Query Fan-Out: One Prompt, Twenty Searches

What is query fan-out?

Query fan-out is Google's own name for a retrieval technique it confirmed at I/O in May 2025 and has since written into the AI Overviews and AI Mode documentation: the engine breaks a question into subtopics and issues a multitude of searches simultaneously, then assembles the answer from what those searches return. Deep Search, the heavier variant, runs what Google describes as dozens or even hundreds.

In plain terms, query fan-out means one search query becomes many: the engine expands a single query into multiple related queries and searches for all of them at once. The honest caveat first: nobody outside Google knows the per-query count. The twenty in this article's title is an illustration of the order of magnitude, not a measurement. What is measurable is the effect on which pages get cited, and that evidence is unusually strong.

Why do AI systems use query fan-out?

Because one query rarely contains the whole question. An AI system answering a buying question needs criteria, alternatives, prices and caveats, and no single page of search results contains all of that, so the engine runs the fan-out process instead: expand the user query, retrieve for each sub-query, then synthesise one comprehensive answer from everything that came back. Query expansion is older than AI search, search engines have quietly rewritten queries for years, but LLMs made the expansion cheap, fluent and dozens of branches wide.

The design also explains the economics of being cited by AI engines: every sub-query is a separate retrieval with its own results, so a page that would never rank for the head term can still be the best answer to one branch. That is what makes understanding query fan-out worth a content team's time. It moved the competition from one search to many.

What types of fan-out queries does one prompt produce?

Synthetic queries are generated, not typed, and each fan-out query the engine produces falls into a few recognisable types:

• Related queries: reformulations and synonyms of the original query, padded with modifiers nobody typed, comparatives, qualifiers, the current year.
• Implicit queries: the questions the prompt implies but never states. A prompt about choosing a product fans out into pricing, alternatives, complaints and compatibility even when none of those words appeared.
• Comparative queries: a query like best rank tracker reliably produces versus-style branches against the category's known names, which is why comparison pages keep getting cited.
• Session and context queries: branches conditioned on earlier turns in the conversation, which is one reason the same prompt fans out differently for different people.

The set is also probabilistic: the same prompt fans out differently on different runs, so a branch that existed today may not exist tomorrow. You cannot target a moving, invisible query set exactly. You can cover it, and coverage turns out to be what the citation data rewards.

What does fan-out reward?

The largest relevant dataset is Ahrefs' study of 1.4 million ChatGPT prompts, where the strongest predictor of which pages get cited was semantic similarity between the page (and its title) and the sub-queries, not the user's original query. Descriptive, natural-language URL slugs correlated with citation in the same study, which fits: a slug is one more surface the sub-query can match.

The consequence shows up in ranking data as an apparent paradox: Moz measured 88 percent of Google AI Mode citations coming from pages outside the organic top ten for the visible query. Those pages are not beating the ranking system; they rank for the sub-queries the visible query fanned out into. The competition moved one level down, to the question space around every head term.

How does query fan-out impact SEO and content strategy?

Query fan-out rewrites two SEO habits. Keyword research first: traditional keyword research ranks terms by search volume, but sub-queries have no search volume, because nobody types them, so volume-first planning systematically misses the search queries the engines actually run. The working unit of a content strategy becomes the topic cluster: a pillar page on the head term, cluster pages on the questions around it, which is fan-out coverage expressed as site architecture. Second, technical SEO still gates everything, because a sub-query can only retrieve pages the engine can crawl and read; schema markup, for what it is worth, moved no AI citations in controlled testing, so coverage cannot be bought with markup.

The habit that transfers unchanged is intent mapping. Every fan-out branch is an intent, and covering intents rather than keywords is what the best SEO strategies already did. The difference is that AI-powered search platforms run the mapping automatically, at query time, against your actual pages, and your AI search visibility is decided by how many branches find an answer on them, because that is what decides whether you appear in AI answers at all.

How do you cover a fan-out space?

Treat every query to cover as a question space rather than a keyword:

• Give each sub-intent its own heading, phrased the way someone would ask it. Question-format headings measured a 12 percent organic session lift in a SearchPilot split test, and they are what sub-queries match against.
• Answer the obvious adjacent questions on the same page: every sub-query satisfied elsewhere is a citation you did not get.
• Cover comparisons, specifications and criteria explicitly. Fan-out reliably generates commercial and comparative sub-queries around any product term.
• Do not shred the topic into micro-pages. Google states multi-topic pages are understood fine, and fragmentation has no evidence behind it; one well-sectioned page covers a fan-out space better than ten thin ones.

Where fan-out meets extraction

Fan-out decides which pages are retrieved; chunking decides what gets lifted from them, and the two reward the same structure. Measurements of Gemini's grounding behaviour show a roughly 2,000-word budget per query split across four to six sources, with extraction from any single page plateauing around 540 words. Mike King's chunk experiments found that prepending the heading to a passage improved its similarity to the query by around 17 percent, which is a direct mechanical reason section headings matter.

So each heading section has to stand alone: one sub-question, its answer first, the entity named rather than pronouned. A section built that way is simultaneously a fan-out match and an extractable chunk.

Which tools help with query fan-out research?

No query fan-out tool shows you Google's real sub-queries, so every one of them approximates. Dan Petrovic has published a fan-out simulator, Qforia, that generates plausible sub-query sets for a prompt, and the big SEO suites describe manual LLM prompting for the same job. Treat all of it as brainstorming for query fan-out analysis rather than ground truth, because the real set is probabilistic and private. People Also Ask boxes and your own Search Console queries remain the free sources of real question demand, and the raw material is always the same: what buyers ask, what surfaces alongside, which comparative angles recur.

RankX AI approaches it from both ends. Keyword Research finds the demand and the question variants AI engines expand prompts into, so you can use query fan-out logic inside ordinary keyword planning, and the tracked-prompt runs show which real questions your brand already appears for and where rivals appear instead. For the Google side specifically, the free AI Overview Checker shows whether a keyword triggers an AI Overview and who is being cited in it, which is the fan-out output you can actually observe. The wider practice this sits inside is covered in the GEO guide, and measuring the result is its own discipline.

Sources

• Google, AI features and your website (query fan-out documentation), checked 20 Aug 2026
• Ahrefs, why ChatGPT cites pages, 1.4M prompts, checked 20 Aug 2026
• Moz, AI Mode citation study, ~40,000 queries, February 2026, checked 20 Aug 2026
• SearchPilot, question-format heading split test, checked 20 Aug 2026
• DEJAN, How big are Google's grounding chunks, 7,060 queries, checked 20 Aug 2026
• iPullRank, A refutation of misinformation about chunking, checked 20 Aug 2026
• Ahrefs, schema added to 1,885 pages vs controls, checked 20 Aug 2026