Fan-out queries: the searches an AI assistant runs behind your prompt

What are fan-out queries in AI search?

When an assistant decides a question needs current information, it grounds the answer in web search instead of recalling it. Before it can search, it has to turn conversational language into something a search index will respond to. That rewriting step is the fan-out.

Vendors publish little about how the rewriting works, so treat the following as the observable shape rather than a specification. Across the assistants we watch, fan-outs typically share four traits.

• There are several. Commonly two to six per prompt, more for comparison questions that need one search per option.
• They are shorter than the prompt. Personal context, tone and politeness are stripped. What remains is a noun phrase plus constraints.
• They are often dated. Assistants frequently append the current year or month, because recency matters for recommendations, prices and opening hours. This is why a page with no visible date, or one that never mentions the current year, can quietly underperform.
• They split the constraints. A prompt with a location, a time and a quality bar tends to become separate queries rather than one long one, because long queries return thin results.

Why fan-out queries decide your visibility

The practical consequence is short enough to put on a wall. You are retrieved by the fan-out, not by the prompt. Nobody optimises a page for "we need a physiotherapist in Business Bay who can see my dad on a Saturday". The queries that actually run are ordinary, and they are the ones you can compete for.

That reframes three familiar problems. First, keyword research is not dead in this channel, it has moved: the target is the machine's reformulation rather than the human's question. Second, a business can be an excellent match for the prompt and still be absent, because it was never returned for any of the searches the assistant chose to run. Third, the answer has room for three to five names, so the retrieval stage is where almost all of the loss happens, as set out in how AI assistants choose businesses.

The fan-out also explains why AI visibility work and search work overlap so heavily without being the same thing. The queries run through search infrastructure, so ranking still gates the candidate pool. What differs is which queries run, and what happens to the results afterwards. That distinction is the subject of GEO vs SEO.

How to see fan-out queries yourself

You do not have to take this on trust. Both major assistants expose the searches they ran, one awkwardly and one plainly. Interfaces change often, so expect the details below to drift.

ChatGPT: the network tab

Ask a recommendation question in a desktop browser with web search active. Open your browser developer tools before you send it, go to the Network tab, and watch the streaming conversation response. Inside that stream the search-related events carry the reformulated queries, in recent builds under a field named search_model_queries. Those strings are the fan-out. Field names are an internal detail and OpenAI can rename or remove them without notice, so if you cannot find it, search the response body for a fragment of a query you expect to see.

Claude: expand the search chips

Claude is easier. When it searches, the answer carries a "Searched the web" line. Click it and the individual queries are listed in the interface, along with the sources fetched for each. No developer tools required. This is the fastest way to show a client, in one screen, that their prompt became four searches none of which named them.

Perplexity and Google AI Overviews

Perplexity displays related searches and cites sources prominently, which makes the retrieved set easy to audit even where the exact internal queries are not shown. Google AI Overviews cite links inline. Neither is a full view of the fan-out, but both tell you which pages survived it, which is the part you can act on.

How to track if ChatGPT mentions my brand

Here is the method we run daily against our own site. We are describing it so you can copy it, not selling it. It needs a search tool, a text editor and about thirty minutes a day, and it is deliberately boring.

1. Freeze a seed prompt list. Write the prompts a real customer would type to find a business like yours. Full sentences with real constraints. Never include your own name. Keep the list fixed, because a changing list makes every comparison meaningless.
2. Simulate the fan-out, dated.For each seed prompt, write the two to four searches an assistant would plausibly issue for it today, with today's real date or month in the phrasings where an assistant would naturally put one. This is a reconstruction, not a capture, and it should be labelled as such in your own notes.
3. Run those searches for real. Issue each reconstructed query as an ordinary web search. Record three things: the query, whether one of your URLs appears and which one, and which competitors appear.
4. Diff against yesterday.Store one small file per day. The interesting output is not today's total, it is the change: query shapes that are new, results you gained, results you lost.
5. Feed the gaps back into the site. Any fan-out phrasing where you are absent is a content brief. Give it a heading in your own words, answer it in the first two sentences, and publish. Then wait for the phrasing to reappear in a later run and see whether you now show up.
6. Sample the assistants directly, less often. Separately, ask the seed prompts in ChatGPT, Claude and Perplexity, several times over several days, and log the appearance rate and any factual errors stated about you. A confident wrong answer is worse than absence.

Be honest about what this measures. Presence in the results of reconstructed fan-out queries is a leading indicator, not proof that ChatGPT named you. Crawler-hit trending and real impression data are separate instruments, and if you have not wired them up, say so rather than implying them. We hold ourselves to the same rule in our own reports.

AI visibility monitoring tools

A market of commercial trackers now exists, and some of it is good. The question to ask any of them is methodological rather than commercial.

No tool can guarantee a mention, and any vendor implying otherwise is describing something they do not control. What tooling buys you is frequency and record keeping. The judgement about what to publish next is still yours.

What to do with a fan-out you keep losing

Treat a repeatedly missed query the way you would a missing product page. Check the mechanics first: confirm AI crawlers can reach you, which is covered in the AI crawler robots.txt guide, and confirm the relevant facts exist as server-rendered text rather than inside an image. Then check the language. If the fan-out says "open Saturday" and your site says "flexible weekend appointments", you are not answering the query. Then check corroboration, because a claim only you make is weak evidence. The rest of the sequence is in the GEO checklist, and the wider definitions live in what generative engine optimization is.