Q
What is query fan-out?
A
Query fan-out is how an AI turns one prompt into several internal web searches. Today's ChatGPT tends to draft its own answer first, then run searches that verify the claims in that draft.
Free, no sign-up required
Query Fan-Out Analyzer
See how ChatGPT builds its answer to a prompt, from the hidden draft it writes before searching to the final cited answer.
Draft answer, research plan, real search queries, retrieved-page scoring, and a brand check. See why the answer came out the way it did, which prompt trackers can't show you.
Answer before searchReal fan-out queriesPassage scoringBrand funnelCSV export
How an AI search answer is built
ChatGPT doesn't just search your question. It first decides what it thinks the answer is, then searches to back it up. This tool shows that chain step by step.
- Step 01
The person and the prompt
Along with the prompt, the model gets context about the user, like their role and goals. Change the user and everything downstream changes.
- Step 02
Answer before search
Before searching, the model drafts an answer from training knowledge. If you're not in that draft, search has to win you back in.
- Step 03
Research plan
The draft's claims are decomposed into broad discovery, claim verification, and nice-to-have searches: the query fan-out.
- Step 04
Evidence targets
For each search, the model imagines what ideal evidence looks like (HyDE), and retrieved pages are compared against it.
- Step 05
Retrieval and selection
Many pages get retrieved, but only some get used. Being found and being used are two different things.
- Step 06
Citations and final answer
The final answer is written from the surviving evidence, with citations. That last step is all a prompt tracker sees.
- Step 07
Brand diagnosis
See exactly where your brand appears and where it drops out, so you know what to fix next.
Why looking at the final answer isn't enough
"We weren't cited" doesn't tell you what to fix. Knowing which stage you lost at makes the next step concrete.
1
Were you missing before search started?
If the model's draft doesn't include you, searching mostly confirms the competitors it already had in mind. Check this stage first.
2
Retrieved isn't the same as used
Pages that get retrieved but not cited usually score low against the evidence targets. See which passages lost.
3
The plan checked against real searches
Planned searches that ChatGPT actually ran are marked, and real queries outside the plan are flagged.
4
Directly actionable for content
Evidence targets show the shape of text the AI is looking for, so you can write pricing, comparison, and proof passages that match.
Other Free Tools
Try our other free tools for AI search optimization
NEW
AI Search Content Audit
See which facets of a question your page answers, measured with vectors.
NEW
AI Pre-training Data Checker
See whether your site is in Common Crawl, where many AI training sets start.
NEW
AI Recruitment Journey Mapper
Break down candidates' AI search and auto-generate a recruitment journey map.
Frequently Asked Questions
Q
What is an evidence target (HyDE)?
A
For each search, the model writes a hypothetical passage that looks like ideal evidence. Retrieved passages are scored by how semantically close they are to it. The figures can be made up; what matters is the shape of the passage.
Q
What changes when I add who is asking?
A
The same prompt from a different person produces a different draft, different searches, different retrieved pages, and a different final answer. We pass the user context separately, the way memory works, rather than stuffing it into the prompt.
Q
Why does ChatGPT sometimes not search at all?
A
When ChatGPT decides it can answer from training knowledge alone, it skips web search. You still see the draft answer and research plan, so you can tell what the model is assuming.