Query fan-out is the step where an AI search engine takes one question, breaks it into several narrower searches, runs them, and writes a single answer from everything those searches return. Google says both AI Overviews and AI Mode may use it, and OpenAI describes ChatGPT search rewriting a prompt into one or more targeted queries. The practical result is that the pages an engine cites are often not the pages that rank for the question you typed. They rank for the sub-questions the engine asked on your behalf. If you track AI visibility prompt by prompt, this is the reason a citation can come from a page you would never have connected to that prompt.
What query fan-out is
A classic search runs one query and returns one ranked list. An engine that fans out treats your prompt as a bundle of questions. “Which AI visibility tool includes Google Search Console?” contains at least three: what AI visibility tools exist, which of them connect to Search Console, and whether Search Console reports AI traffic itself. The engine searches for each part, reads what comes back, and stitches the parts into one reply with links.
Google’s documentation for site owners describes it in those terms: 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 (Google Search Central, AI features and your website, last updated December 10, 2025). Google’s May 20, 2025 AI Mode announcement describes AI Mode breaking a question into subtopics and issuing many queries at the same time, and says its Deep Search mode can issue hundreds of searches for one question (Google, AI in Search: going beyond information to intelligence).
Google is not the only engine that does this. OpenAI’s help article on ChatGPT search says that when ChatGPT uses a search partner, it typically rewrites your prompt into one or more targeted queries, and after reviewing the first results may send additional, more specific queries (OpenAI Help Center, Searching the web with ChatGPT). The name differs by vendor. The behaviour is the same: your words go in, other words go to the index.
How fan-out works, step by step
- The prompt is read for intent. The engine works out what the person is trying to decide, not only the literal words. A question with a constraint (“without enterprise pricing”, “for a small team”) becomes a filter it has to satisfy.
- It is split into sub-questions. Each part of the decision becomes its own search: the category, the constraint, the named products, a definition the answer will need.
- The sub-searches run in parallel. Google describes AI Mode issuing its queries simultaneously rather than one after another, which is why a fan-out answer arrives in seconds.
- Results are read and filtered. Pages from each sub-search are read, and the passages that answer a part of the question are kept.
- One answer is written, with links. The reply cites the pages it leaned on. Because those pages came from different sub-searches, the link list can be wider than a single results page would be.
- Some engines search again. OpenAI says ChatGPT search may send more specific queries after reviewing the first results, so the fan-out can have more than one round.
None of the engines publish the exact sub-queries they ran for a given answer, and the split changes from one run to the next. That matters for measurement, as the next sections show.
What fan-out looks like in real citations
We cannot see an engine’s private sub-queries, but we can see their footprint: the pages it cites. In Cituna’s own tracking of cituna.com, in the scan of October 4, 2026 across seven engines, several answers cited pages that answer only one slice of the prompt.
How to read this table
| Tracked prompt | A page the engines cited | The narrower question that page answers (our reading) |
|---|---|---|
| which AI visibility tool includes Google Search Console | developers.google.com, the Search Central blog post “Introducing Search Generative AI performance reports in Search Console” | Does Search Console report AI search traffic itself? |
| I run marketing at a small B2B SaaS. Which AI visibility tracker covers ChatGPT, Claude and Grok without enterprise pricing? | searchable.com/pricing | What does one specific tool cost? |
| how can I check if Claude recommends my brand | support.claude.com, an article on enabling Claude’s web search | Does Claude search the web at all? |
| how do I know if my website is cited in AI answers | support.google.com, the Search Console Help article “Generative AI performance report (Search)” | What does Google’s own reporting show? |
| AI visibility tool with an API for developers | finseo.ai/developers/api and reachd.ai/developers | Which vendors publish API documentation? |
Source: Cituna’s scan of its own tracked prompts on October 4, 2026, cited pages across all seven engines. Sub-questions are Cituna’s reading, not engine output.
Pages like these turn up across the 34 prompts in that scan. Roundups that answer the whole question are cited too, but the vendor pricing page, the help-centre article and the developer docs get in because they are the best answer to one sub-question.
The engines also differ in how many pages they cite per answer. On the AI-visibility buyer questions Cituna tracks for its own brand, over the 30 days to October 5, 2026, Perplexity cited 21.6 third-party pages per answer, Claude 8.9, Google AI Mode 8.4 and ChatGPT 3.9. A long link list per answer is what you would expect from an engine that runs many searches per prompt, but link counts alone do not prove how many searches ran: ChatGPT also rewrites prompts into several queries and still cites the fewest pages here.
Why fan-out changes how you read AI visibility
A citation does not mean you rank for the prompt. If an engine cites your pricing page for a broad “best tools” prompt, it most likely found that page through a sub-search about price, not through the broad query. Optimising the page for the broad phrase can miss the reason it was picked.
A missing citation is often a missing sub-answer. When a rival is cited and you are not, compare what the cited page answers with what your page answers. Often the cited page holds one fact yours lacks: a public price, a supported-engines list, a plain yes or no on an integration. The guide on what sources ChatGPT cites covers the page types that keep coming up.
Engines fan out differently. On the same tracked questions, the five most-cited sites for Google AI Mode shared no host with the five most-cited for Claude, Grok or Perplexity. Different engines split the same prompt differently and search different indexes, so one engine’s citations tell you little about another’s. Track each engine on its own, as set out in how often AI answers change.
Google rankings still feed the system. For AI Overviews and AI Mode the sub-searches run against Google’s index, so ordinary ranking for the narrower query still matters. Google says no special optimization is needed beyond its normal guidance. Our piece on whether Google rankings affect AI visibility looks at where that link holds and where it breaks.
How to measure fan-out in your own data
You cannot export an engine’s sub-queries, so work backwards from what it cited.
- Track the prompts your buyers ask, daily, on every engine. One run is a sample of one fan-out; the split varies between runs. Repeated runs show which pages keep coming back.
- List every page cited for each prompt, not only the domains. The URL tells you which slice it answered: a /pricing page, a /docs page and a comparison post answer different sub-questions.
- Label each cited page with the sub-question it answers. Keep the labels short: price, integration, definition, comparison, “is it possible”. Within a few days you will have a list of the sub-questions each engine asks for your category.
- Check your own page for each sub-question. For every label, is there a page on your site that answers it plainly in the first paragraph, in server-rendered HTML? If not, that is the gap.
- Re-measure after you publish. A new page that answers a sub-question should start appearing among the cited pages for the parent prompt. If it does not after a few weeks, compare it with the page that is cited instead.
In Cituna, steps 1 and 2 are done for you. Each daily scan stores the answer text each engine gave and the source URLs it cited for every tracked prompt, and the Sources page rolls those citations up by site, with the pages behind each one. It also shows your own pages the engines cite, which is where a page picked up through a sub-question shows itself: on October 4, 2026 our /developers page was cited for a prompt about the cheapest multi-engine tool, a question that page does not set out to answer.
Common mistakes with query fan-out
- Writing one page per guessed sub-query. Fan-out generator tools produce plausible sub-queries, but they are a model’s guess, not the engine’s log. Start from pages that are actually cited.
- Treating the cited domain as the lesson. “They cite G2” is less useful than knowing which G2 page and which sub-question it answered.
- Reading one engine as all engines. In our data, some engine pairs share none of their five most-cited sites.
- Expecting one run to be stable. The same prompt can fan out differently tomorrow. Judge on repeated runs.
Measure the pages behind every answer
Query fan-out is why the page an engine cites so often surprises you: it was the best answer to a question you never typed. To see those pages for your own prompts, Cituna asks ChatGPT, Perplexity, Gemini, Claude, Grok, Google AI Overviews and Google AI Mode your buyers’ questions every day, records the sources behind each answer, and writes the fix for every answer you are missing from.
Drafted with AI assistance from our own research and Search Console data, and reviewed by Rahul A before publishing. Rules and prices change; check the linked official source before you act.