ‘query fan-out’ technique” that works by “issuing multiple related searches across subtopics and data sources” to develop a response.¹

Two words in that sentence do a lot of work. Subtopics means the expansion is semantic: the engine decomposes the question into aspects a good answer would cover. Data sources means the expansion is also lateral: different searches may be aimed at different corpora rather than all at the same web index. And may means Google is describing a technique its systems can use, not a guaranteed step in every response. The rest of the pipeline this step sits inside is in the how AI search works stage.

Where those generated searches land is documented on a second Google page, its AI optimization guide, which defines grounding as improving the quality, accuracy and freshness of AI responses “by relying on our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index.”⁹ Fan-out therefore does not route around Search. It multiplies the number of ordinary Search retrievals a page has to be a candidate for.

Mechanism

Why does one question become several searches?

Because a conversational question usually asks for a judgement, and a judgement needs several kinds of evidence that no single search returns. “Is this worth it for a two-person team?” is not answerable from one result set. Answering it well requires what the thing costs, how long it takes to set up, what the alternatives are, and what goes wrong: four different searches with four different vocabularies, none of which contains the words “worth it”. An engine that issued one literal search for that phrase would retrieve forum posts arguing about the phrase.

This is the mechanical reason AI answers cite pages that never mention your category. If your question expands into a sub-query about subscription pricing models in general, a page about subscription pricing models in general is a legitimate candidate, whatever industry it was written for. It is also the mechanical reason the reverse happens: a perfectly on-topic page that answers the visible question and none of its sub-questions competes at exactly one retrieval point out of several, and loses the other entry points to whoever did answer them.

Consequence

How does fan-out change what gets retrieved?

Fan-out multiplies the number of independent chances a page has to enter the candidate pool, and it decouples the citation set from the ranking set for the original query. The academic measurements of that decoupling are large. A 4,706-query audit of Google AI Overviews published in Findings of ACL 2026, on data collected in September 2025, found that 53% of the domains AI Overviews consults do not appear in the organic top 10 for the query at all, and 27% are absent from the top 100.² A 55,393-query preprint from March and April 2026 measured the comparable gap at 29.8% of AI Overview reference domains absent from the corresponding first page.⁸ Read the pair as a range of roughly 30% to 53%, and note that Kirsten counts domains the system consults, which is not the set it cites. An 11,500-query study presented at SIGIR 2026 measured URL-level Jaccard similarity of 0.11–0.18 between Google organic results, AI Overviews and Gemini.³

The same audit found that the retrieved set is not even stable for one query: running the same queries two months apart, only 18% of the web pages AI Overviews returned were common to the two runs, against 45% for organic search.² A screenshot of who got cited is a sample of one, so any before-and-after on your own coverage needs repeated measurements before it says anything.

Fan-out is the most plausible mechanism for that gap, and it is worth saying plainly that plausible is what it is: neither study instruments the fan-out step, because nobody outside Google can. What the studies establish is the size of the effect to be explained. What Google’s documentation establishes is that a mechanism of the right shape exists. Those are two different kinds of evidence and they are stronger together than either is alone.

Ranking has not stopped mattering. It is now necessary for the high-frequency citations and increasingly insufficient for the long tail: the pages that get quoted over and over still tend to be pages that rank, while the tail of one-off citations increasingly comes from pages that rank nowhere near page one for the visible query but match a fan-out sub-query precisely.

Breadth, not length

Does fan-out mean I should write longer pages?

No. Fan-out rewards the number of sub-questions you genuinely answer, and word count is close to unrelated to citation outcomes. A vendor study published December 2025 analysed 560,346 AI Overviews resolving to 1,677,876 cited URLs, of which 174,048 pages had extractable content, and measured the Spearman correlation between word count and citation position at 0.04. In the same dataset 53.4% of citations went to pages under 1,000 words, and pages cited in positions four to ten averaged 1,690 words against 1,270 words at position one, so longer pages sat lower if anything.⁶

Set that beside the fan-out coverage finding from a different vendor dataset published January 2026, drawing on 146 million search result pages and 1.9 million AI Overview citations, which reported that ranking across a topic’s fan-out queries correlated with AI Overview citation at Spearman 0.77, and that pages ranking across fan-out queries were 161% more likely to be cited.⁷ Both figures are vendor-published and correlational, and both point the same way when read together: the thing that moves is coverage of distinct questions, not volume of prose. Ten thin pages that each answer a fragment compete once each. One page that answers the cluster competes at every entry point.

Visibility of the step

Can you see the fan-out queries for your topic?

No engine exposes the queries it generated, and Google’s reporting is aggregated rather than itemised. Its documentation states that “sites appearing in AI features (such as AI Overviews and AI Mode) are included in the overall search traffic in Search Console” and are “reported on in the Performance report, within the ‘Web’ search type.”¹ You can see that AI surfaces sent impressions and clicks. You cannot see which sub-query retrieved you, because the fan-out queries are internal.

The step is also narrower than it sounds. Across 55,393 trending queries collected in March and April 2026, only 13.7% returned an AI Overview at all, rising to 64.7% for question-formatted queries.⁸ Nothing fans out where the surface never fires, so the queries worth reasoning about this way are the ones phrased as questions.

So any tool or consultant offering “your fan-out queries” is offering a reconstruction. That is not necessarily useless, because a good reconstruction of the questions a buyer would ask next is a genuinely helpful artefact, but it is inference presented as telemetry and you should price it accordingly. The cheapest honest substitute costs nothing: read the last twenty questions real prospects asked you in sales calls and support tickets, and the follow-up questions the engines themselves suggest under an answer. Those are the sub-questions, in the buyer’s own words.

Application

How do you write for fan-out without padding?

Give each sub-question its own heading phrased as the question, answer it in the first sentence beneath, and keep that answer self-contained enough to survive being lifted out with no surrounding context. That structure is what makes one page eligible at several retrieval points instead of one, and it is the same structure that helps at stage four, where selection operates on passages rather than pages. Adding a paragraph of throat-clearing before each answer does the opposite of both.

The discipline that keeps this honest is subtraction. Only add a sub-question you can actually answer with something specific; a section that says nothing is a section that dilutes the page’s topical signal without adding an entry point. An end-to-end benchmark accepted at KDD 2026, run over 171,003 documents and 2,700 queries, measured ten body-only rewriting strategies against a top-20 hit rate of 0.58: averaged, it fell to 0.53, the worst strategy to 0.37, and the paper concludes that body-only optimisation is insufficient on its own, because the rewrites drift from the vocabulary real queries use.⁵ Coverage that is real earns entry points. Coverage that is padding costs relevance.

The honest limit of this article

Fan-out is documented by exactly one engine, in one sentence, hedged with the word “may”. Everything in this article about how other engines expand a question is inference from their observed behaviour, not from anything they have published. Nobody outside Google can see a fan-out query, so the coverage effect cannot be measured directly by anyone at all: the 0.77 correlation and the 161% figure above come from a vendor dataset, are correlational, and use a reconstruction of fan-out rather than the real thing. What is solid is that the mechanism exists, that citation sets diverge sharply from ranking sets, and that the divergence needs a mechanism of roughly this shape to explain it.

Where a product fits, and where it does not

The fan-out method costs nothing to run by hand: list the six questions a buyer asks between first hearing of your category and paying for it, search each one, note who is cited, and check whether one page of yours could honestly answer all six. If not, that is your next piece of work, and no software is required to do it. Bavior sits one step later. It runs a fixed prompt set across five engines on a schedule and records which sources each answer cited, so you can watch whether your coverage work changes who gets retrieved, rather than guessing from a single screenshot; where a cited source is a live discussion thread it drafts a reply on an account you control, which you approve before anything posts. It does not reconstruct fan-out queries, because that would be inference dressed as data, and it cannot make an engine issue a search that favours you. The free AI visibility check and the free GEO audit run without a paid plan; paid plans are from $99/mo billed monthly, or $79.17/mo billed annually (as of 30 Aug 2026).

Sources, all checked 30 Aug 2026

Google Search Central, “AI features and your website” (query fan-out, Search Console reporting, eligibility; first-party, last updated 10 Dec 2025): developers.google.com/search/docs/appearance/ai-features
Kirsten et al., Findings of ACL 2026; 4,706-query audit of Google AI Overviews, data collected September 2025 in the US and Germany; “on average 53% (27%) of domains that AIO consults are not contained in top-10 (top-100) Organic search results”; 18% page overlap between two AIO runs two months apart, against 45% for organic: aclanthology.org/2026.findings-acl.526
Grossman et al., SIGIR 2026; 11,500 queries; URL-level Jaccard 0.11–0.18 across Google organic, AI Overviews and Gemini: arxiv.org/abs/2604.27790
“Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023–2026)”, 15 Jul 2026, arXiv:2607.14035 (survey preprint): arxiv.org/abs/2607.14035
Kim et al., “SAGEO Arena: A Realistic Environment for Evaluating Search-Augmented Generative Engine Optimization”, accepted at KDD 2026, arXiv:2602.12187v2 revised 7 Aug 2026; 171,003 documents and 2,700 queries; ten body-only strategies drop Hit@20 from 0.58 to 0.53 on average, worst 0.37: arxiv.org/abs/2602.12187
Content-length study, Dec 2025: 560,346 AI Overviews, 1,677,876 cited URLs, 174,048 pages with extractable content; word count versus citation position Spearman 0.04; 53.4% of citations to pages under 1,000 words. Vendor-published; described, not linked, per this curriculum’s rule.
Fan-out coverage study, Jan 2026: 146 million search result pages, 1.9 million AI Overview citations; ranking across reconstructed fan-out queries correlated with citation at Spearman 0.77, and such pages were 161% more likely to be cited. Vendor-published and correlational; described, not linked, per this curriculum’s rule.
Xu, Iqbal & Montgomery, 2026; 55,393 trending queries collected 13 March to 21 April 2026; 13.7% AIO activation overall and 64.7% for question-formatted queries; “29.8% of AIO reference domains do not appear anywhere on the corresponding first page” (preprint): arxiv.org/abs/2605.14021
Google Search Central, “Google’s Guide to Optimizing for Generative AI Features on Google Search” (grounding defined as improving responses “by relying on our core Search ranking systems”; first-party, last updated 10 Jul 2026): developers.google.com/search/docs/fundamentals/ai-optimization-guide

Continue reading

Learn GEO GEO roadmap Learn GEO Previous: Query interpretation Learn GEO Next: Retrieval

FAQ

Frequently asked questions.

Is query fan-out only a Google thing?

Google is the only engine that documents fan-out in its own words, describing AI Overviews and AI Mode as possibly "issuing multiple related searches across subtopics and data sources" to develop a response. Other assistants visibly issue multiple searches for a single question, and several show the searches they ran in the interface, but none publishes a description of how the expansion works. So the mechanism is confirmed for one engine and observed for others, and any article that describes a specific fan-out algorithm for a non-Google engine is describing a guess.

How many sub-queries does one question fan out into?

Nobody outside the engines knows, and no engine publishes the number. Any specific figure you see, whether five, ten or dozens, is either an inference from an interface that happens to display some of its searches, or an invention. What is reportable is the consequence rather than the count: a 4,706-query academic audit of Google AI Overviews found, on September 2025 data, that 53% of the domains it consults do not appear in the organic top 10 for the query, while a 55,393-query preprint collected in March and April 2026 put the comparable first-page gap at 29.8%. Roughly 30% to 53% is the honest range, and it is the size of the gap that fan-out is invoked to explain. Design your content for coverage of real sub-questions rather than for a number of them.

Should I split my topic into many pages to catch more fan-out queries?

Usually the opposite. Fan-out rewards a page that genuinely answers several of a topic's sub-questions, because that page is a legitimate candidate at each of those retrieval points, while ten thin pages each compete at one. Length itself is not the lever: a December 2025 vendor study of 174,048 cited pages measured the correlation between word count and citation position at Spearman 0.04, with 53.4% of citations going to pages under 1,000 words. Add a section only when you can answer that sub-question with something specific: a section that says nothing dilutes the page without adding an entry point.

Where can I see which fan-out query cited my page?

Nowhere: the fan-out queries are internal and no engine reports them. Google's documentation says sites appearing in AI features "are included in the overall search traffic in Search Console" and reported "within the 'Web' search type", so you see aggregate impressions and clicks from AI surfaces, not the sub-query that retrieved you. Any product presenting "your fan-out queries" is presenting a reconstruction. The free substitute is the last twenty questions real prospects asked you in calls and tickets, plus the follow-up questions engines suggest beneath their own answers.

Bavior Editorial

The team that researches and maintains Bavior’s writing on Reddit marketing and AI search visibility. Every figure here is attributed to a named source with the date it was checked, and none of our links are affiliate links.

Found a number that looks wrong? Tell us and we will re-check it: support@bavior.com