Answer Radar / Query fan-out

Query fan-out: the searches AI engines ran before they answered

One buyer question becomes several web searches inside ChatGPT, Gemini or Claude, and those searches pick the pages that get cited. Answer Radar records them per answer, so you can see the retrieval layer nobody types.

Start tracking What fan-out is, in depth

Updated 2026-09-04

What the fan-out view records

The closest thing to a keyword list an answer engine will give you, with its limits printed on it.

For every answer in a run, Answer Radar stores the searches the engine reported issuing, next to the answer text and the citations. The fan-out view groups those searches by prompt and then by engine, with how many answers issued each query and how many times. You scope it to one run or to a date window. Where an engine does not report executed searches, the segment says so and gives the reason; it never fills the space with a list we cannot stand behind.

A real fan-out, run on 4 September 2026 on our own project

The prompt was one of 21 we track for linkeddit.com. Three engines reported the searches they ran; the fourth is shown unmeasured with its reason.

answer_fanout_get({
  project_id: "<your project id>",
  run_id: "19910a90-4d5d-4549-8b7f-1116184f9b5a",
  mode: "summary"
})

The exact call, as issued through the Linkeddit MCP connector. project_id is your project; the run id is ours.

Prompt: "Which competitive intelligence tools use Reddit and review site data?"

ChatGPT: Five searches shown, eight distinct: "Reddit" "competitive intelligence" tool social listening / Brandwatch competitive intelligence reddit / G2 reviews competitive intelligence software social listening Reddit / Klaviyo competitive intelligence uses reddit / Kompyte Reddit data. Three omitted in summary mode.
Gemini: Two searches: competitive intelligence tools reddit review site data / social listening sentiment analysis competitive intelligence reddit G2 Capterra.
Claude: One search: competitive intelligence tools Reddit review site data.
Perplexity: Unmeasured. Reason: related_queries_not_executed_searches. One sample counted so the segment is not silently missing.

Run 19910a90, 15 prompts with fan-out, 58 engine segments, 35 usable. Provenance executed_searches, parser fanout-v1. Quoted verbatim from the tool response.

How the four engines fanned out on the same 15 prompts

Same run, same prompt set, 4 September 2026. The shape of each engine's retrieval is different enough to change what you write.

Engine	Usable segments	Queries reported	Shape
ChatGPT	8 of 15	44 distinct across 9 answers	About five searches per answer. Names vendors and review sites: G2 Klue alternatives, Klue pricing, Kompyte Reddit data.
Gemini	15 of 15	34 distinct across 16 answers	About two per answer. Restates the prompt, then adds a year or a category word: best Klue alternatives 2025 2026.
Claude	12 of 15	12 distinct across 12 answers	One search per answer, close to the prompt as typed.
Perplexity	0 of 13	none reported	Reports related topics it suggests, not searches it ran. Shown as unmeasured with the reason, not as a list.

A segment is one engine on one prompt. ChatGPT reported searches on 8 of 15 prompts and answered the other seven without reporting any, which is the first limit above in practice. Where it did search, it did the most work per answer and was the only engine to put a competitor name and the word pricing into the same search.

What to do with it

Three reads that turn a list of searches into a page to write or fix.

Recurring phrasings

On our run, every engine that searched for Klue alternatives added a year or the words competitive intelligence tools. A page whose heading answers that phrasing is a page the engine can find. A page titled for the brand alone is not.

Competitor names inside searches

ChatGPT searched Kompyte Reddit data and Brandwatch competitive intelligence reddit on a prompt that named no vendor. The engine framed the question around rivals before it read a single page. That is the moment to be present, and the content gap view says whether you are.

Review sites and formats

Searches that name G2 or Capterra, or ask for alternatives, tell you what kind of source the engine expects to cite. If your category is retrieved through review sites, your review profile is a ranking page whether you maintain it or not.

Two limits, reported rather than hidden

Both come back in the tool's own response. A list that looks complete and is not would be worse than an honest gap.

A provider may issue searches it does not report, and the queries it chooses vary between runs and between models. The list is evidence of how a prompt was retrieved against, not a complete or repeatable set.
Perplexity reports related topics it suggests rather than searches it ran. Its segments carry a reason and no queries. On our run that was 13 of 58 segments.
Counts are answers issuing a query, not people searching it. No search volume is attached because no engine publishes one.
Fan-out is per engine. Pooling queries across engines produces a list nobody ran.

Frequently asked questions

What is query fan-out?+

Query fan-out is the set of web searches an answer engine issues while answering one question. A buyer types one prompt; the engine turns it into several searches the buyer never typed, and those searches decide which pages get read and cited. Answer Radar records them per answer and groups them by prompt and by engine, so you can see the retrieval layer between the question and the answer.

Which engines report their fan-out?+

On our 4 September 2026 run, ChatGPT, Gemini and Claude reported executed searches. Perplexity reports related topics it suggests rather than the searches it ran, so Answer Radar marks its segments unmeasured with that reason instead of showing a plausible list. Google AI Overviews is not in the run yet.

How is this different from keyword research?+

Keyword research estimates what people type into Google and how often. Fan-out is observed behaviour: the phrasings an engine actually reached for on your prompts. No engine publishes search volume for its internal queries, so none is attached. Use fan-out to phrase pages the way engines search, and Search Console for what real buyers typed.

Is the fan-out list complete?+

No, and the tool says so in its own response. A provider may issue searches it does not report, and the queries change between runs and between models. Treat the list as evidence of how a prompt was retrieved against, not as a repeatable set. In summary mode the tool also omits some queries per segment and reports the omitted count; evidence mode returns every query with the sample ids that issued it.

What do I do with the fan-out once I have it?+

Look for phrasings that recur across answers, competitor names inside searches, and searches that name a review site or a comparison format. Each one is a page the engine is trying to find. The content gap view then checks whether a page you already have covers the requirement, so the fix is aimed at a page and not at a score.

Can I get the fan-out from Claude or Cursor instead of the app?+

Yes. The view is the answer_fanout_get tool in the Linkeddit MCP connector, scoped by run_id or by a start and end date. The browser app and the connector read the same records.

See which searches decide what the engines say about you

Fan-out is part of Answer Radar, included with the Compete plan, in the browser and through the MCP connector.

See pricing How Answer Radar works