How does ChatGPT build its query fan-out about your brand?
When ChatGPT answers a question, the query fan-out it runs doesn’t target what you asked. It quietly rewrites your question — often into another language — and searches for that instead. Vercite analysed 17,806 ChatGPT fan-out queries from our own platform data to see exactly what it looks for.
What is a query fan-out?
It’s ChatGPT’s query fan-out: the set of background searches it runs to answer you, queries nobody typed. Your brand isn’t competing for the user’s question; it’s competing for the queries ChatGPT derives from it.
A user types one prompt into ChatGPT, in their own words and their own language.
Before answering, ChatGPT silently issues its own background search queries. That set is the query fan-out.
Each fan-out query pulls sources from the web. Whatever ranks for those strings becomes the raw material.
The reply cites the brands and pages the fan-out surfaced — not the ones that rank for your original prompt.
How many fan-out queries does ChatGPT run per answer?
Fewer than you’d think — about two. In a single answer ChatGPT’s query fan-out is roughly two distinct queries. Ask the same question repeatedly over weeks, though, and it explores a far wider set as it keeps re-phrasing — about 38 distinct fan-out queries in total.
1.9 distinct fan-out queries per response, on average.
The wide number is a vocabulary built up over time, not a single burst. ChatGPT keeps your topic but varies how it phrases the fan-out from run to run.
Does ChatGPT’s query fan-out use your language?
Often not. 41% of its fan-out queries are in a different language than your prompt — and when the prompt is in a local language, that means the question is rewritten into English. To rank in the query fan-out ChatGPT actually runs, English content can matter even when your customer typed Swedish.
Roughly four in ten fan-out queries are rewritten into another language before the fan-out runs.
How much does ChatGPT rewrite your prompt?
Heavily. 64% of the words in ChatGPT’s fan-out queries never appeared in your prompt. Part of that is translation — but not all: even when it keeps your language it swaps out roughly half the words, and when it translates, almost all of them.
When the fan-out stays in your prompt’s language, about half the words are still new. It genuinely reformulates.
When it translates, nearly every word changes — translation and reformulation compound.
Only 16% of fan-out queries stay close to your wording. The other 84% reword it meaningfully — so exact-match optimisation to your customer’s phrasing isn’t enough. 64% new words · 30% word overlap on average.
What is ChatGPT actually asking for?
It works the whole funnel — and it likes to be taught. Location and best-of questions lead, but a striking share of fan-out queries ask ChatGPT to define, explain or guide. Share of distinct fan-outs that signal each intent; a query can carry more than one.
Which words does ChatGPT use most?
How, best and what lead — followed by a distinctive guide-and-define streak that shows ChatGPT asking to be taught the topic. The mint figure is each word’s share of all distinct fan-outs; multilingual variants are folded into one label.
What do ChatGPT’s fan-out queries look like?
Long, descriptive and keyword-shaped — not the question your user typed.
Long, qualified phrases — not short keywords.
Almost always a keyword string, not a natural question.
The rest reword your prompt in the fan-out.
How do you get cited by ChatGPT?
Optimise for the query ChatGPT writes, not the one your customer types.
- What we measured
- A query fan-out is a background search ChatGPT issues while answering. “Distinct” counts unique query strings; we deduplicate identical strings within a prompt.
- Per answer vs cumulative
- “~2 per answer” is the average within a single response. “~38” is the cumulative distinct set a prompt accumulates across repeated checks over the window — not queries fired in one answer.
- Language
- Detected algorithmically. On very short strings this is approximate; the direction (heavy translation into English) is robust, the exact percentage is indicative.
- Intent & words
- Tagged with multilingual keyword patterns, which lean English and can overlap, so intent shares are directional. New-word ratios compare each fan-out query’s words against its prompt.
- Scope
- This study covers ChatGPT only. Vercite tracks Gemini, Perplexity and Google AI separately; engine behaviour changes over time, so figures are a snapshot of this window.
Try Vercite for free today
Monitor your brand across ChatGPT, Gemini, Perplexity & Google AI