What is a grounding query?

A grounding query is the search an AI system sends to an index so it can ground its answer in live sources. The model writes it. The user never sees it.

The word "grounding" refers to tying an answer to retrieved documents rather than to what the model remembers from training. A grounding query only exists when retrieval fires. If the model answers from memory, no grounding query runs and no page gets fetched.

The sequence runs in four steps:

1. A user types a prompt.
2. The system decides whether to search at all. Retrieval is per-query and discretionary on every major platform.
3. If it searches, the model generates one or more grounding queries and sends them to an index.
4. The index returns candidate pages. The model reads some of them and cites some of those.

The vendors document step 3 themselves. OpenAI's help centre says ChatGPT search "typically rewrites your query into one or more targeted queries" that it sends to search providers, and that saved memories may shape the rewrite. The Gemini API documentation says grounding with Google Search automatically generates one or more search queries and runs them. Google describes a "query fan-out" technique for its AI features, issuing multiple related searches across subtopics.

So a single conversational prompt can become several short, specific searches. For the full path from query to citation, see how AI search picks sources.

Most AI visibility tracking sends prompts the tool or the marketer chose. Those prompts are a proxy. The engine searches with its own rewritten queries, and nobody outside the operator observes them. That gap is covered in more detail in whether tools can see real AI prompts.

Bing is the exception. Its AI Performance report in Bing Webmaster Tools, in public preview since February 2026, lists grounding queries: in Microsoft's words, "the key phrases the AI used when retrieving content that was referenced in AI-generated answers." Coverage spans Microsoft Copilot, AI summaries in Bing and select partner integrations. You can read it in the dashboard and it costs nothing.

What to do with it:

• Verify your site in Bing Webmaster Tools and open the AI Performance report.
• Export the grounding queries by hand, since there is no API as of October 2026.
• Compare them with your prompt set and note topics the engine searches that you never test.

This matters because topical fit to the retrieval query is one of the few factors with repeated support. A 2026 critical survey of 45 GEO studies (arXiv 2607.14035) names topical relevance and context position as the most reproducible levers. Grounding queries tell you which topics the engine actually searches for.

"The grounding query is the prompt"

No. The prompt is the user's input. The grounding query is the model's search. They can share words, but OpenAI and Google both document rewriting.

"Bing's list shows every query that retrieved my page"

No. Microsoft describes the list as queries behind content that was referenced in answers, and it labels this metric a sample. Pages retrieved but not cited do not appear. See whether the data is sampled.

"Bing's grounding queries show what ChatGPT searches"

No. The report covers Microsoft surfaces and select partners. It is not a window into ChatGPT, Gemini or Google AI Overviews.

"I should write a page for every query variant"

Google's AI optimization guide says you don't need to worry about capturing every long-tail variation. Use grounding queries to find topic gaps, not to stuff keywords.

• Query fan-out: splitting one prompt into several related searches across subtopics.
• Retrieval: the answer-time step where grounding queries pull candidate pages into the model's context.
• Parametric answer: a reply from training data alone, with no grounding query. Test the difference with search on and off.
• Retrieved vs cited sources: the retrieved set is usually larger than the cited set, per OpenAI's API documentation.

Can I see grounding queries for ChatGPT or Google AI Overviews? Not as of October 2026. OpenAI and Google document that they rewrite prompts into searches, but neither shows site owners those queries. Google Search Console's generative AI report shows impressions only. Bing Webmaster Tools is the only first-party source of real grounding queries, and it covers Microsoft surfaces and select partner integrations, not ChatGPT or Google.

Does one prompt always produce one grounding query? No. OpenAI says ChatGPT rewrites a prompt into one or more targeted queries. Google describes fanning out across subtopics, and the Gemini API says it may generate one or several searches. A prompt can also produce none, because each platform decides per query whether to search at all. Then the answer comes from training data.

Is Bing's grounding query data complete? No. Microsoft states the grounding query data represents a sample of overall citation activity. Treat it as a directional list of real retrieval phrasings, not a full count. Other metrics in the same report, such as total citations, carry different caveats, so read each one on its own terms.

Do grounding queries replace a prompt set? No. A prompt set lets you test the same questions across engines over time. Grounding queries show what one family of engines actually searched. Use both: keep a stable prompt set for tracking, and use exported grounding queries to find topics your prompts miss.