Query Fan-Out: the searches AI models run to answer prompts

When ChatGPT or another AI model answers a question, it frequently runs several internal web searches to gather information first. Those internal searches are called fan-out queries, and they are a goldmine: they show the exact phrasing AI models use to research your category.

With Query Fan-Out, you can:

• See the real internal search queries ChatGPT and other models run while answering your prompts
• Turn high-frequency fan-out queries into a keyword list your content should target
• Check who ranks for each fan-out query today and spot the ones where your site is missing
• Uncover the narrow, specific sub-questions that never rank on Google but decide AI answers

Open Prompts, then select the Query Fan Out tab. It is available on all plans.

Query Fan Out is available in embedded portals and only shows queries from the embedded project.

What it does

• Lists the internal search queries AI models executed while answering your tracked prompts
• Example: for the prompt "how to run targeted ads", ChatGPT might internally search "targeted local advertising strategies"
• Fan-out queries are also visible inside each individual response detail view, so you can see what was searched for one specific answer
• Click any query to open its Google or Bing results and see which pages currently win that search

How to use it

1. Open Prompts, then select Query Fan Out
2. Filter by date range, tag, country, or language to focus on a topic area
3. Look for recurring fan-out queries across many prompts; these are the searches AI models rely on most in your category
4. Check who ranks for those queries today. If your site is absent, that is a concrete content gap: a page targeting that query has a direct path into AI answers.

How many fan-out queries are normal?

It varies more than you might expect. AI models frequently run several internal searches for one answer, but an execution can also surface just one, or none at all. The count depends on the model and on whether the provider exposes its internal searches, so seeing a single fan-out query on a prompt is normal, not a tracking problem. Coverage also differs per AI model, which is why some models contribute many more fan-out queries than others.

Tips & notes

• Fan-out queries often differ from the original prompt phrasing. Optimizing only for the prompt wording misses the searches the model actually runs.
• Treat high-frequency fan-out queries as a keyword research source alongside your traditional SEO tools
• The fan-out data is exportable from the Exports page (profile menu > Exports), and it is included in the Master XLSX export

Synthetic query ideas

The Synthetic tab shows inferred search ideas based on your tracked responses. Each result starts from one saved, successful response and uses its prompt, answer, country, language, and observed queries. It can suggest up to five related queries, removes duplicates, and excludes the original prompt and observed queries. Long answers use a labeled excerpt.

Synthetic ideas are labeled Synthetic. They show possible topics to explore, not searches an AI model actually ran.

Open Prompts > Query Fan Out, switch to Synthetic, and choose Generate queries. The picker opens over the results and shows the latest successful, non-empty response for each prompt and model in your filters, leaving out the ones you have already generated. It mixes prompts and models as it lists them, so the first page covers your tracking broadly instead of repeating one model. Hover a response to preview the full answer, or open it in a new tab, without losing what you already selected. Select up to 10 responses per request. Each project can generate synthetic queries for 100 responses per UTC calendar month, shared by all members and models. Requests count when queued, including failed generations. Reopening a saved result or retrying its job does not use another slot. Copilot and Alexa responses cannot be used for generation. Embedded portals and response details only show saved results; generate queries from the Synthetic page in the main app.

Export synthetic ideas

Use the Export table button on either Synthetic table to download the current filtered view as CSV or Excel. Synthetic exports are separate from observed fan-out exports, so inferred suggestions never mix with model-reported searches.

By prompt

In By prompt, expand a prompt to see its saved query suggestions and how many source responses contain each one. Repeated queries count as one suggestion, ignoring differences in case and spacing. Each response counts once per suggestion. The list uses your current filters and has pagination when needed. Expanding a prompt or choosing Try again after a loading error only reads saved results; it does not make another AI request.