KnitKnot

# Inspect ChatGPT retrieval queries

See the search queries ChatGPT exposed during a benchmark, trace each one to a tracked question, and avoid reading query counts as search demand.

## At a glance

  1. 1. **Open Query Fan Out**: Look at the topics and content needs behind captured ChatGPT searches.
  2. 2. **Read a query group**: Open a row to see the observed queries and the tracked questions that led to them.
  3. 3. **Check the answer**: Open an observed query to read its captured answer before deciding what content is missing.

If **Intelligence → Query Fan Out** appears in your workspace, open it in the [console](https://app.knitknot.ai/intelligence/query-fan-out). It shows retrieval queries exposed by ChatGPT during completed benchmarks. A retrieval query is the wording ChatGPT sent while looking for material to answer a tracked buyer question. This view does not include searches inferred by KnitKnot or retrieval from other AI engines.

The page uses the last two completed full benchmarks. Older observations leave this view when they fall outside that window.

## Find what ChatGPT looked for

  1. 1. Read **Search trigger rate** as the share of eligible answers that exposed a search. The line under it shows the searching and eligible answer counts.
  2. 2. Search for a phrase in the captured queries, or filter the table by **Topic** and **Content need**. Each row groups one topic with one kind of content need.
  3. 3. Check **Observed query** for a recent example. **Prompts** counts tracked questions that expanded into this group; **Answers** counts executions that searched; **Queries** counts observed retrieval evidence.
  4. 4. Open the row to see the group's query receipts. Search within that list or filter by competitor, capability, or buying-role lineage. Select a tracked question to open its question record, or select a query row to open the captured answer.

For example, a buyer question about whether two products support the same integration might lead ChatGPT to search for each product's integration page. Read the recorded query and answer before deciding that your own page is missing. A query can show what ChatGPT tried to retrieve; it does not prove which page influenced the answer.

## Read the counts correctly

**Observed queries** are not Google search volume, buyer demand, or a count of people asking the question. They are searches captured while your benchmark questions ran. Prompt-verbatim echoes are excluded because they add no wording beyond the submitted question.

A group detail page can show up to 200 query receipts. When the total is larger, its header states how many are shown. If the main page says it is showing the highest-reach groups, use search or a scoped link to narrow the view. Its summary still describes the full qualifying result.

The comparison at the bottom appears only when two benchmarks have captured comparable retrieval data. One captured benchmark gives you a snapshot, not a trend. A blank table can mean there is no completed run, no eligible ChatGPT capture, no exposed retrieval query, only prompt echoes, or no group matching your filters. Read the empty-state message before drawing a conclusion.

For the questions sent to the engines, see [Manage benchmark questions](/docs/manage-questions/). To inspect a cited page or an incorrect statement in the final answer, use [Investigate a claim in an AI answer](/docs/investigate-a-claim/).

Docs navigation
Docs › Guides

Inspect ChatGPT retrieval queries

See the search queries ChatGPT exposed during a benchmark, trace each one to a tracked question, and avoid reading query counts as search demand.

Updated

At a glance

  1. Open Query Fan Out

    Look at the topics and content needs behind captured ChatGPT searches.

  2. Read a query group

    Open a row to see the observed queries and the tracked questions that led to them.

  3. Check the answer

    Open an observed query to read its captured answer before deciding what content is missing.

If Intelligence → Query Fan Out appears in your workspace, open it in the console. It shows retrieval queries exposed by ChatGPT during completed benchmarks. A retrieval query is the wording ChatGPT sent while looking for material to answer a tracked buyer question. This view does not include searches inferred by KnitKnot or retrieval from other AI engines.

The page uses the last two completed full benchmarks. Older observations leave this view when they fall outside that window.

Find what ChatGPT looked for

  1. Read Search trigger rate as the share of eligible answers that exposed a search. The line under it shows the searching and eligible answer counts.
  2. Search for a phrase in the captured queries, or filter the table by Topic and Content need. Each row groups one topic with one kind of content need.
  3. Check Observed query for a recent example. Prompts counts tracked questions that expanded into this group; Answers counts executions that searched; Queries counts observed retrieval evidence.
  4. Open the row to see the group’s query receipts. Search within that list or filter by competitor, capability, or buying-role lineage. Select a tracked question to open its question record, or select a query row to open the captured answer.

For example, a buyer question about whether two products support the same integration might lead ChatGPT to search for each product’s integration page. Read the recorded query and answer before deciding that your own page is missing. A query can show what ChatGPT tried to retrieve; it does not prove which page influenced the answer.

Read the counts correctly

Observed queries are not Google search volume, buyer demand, or a count of people asking the question. They are searches captured while your benchmark questions ran. Prompt-verbatim echoes are excluded because they add no wording beyond the submitted question.

A group detail page can show up to 200 query receipts. When the total is larger, its header states how many are shown. If the main page says it is showing the highest-reach groups, use search or a scoped link to narrow the view. Its summary still describes the full qualifying result.

The comparison at the bottom appears only when two benchmarks have captured comparable retrieval data. One captured benchmark gives you a snapshot, not a trend. A blank table can mean there is no completed run, no eligible ChatGPT capture, no exposed retrieval query, only prompt echoes, or no group matching your filters. Read the empty-state message before drawing a conclusion.

For the questions sent to the engines, see Manage benchmark questions. To inspect a cited page or an incorrect statement in the final answer, use Investigate a claim in an AI answer.