How Query Fan-Outs Work in AI Search
The four stages of a fan-out, from prompt to decomposition to retrieval to citation, and how each stage shows up in the data a tool can give you.
A fan-out is not one event. It is a short pipeline, and each stage leaves different evidence. If you want to judge which tools measure what, it helps to lay the pipeline out and mark where the data comes from.
Stage 1: does the prompt trigger search?
The model first decides whether it needs the live web. Questions about current products, comparisons, prices, and "best" lists tend to trigger search. Timeless questions may not. No search means no fan-out, and no citations to earn. When a tracked prompt shows a mention but no sources, this is a likely reason.
Evidence: the final answer and whether it carries citations.
Stage 2: the prompt is split into queries
The engine writes its own queries. According to Promptwatch's ChatGPT query fanouts report, one prompt can produce 3 to 8 or more, though tracked averages ran lower: about 2.15 per response in early December, about 1.84 in early March, and 1.0 in April. Those queries also got terse. Average length dropped from about 117 characters to roughly 53.
The report reads this as ChatGPT searching more like someone typing keywords into Google. The practical upshot is that titles and headings phrased as short, entity-first search terms match better than conversational ones.
Evidence: the list of generated queries, which only fan-out aware tools capture.
Stage 3: retrieval
Each query returns sources. How many make it into an answer varies by engine. The Promptwatch average-sources data has ChatGPT at roughly five per web-search response, and Google AI Overviews and Perplexity at roughly ten. A page that is sixth for a query is invisible in one engine and potentially cited in another.
This stage also depends on access. A crawler that never fetched your page, or hit an error doing so, cannot retrieve it.
Evidence: crawler logs and the set of cited sources per response.
Stage 4: synthesis
The model writes an answer from what it read. It may cite some sources and ignore others. The user sees only this stage.
Evidence: mentions, sentiment, and citations in the answer text.
Mapping stages to tool categories
| Stage | What it tells you | Needs |
|---|---|---|
| 1. Trigger | Whether search ran | Response capture |
| 2. Decomposition | Which queries were generated | Fan-out tracking |
| 3. Retrieval | Who was fetched and cited per query | Citation analytics, crawler logs |
| 4. Synthesis | What the user read | Prompt tracking |
Most tools in our directory cover row four. Fewer reach row two. Very few connect all four rows to traffic and content, and that is why Promptwatch leads our ranking. Its prompts carry fan-outs, volumes, and difficulty. Its citation analytics break down pages, domains, Reddit, and YouTube. Agent Analytics covers crawlers such as ChatGPTBot and ClaudeBot, including the path from crawl to citation. Visitor analytics attach conversions to AI-referred traffic.
What changes between engines
Do not assume ChatGPT's behavior transfers. The source counts above differ, and Google describes its own fan-out inside AI Mode. The reliable approach is to track the same prompt across engines and compare. Promptwatch monitors ChatGPT, Gemini, Claude, Perplexity, and others through real UI monitoring plus Google AI Overviews and AI Mode, which is what makes that comparison possible on paid plans. The free Explore plan is ChatGPT only with 10 prompts.
A caution on precision
Fan-outs are sampled, not guaranteed. Running the same prompt twice can produce different queries. Treat a single run as an example and look for patterns across repeated checks. That is also why trends matter more than snapshots, and why prompt trends and citation trends, which show what changed between checks, are worth more than a one-time export.
How to use this
When a page fails to get cited, walk the stages. Did search trigger? Which queries ran? Was your page fetched? Who won retrieval? The answers point to different fixes: rewrite the heading, fix the crawl error, or build a page for the missing query. For the content side, see how to use query fan-outs to write blog posts. For definitions, see what query fan-outs are.