Query fan-out: how AI assistants really search
An assistant does not search the way you do. It takes the question apart, retrieves fragments from many pages and only then writes an answer. That single mechanism decides which pages get quoted.
If you want to end up inside an AI answer, you have to understand how that answer is produced. It is nothing like a ranked list of links, and the difference has practical consequences for every brief you write.
In short
- An assistant breaks one question into several sub-queries. That step is called query fan-out.
- It retrieves fragments of many pages and assembles them into a single answer. That step is retrieval-augmented generation, or RAG.
- Prompts written for an assistant are far longer than search queries, so they carry context a keyword never did.
- Modular pages win: a heading, a short answer, then the detail. Long warm-up paragraphs get skipped.
What a search engine does, and what an assistant does instead
A classic search engine matches a page to a phrase and returns a ranked list. The person does the rest of the work: opening tabs, comparing, deciding.
An assistant does that work itself. It breaks the question into several sub-queries, runs each one separately, pulls back passages rather than whole documents, and writes one answer out of the pieces. The splitting step is query fan-out. The retrieval and reuse of passages is RAG, retrieval-augmented generation.
The consequence is uncomfortable for anyone used to rank tracking. Your page is no longer competing for a position against nine other pages on one query. It is competing for inclusion, one passage at a time, across a handful of sub-queries you never see and cannot look up.
Prompts are not keywords
The input has changed shape, not just length.
- Prompts sent to an assistant are noticeably longer than a typical search query, closer to a full sentence than to a phrase.
- Sessions last longer, because people follow up in conversation rather than starting a new search.
- People describe a whole situation, including budget, constraints and what they already rejected, instead of typing a fragment of a phrase.
All of that context becomes fuel for the fan-out. A question like "which supplier should we pick if we have a small team and no in-house developer" turns into sub-queries about pricing models, implementation effort and support, and each of those goes looking for a different passage. Keyword volume tells you very little about any of them.
What this means for your content
Since the assistant retrieves fragments, the pages that win are the ones a useful fragment can be cut out of. A section that only makes sense after reading the two sections above it is close to useless here. Long introductions that circle the topic before saying anything are skipped entirely.
Keep it modular. Clear heading, a short and specific answer directly underneath, then the elaboration for the human reader who clicked through. That ordering costs you nothing with people, who also prefer to see the answer before the essay, and it is the difference between being retrievable and being invisible. Structured data helps the same cause by making the relationships on the page explicit, which is why schema markup earns its place in an AI-era brief.
Every section has to survive being cut out of the page and read alone.
How to plan for sub-queries
- Write down the customer question as they would actually ask it, in full sentences, with the context they would include. Not the keyword, the question.
- Break it into 5 to 10 sub-questions the assistant would plausibly need answered before it could reply. Price, fit, effort, risk, alternatives, proof.
- Give each sub-question its own section, with the heading phrased as the question and the answer in the first two sentences.
- Add data and sources. A passage with a specific figure and a named source is far easier to quote than a passage of adjectives, and it is what makes a model treat you as worth citing at all.
Done properly this replaces the old keyword map with a question map. It is also why long-form content is not dead: a long page split into self-contained sections gives the retriever many chances to find something usable, while a long page written as one continuous argument gives it none.
How to tell whether it is working
Rank tracking will not show you this. The practical method is to keep a fixed list of the real questions your customers ask, run them through the assistants your market uses, and record whether your domain appears and in which sub-part of the answer. Run the same list on a schedule, because answers drift. More on the sources those answers come from in where AI citations come from, and on how the disciplines fit together in SEO vs AEO vs GEO.
Common questions
What is query fan-out?
It is the moment an assistant breaks one user question into several sub-queries, searches each of them separately and then assembles a single answer out of what came back.
What is RAG?
Retrieval-augmented generation: the model retrieves fragments of pages from the web and uses them as context for the answer it writes, instead of relying only on what it memorised during training.
How should we write so that an assistant quotes our page?
Modular and answer-first: a heading phrased as the question, a short and specific answer underneath it, then the detail with data and sources.
Do long articles still make sense with query fan-out?
Yes, as long as they are split into sections that stand on their own. The assistant lifts single fragments, so every section has to defend itself without the rest of the page.
Read next
Find out whether AI recommends your company.
Start with the free SEO and GEO audit, delivered in 5 working days. We check how the models describe your brand and hand back a prioritised list of changes.