How to find ChatGPT's fan-out queries

Find the queries in your browser

Use ChatGPT in your own desktop browser. The instructions below use Chrome DevTools. You do not need to install an extension or paste code into the console.

1. Open a chat that uses web search

   In ChatGPT, choose View all tools, then Search, and send your question. Wait for the answer. An existing conversation with a searched answer also works as a starting point. OpenAI documents how to select Search.

2. Copy the conversation ID

   In a saved chat URL such as chatgpt.com/c/<conversation-id>, copy the part after /c/. You will use it to identify the chat's network response, not as a search query.

3. Open Network, then reload

   Right-click the page and choose Inspect. Open Network, make sure recording is on, and reload the chat with DevTools still open. Enable Preserve log to keep requests across reloads. Select Fetch/XHR and paste the conversation ID into the filter.

4. Select the conversation response

   Look for a request for the matching conversation. Aiso's extractor uses the path /backend-api/conversation/<conversation-id>. Open the request's Response tab. The endpoint is an implementation detail and can change.

   Do not type a field name into the request filter and expect it to search the response body. Use the Network search control to search response contents, or copy the response into a local text editor.

5. Find the field and copy the query strings

   Search for search_model_queries. In the string-list format below, the queries are inside metadata.search_model_queries.queries. Also search for search_queries and inspect the q values. Keep the query text with the message it belongs to.

Chrome's Network reference covers recording, searching responses and Copy response. Inspect the response body, not just the request headers.

What the query metadata looks like

Format 1: query strings inside search_model_queries

Expand metadata, then search_model_queries, then queries. The strings in that array are the values to copy.

{
  "metadata": {
    "search_model_queries": {
      "queries": [
        "CRM Gmail integration five users",
        "HubSpot vs Pipedrive Gmail integration"
      ]
    }
  }
}

Format 2: objects inside search_queries

Read the q value from each object. Aiso's extractor also checks for a search_queries array outside a metadata object.

{
  "metadata": {
    "search_queries": [
      {
        "q": "CRM Gmail integration five users"
      },
      {
        "q": "HubSpot vs Pipedrive Gmail integration"
      }
    ]
  }
}

The query text in both examples is invented; neither is a live ChatGPT capture. The structures match formats handled by Aiso's extractor, inspected September 8, 2026. Inspect the example JSON and expected extracted values.

Save the message context, not just a list

A conversation response can contain multiple messages and alternative branches. Do not assign every query you find to the newest prompt. Keep the relevant message ID and user turn alongside the exact query, capture date and visible model. When you cannot establish the association, label it conversation-level rather than guessing.

To copy the response in Chrome, right-click the request and choose Copy → Copy response. Save it locally. Before sharing any extract, remove unrelated chat text and personal details. Do not share cookies, authorization headers or an unreviewed HAR export.

When the queries are missing

Asking ChatGPT to reconstruct the searches is not a substitute for capturing them. Unless you can match its list to exposed tool or response metadata, label it generated query suggestions, not observed fan-outs.

OpenAI describes rewriting requests into one or more searches, sometimes followed by additional searches after results arrive. There is no universal query count in that documentation. The browser response is also not a guaranteed complete record of server-side search activity.

Check what changed between the prompt and the searches

Suppose your test prompt is “Find a CRM for five people that works with Gmail.” In the illustrative JSON above, one query introduces HubSpot and Pipedrive even though the user named neither.

That gives you a specific next check: did your captured queries introduce named competitors, a comparison or an integration requirement? Compare those exact strings with the pages cited in the same answer. You may need a Gmail integration page or a factual comparison, rather than another broad “best CRM” article.

This is a way to choose a content task, not proof of ROI. A query's presence does not by itself show which result caused the final recommendation. Repeat the capture before investing in a pattern that appeared only once.

Find fan-out queries across conversations with Aiso

The manual method lets you inspect a conversation you can access. For category research, Aiso combines observed conversation data with fan-out and citation analysis. Use it to investigate the questions and search wording around your category rather than building a strategy only from prompts you invented yourself.

Keep observed fan-outs separate from generated suggestions and controlled tests. Check the available model, market and date coverage in Aiso's methodology. An extension can reduce copying work, but it cannot recover metadata that the interface does not expose.

Method notes

The opening 20× calculation uses Vercite’s reported averages from its June 2026 study of 17,806 fan-out queries. It compares a cumulative vocabulary with a per-answer average, not two measurements of a single answer.

The field paths above were checked against Aiso's ChatGPT extractor implementation. This was a code review, not a new live test across ChatGPT accounts. OpenAI does not document these internal fields as a public API.

For browser controls, use Chrome's Network documentation. For search availability, query rewriting and source links, use OpenAI's ChatGPT Search documentation. The downloadable examples contain no customer conversations.