How do you use fan-out queries for GEO?

Published 3 November 2026. Updated 4 November 2026.

To use fan-out queries for GEO, record the searches ChatGPT, Claude and Grok actually run for the prompts your buyers ask, clean and group them into themes, and match each theme to a page on your site. Then read each turn to see whether your page was fetched, cited or missed, and fix the gap that turn shows.

This page sets out that workflow step by step, with what each assistant's search data tells you and what to fix first. If fan-out is new to you, start with our guide to query fan-out, and see how ChatGPT, Claude and Grok search the web for the search step itself.

What does GEO with fan-out data mean?

It means planning GEO work from the search queries an assistant wrote and ran, instead of from the prompt a person typed or a keyword list. GEO (generative engine optimisation) is the work of getting a site found, read and cited in AI answers. Fan-out queries are the searches an assistant sends to the web before it answers, often several for one prompt.

Those searches decide which pages get into the answer. An assistant reads the pages its searches return, so a page that never comes up for those searches is never read, however well it answers the prompt.

Why work from fan-out queries instead of prompts or keywords?

Because the assistant cites from what its own searches retrieve, and those searches are not the ones you would guess. Mahe Chen and Nick Koudas studied 158,099 answers from ChatGPT, Gemini and Perplexity in "What Drives AI Citations?" (October 2026). Every page ChatGPT cited was one its own search had retrieved, and the sub-queries it wrote rarely matched what Google or Bing returned for the same question.

Which pages the search returned explained 32% of the difference between cited and uncited pages on the same site, against 2.8% for the pages' own properties. A prompt or a head keyword tells you the topic. The fan-out queries tell you the exact searches a page has to come up for. Our page on how AI citations work in ChatGPT, Claude and Grok covers these studies in more depth.

Where can you get fan-out queries?

From four places, and they are not the same kind of data. Know which one each line of your list came from before you plan work around it.

Source	What you get	Watch out for
Recorded in the chat app while you send	The searches ChatGPT, Claude or Grok ran for that turn, with the pages fetched and cited	It has to be running when you send. It shows one account's session, not everyone's
Runs through a model's API	Repeatable runs at scale, easy to log	The API can behave differently from the app people use
Tools that predict sub-queries	A quick list of likely searches for a topic	Predictions, not searches any assistant ran. Good for ideas, weak as evidence
Bing Webmaster Tools, AI Performance report	Grounding queries where your site appears and the pages cited for each	Bing's data about your own site only. It does not show what ChatGPT, Claude or Grok searched

The API caveat comes from the studies themselves. MJ Cachón's study of branded ChatGPT prompts (August 2026) ran GPT-5.5 through OpenAI's API and notes that the results may differ from what real users see. Bing's Topics Portfolio post (13 October 2026) describes the report's beta view, which groups grounding queries into topics and lists opportunity queries where other sites are cited and yours is not. It only appears "If there is sufficient data for your site."

What does each assistant's search data tell you?

Each assistant runs more than one kind of search, and each kind points to a different fix. These are the kinds Only Fan Out records from ChatGPT, Claude and Grok.

Assistant	Kind of search	What it tells a site owner
ChatGPT	Web	The phrasing a page needs to come up for in an ordinary search
ChatGPT	Business	A local search, sometimes with a place name and several queries bundled together. Check your business details and location pages
ChatGPT	Image	ChatGPT looked for pictures. Check your images have descriptive file names and alt text
ChatGPT	Find	A look-up inside results ChatGPT already had. No new search, so nothing to rank for
ChatGPT	Product cards	Price, rating and merchant shown apart from the cited sources. For shops, check your product data
Claude	Web	Searches grouped by turn, with citations attached to the sentences that use them
Claude	Places	Claude plots places on a map without citing them, so a map place never shows up as a citation
Grok	Web and page browses	A page browse means Grok opened a specific URL. Check that page loads and states the facts plainly
Grok	X search	Grok searched posts on X. What your account and others say about you there feeds the answer
Grok	Connected Gmail, Drive or Calendar	Counted, but the query text is never stored. Leave these out of your list

A turn can mix kinds. A question about where to buy running shoes in Leeds might send ChatGPT a web search, a business search and a find, and only the first two are searches your pages can come up for.

How do you run a GEO workflow with fan-out data?

In six steps, repeated each month on the same set of prompts.

1. Write 10 to 20 prompts in your buyers' words. Include comparisons, price questions and, if you serve a place, local ones.
2. Run each prompt in ChatGPT, Claude and Grok with recording switched on, in a fresh chat, more than once. The queries change from run to run.
3. Export each turn and clean the list, using the rules in the next section.
4. Group the queries into themes, such as pricing, comparisons or setup, and note the page on your site that should answer each theme. Write "no page" where there isn't one.
5. For each theme, check whether your page comes up when you run those queries yourself, and whether the assistant fetched it, cited it or did neither.
6. Fix in the order set out below, then run the same prompts again next month and compare by theme.

Themes matter more than exact strings because exact strings rarely repeat. Seer Interactive pulled fan-out queries from Gemini 3's API for 100 prompts, twice a day for a week. Across 13 runs, only 8 exact queries came back every time. Our fetched vs cited sources post shows how to log the fetched and cited pages for each turn. For a first baseline, our GEO audit with fan-out data adds the crawler checks, an audit sheet and a scorecard by theme.

How do you clean a raw list of fan-out queries?

Take out the lines that are not searches you can come up for, and split or mark the rest so each one counts once.

• Drop ChatGPT's find lines. They search inside results ChatGPT already had.
• Split business searches that bundle several queries with semicolons, and keep the place name with each.
• Drop Grok's searches of connected Gmail, Drive or Calendar. There is no query text to work with.
• Mark any search with site: and your own domain. The assistant was checking your site for one fact, and that needs a different fix.
• Mark turns where ChatGPT withheld the queries and showed only a label such as "Searching the web". Count them as missing data, not as turns with no searches.
• Keep the run date and the assistant on every line, so next month's list compares like with like.

What should you fix first?

Start where the assistant searched and found nothing of yours, then move down the list. Each state in a turn points to its own fix.

What the data shows	What to do
A theme the assistant keeps searching, and no page of yours answers it	Write one page for it, with a title and opening that answer the query in plain words
You have a page, and it was not fetched	It did not come up for the query. Work on ranking for that phrasing: title, headings and links from your related pages
Fetched, and not cited	Put a short, direct answer to the query near the top of the page
A site: search on your own domain	Make sure that fact sits in plain HTML on its own crawlable page
Your brand named, and no page of yours fetched	The model knew you without reading you. Check what it said is true, and publish a page that states the facts
Cited	Check the cited sentence says what your page says, and keep the page current

Two figures back the first two rows. In Ahrefs' study of 1.4 million ChatGPT prompts, cited pages had titles closer to the fan-out query (0.656 similarity) than to the prompt (0.602). In Cachón's API study, 16.6% of ChatGPT's sub-queries targeted the brand's own domain through site: searches. Both are correlations from one engine. Neither study shows how OpenAI chooses sources internally.

What changed for ChatGPT search data in September 2026?

Since 14 September 2026, ChatGPT sends its query text only in the live answer stream. A saved chat still holds the pages it fetched and cited, but reopening it or exporting it later shows no queries.

For a GEO workflow, that means recording while you send. Going back to old chats gives you sources without the searches that found them. On some turns from ChatGPT's thinking model, the query text is withheld even in the live stream. Our walkthrough on how to see what ChatGPT searched compares the methods that still work.

How do you measure whether it worked?

Run the same prompts again and count, for each theme, how often your pages were fetched, cited and named. Keep citations and mentions as separate counts, because Chen and Koudas found ChatGPT names brands it remembers with no page behind the mention.

If your site is verified in Bing Webmaster Tools, the AI Performance report adds Bing's own grounding queries and the pages it cited for each. Expect the counts to move around from month to month, since the queries do, and judge by theme over several runs. No tool can promise citations. Only Fan Out shows the data and does not change rankings or citations.

Only Fan Out is a free Chrome extension that records, for each turn in ChatGPT, Claude and Grok, the queries the assistant ran in order, every page it fetched and which of those it cited. Add your brand or domain to its watchlist and it highlights every mention across queries, fetched sources, cited sources and answers. Export a turn or a whole conversation as CSV, Markdown or JSON for your theme sheet. Nothing leaves your browser.

Related questions