How Query Fan-Outs Work in AI Search

When a client page is not cited, we debug it the way you would debug a failing pipeline: stage by stage. A query fan-out is that pipeline. Here is how we read it in audits, and what evidence we pull at each step.

Stage 1: did the prompt trigger a search?

Not every prompt does. Stable factual questions can be answered from what the model already knows. Comparisons, "best" questions, pricing, and recent topics usually trigger search. If there is no search, there is no fan-out and nothing to optimize. We check this first, because it saves clients from chasing prompts that do not behave like search.

Stage 2: what queries were generated?

This is the fan-out itself. Promptwatch's ChatGPT query fanouts report says a prompt can produce 3 to 8 or more searches. The tracked averages were lower and moved: about 2.15 per response in early December, about 1.84 by early March, and 1.0 in April. Query length dropped from roughly 117 characters to roughly 53.

We read that as a warning against over-reliance on any one snapshot. When we present fan-outs to a client, we date them.

Stage 3: which pages were retrieved?

Each query returns sources. Promptwatch's average-sources data has ChatGPT using roughly five per web-search response, and Google AI Overviews and Perplexity roughly ten. So the same page can be cut in one engine and included in another. We compare engines instead of assuming ChatGPT's result applies everywhere.

The second question at this stage is crawler access. If the AI crawler never reached the page, nothing else matters. Agent Analytics in Promptwatch logs crawlers such as ChatGPTBot, ClaudeBot, and PerplexityBot, including errors and the path from crawl to citation. Plenty of audits end here, with a blocked bot or a server error nobody noticed.

Stage 4: what got cited?

The model writes an answer and cites some of what it read. The client sees this stage only. We use citation analytics to see which page and domain won each query, which tells us whether we are fighting a competitor, a review site, or a forum thread.

A worked audit, in outline

Suppose a client sells accounting software and is missing from answers to "best accounting software for freelancers."

Check the prompt triggers search. It does.
List the fan-out queries. Say they include a comparison, a pricing query, and an integrations query.
Check crawler logs for the client's comparison page. Clean.
Check citations. A third-party listicle holds the comparison slot.
Decide. Either improve the client's page so it is a better answer than the listicle, or pursue inclusion in the listicle. Usually we do both.

This is a hypothetical to show the shape of the work, not a client result.

Why we use one platform for this

We run client programs on Promptwatch because every stage above has a data source inside it. Fan-outs, volumes, and difficulty on the prompt. Citations at page and domain level. Crawler logs. Visitor analytics for what AI-referred traffic does after the click. Stitching that together from separate tools costs hours per client per month, and the stitching is where mistakes creep in. Tools that only track prompts give us stage four and leave us guessing on the other three.

What this means for content

Every fix maps to a stage. A missing query means a new page. A retrieval loss means a better answer or a stronger source. A crawl error means a technical fix. Only the first involves writing. We cover that in how to use query fan-outs to write blog posts, and the definitions are in what query fan-outs are.

If you want a fan-out audit on your own site, write to hello@1001seomedia.com.