Query Fan-Out

Quick definition

Query fan-out is a retrieval technique used by AI search systems. Instead of running one search for the words a person typed, the system breaks the question into several related sub-queries, runs them across subtopics and data sources, and combines what it finds into a single answer.

Google uses the term for AI Mode and AI Overviews. OpenAI describes a similar step for ChatGPT Search, where a prompt is rewritten into one or more targeted queries. The mechanics differ between systems and are only partly documented, but the consequence for SEO is the same: the searches that decide which pages are considered are often not the searches a keyword tool shows you.

Where the term comes from

The underlying ideas are older than the name. Information retrieval has long used query expansion, query rewriting, and query decomposition to find documents that a literal query would miss. Research on retrieval-augmented language models has studied the same pattern. One example is the 2023 paper "Query Rewriting for Retrieval-Augmented Large Language Models", which trains a small model to rewrite queries before retrieval.

"Query fan-out" became an SEO term in May 2025. When Google expanded AI Mode at its developer conference, it described the system as breaking a question into subtopics and issuing many queries at the same time on the user's behalf. Google also introduced Deep Search, which uses the same technique at a larger scale and, according to Google, can issue hundreds of searches for one report. Google's Search Central documentation now states that both AI Overviews and AI Mode may use query fan-out, issuing multiple related searches across subtopics and data sources to develop a response.

How query fan-out works

No provider publishes its full pipeline, so the following is a simplified model that matches the public descriptions:

1. Interpretation. The system reads the question, the conversation so far, and available context such as language or approximate location.
2. Decomposition. It derives sub-queries for the facets of the question: definitions, options, comparisons, constraints, prices, recent changes, or local details.
3. Parallel retrieval. The sub-queries run against one or more indexes and data sources. Results come back as documents or passages.
4. Selection. The system picks the passages that best support each part of the answer.
5. Synthesis and citation. A language model writes the answer and attaches links to supporting pages.

Step 2 is where SEO assumptions break. A page does not need to match the original question. It needs to be a good result for at least one of the sub-queries, and its relevant passage must be clear enough to be selected in step 4.

An example

Take a hypothetical question: "Which e-bike works for a 20 km commute in a hilly city for under 2,000 euros?"

A fan-out system could plausibly derive sub-queries such as:

• e-bike range on hilly routes;
• motor torque needed for steep climbs;
• e-bike models under 2,000 euros;
• comparisons between hub motors and mid-drive motors;
• maintenance costs for daily commuting;
• reviews of specific models mentioned in earlier results.

This list is an illustration, not a record of what any system actually issued. It shows the practical point: a retailer's category page, a manufacturer's spec sheet, a review site's comparison, and a cycling forum thread could all contribute to one answer, each through a different sub-query.

Query fan-out in Google AI Mode and AI Overviews

Google's documentation for site owners contains several statements worth knowing:

• AI Overviews and AI Mode may use query fan-out across subtopics and data sources.
• Because of this, the response can show a wider and more diverse set of supporting links than classic results.
• To appear as a supporting link, a page must be indexed and eligible to be shown in Google Search with a snippet. Google states that there are no additional technical requirements.
• Traffic from AI features is included in the Web search type of the Search Console Performance report.

In other words, there is no separate fan-out index and no special markup. Fan-out changes which queries a page competes for, not the basic eligibility rules. Normal controls such as nosnippet, max-snippet, and noindex also apply to AI features.

Query fan-out in ChatGPT and other assistants

OpenAI's help documentation for ChatGPT Search says that ChatGPT usually rewrites a prompt into one or more targeted queries before sending them to search providers, and that it may run further searches after reviewing the first results. Location and, where enabled, memory can influence the rewritten queries.

Other assistants with web access use comparable retrieval steps, but they document them in different detail and with different vocabulary. Avoid assuming that one provider's fan-out behaves like another's. If you report on several assistants, describe each one separately.

What changes for SEO

• From keyword to question space. The relevant unit is the set of sub-questions around a topic. A page that covers the main facets of a topic, as part of a site with real topical authority, has more entry points than a page optimized for one phrase.
• Passages matter. Systems select specific passages. Each important section should state its subject, its conditions, and its answer without depending on the rest of the page.
• Comparisons and constraints. Sub-queries often target comparisons, prices, limits, compatibility, and "for whom" questions. Pages that hide these details lose retrieval chances.
• Entities need to be unambiguous. Clear names for products, organizations, and places help a system match a passage to a sub-query. The principles of entity SEO apply directly.
• Original information helps. If ten pages say the same thing, one of them is enough for the answer. First-hand data, tests, or specific experience give a page a reason to be selected.

What query fan-out does not change

Fan-out does not replace the basics. A page still has to be crawlable, indexable, and useful. It is also not a reason to stuff every imaginable sub-question into one page. Thin sections written only to catch sub-queries add little and can make the page harder to read. Real content depth means answering the questions your audience actually has, in a structure they can follow.

There is no reliable "fan-out score" for a page. Any tool claiming to measure how often a page is retrieved by an AI system's hidden sub-queries is estimating, and the method should be disclosed.

Measurement limits

Google does not document a report that lists the fan-out queries behind a user's AI Mode or AI Overview response. OpenAI does not publish the rewritten queries for other people's conversations either. What you can observe:

• Search Console data for your own property, with AI feature traffic counted inside Web search;
• referral traffic from assistants in analytics, such as visits tagged with utm_source=chatgpt.com;
• your own tests, where some interfaces show which searches were run;
• the sources cited in a versioned set of test prompts, repeated over time.

Tools that "simulate" fan-out usually ask a language model to guess plausible sub-queries. That is useful for brainstorming. It is not evidence of what a search system did. If you compare brand presence across such a test panel, you are measuring a form of AI share of voice, and the method needs to be stated.

How to research likely sub-questions

Combine observed search data with knowledge of your audience:

• People Also Ask boxes, related searches, and autocomplete suggestions for the topic;
• long, question-shaped queries in your own Search Console data;
• comparison and alternative searches around your products;
• questions from support tickets, sales calls, and community forums;
• SERP features that reveal which facets a search engine already treats as separate.

Use language model brainstorming as a hypothesis generator and validate the ideas against these observed sources.

A practical workflow

1. Choose one topic and write down the main question a buyer or reader brings to it.
2. Collect observed sub-questions from search suggestions, Search Console, and customer conversations.
3. Group them into facets: definition, options, comparison, constraints, cost, process, and risk.
4. Map each facet to an existing page or section, and mark the gaps.
5. Improve or create content so each facet has a clear, self-contained answer.
6. Check that the pages are indexable, eligible for snippets, and internally linked.
7. Test a fixed set of real questions in AI search surfaces and record cited sources over several runs.
8. Review changes over time without attributing every movement to a single edit.

Common mistakes

• treating an LLM-generated list of sub-queries as real search data;
• writing one page per imagined sub-query and creating thin, overlapping content;
• assuming special markup or a hidden file can influence fan-out;
• hiding comparisons, prices, or limits that sub-queries often target;
• reporting a "fan-out ranking" that no one can verify;
• ignoring basic indexing and snippet eligibility while chasing AI visibility;
• treating Google's and OpenAI's processes as identical.

The Crawl Foundry perspective

Crawl Foundry cannot show you the fan-out queries a search system issued, and no outside tool can. What it can do is make the observable question space easier to work with. Keyword discovery in Crawl Foundry expands a seed through observed search suggestions, including alphabet variants, questions, comparisons, local searches, and related patterns. The ideas stay usable without paid enrichment. For selected candidates, you can request intent, search volume, difficulty, or SERP context, and Crawl Foundry shows the estimated cost before anything runs.

Useful ideas move into lists and tags in your keyword database, so the facet map of a topic stays connected to the pages you plan. The free People Also Ask tool on the site also builds question ideas from live autocomplete data. Both are observations from search, not a recording of what an AI system did internally. That distinction belongs in every brief that uses them.

Related terms

• Google AI Mode
• AI Overviews
• Retrieval-augmented generation
• Passage ranking
• Semantic search
• Search intent
• People Also Ask
• Long-tail keywords
• Topical map
• Information gain

Review sources

• Google Search Central: AI features and your website
• Google: AI in Search, going beyond information to intelligence
• Google Search Central Blog: Top ways to ensure your content performs well in Google's AI experiences on Search
• OpenAI Help Center: ChatGPT search
• arXiv: Query Rewriting for Retrieval-Augmented Large Language Models