Grounding Queries vs Prompts: How to Validate Your Prompt Set

Your prompt set is a list of questions you send to AI engines to see whether they name you. It's useful. It's also not the input that decides which pages get fetched.

Engines don't search for the exact words a user types. OpenAI's help page on ChatGPT search says it "typically rewrites your query into one or more targeted queries" that it sends to search providers. Google may use a "query fan-out" technique, issuing multiple related searches across subtopics. The Gemini API documentation says the model "automatically generates one or multiple search queries" and runs them. This is query rewriting, and it sits between your prompt and the retrieval step.

That creates a gap with two consequences:

• One prompt can become several searches. A question about pricing might also trigger searches about alternatives, reviews or integrations.
• Context changes the rewrite. OpenAI notes that saved memories can shape how ChatGPT rewrites a search query, so logged-in answers vary by user.

So any prompt-based tracker, including the free checker on this site, measures answers to prompts someone chose. It doesn't observe the searches the engine ran. For more on this, see whether AI searches for the exact question a user types. The practical move is to check your prompts against the one source of real retrieval queries that exists.

As of October 2026, Bing Webmaster Tools is the only place a site owner can see grounding queries. Its AI Performance report (public preview since February 2026, expanded in June 2026) lists "the key phrases the AI used when retrieving content that was referenced in AI-generated answers." It covers Microsoft Copilot, AI summaries in Bing and select partner integrations.

Know its limits before you build on it:

• No API. As of our evidence date, the Bing Webmaster API has no grounding query endpoint. You export from the interface.
• It's a sample. Microsoft states that the grounding query data "represents a sample of overall citation activity." Treat counts as directional.
• Cited pages only. It shows queries that retrieved content which then got referenced. It won't show searches where your page was fetched and dropped, or never fetched.
• Microsoft surfaces only. ChatGPT, Perplexity, Claude and Google don't expose their rewritten queries to publishers.

That last point matters. Copilot's queries are evidence about how engines rewrite in your category, not a transcript of every engine. Our guide to tracking Copilot mentions and grounding queries covers setup and verification.

You need two lists: your tracked prompts and a grounding query export. A spreadsheet is enough.

1. Export grounding queries. In Bing Webmaster Tools, open the AI Performance report, set the widest date range available, and export or copy the grounding query list with its intent and topic labels.
2. Export your prompt set. Pull every prompt you track, with the platform and country each one runs on.
3. Normalise both lists. Lowercase them, strip brand names and filler words, and merge obvious duplicates. You're comparing meaning, not wording.
4. Assign a topic to every row. Use short labels such as "pricing", "alternatives", "integration with X", "setup", "compliance". Use the same labels for both lists. Bing's own Topics and Intents columns are a good starting vocabulary.
5. Build a coverage table. For each topic, record how many grounding queries and how many prompts it has.
6. Sort into three buckets. Covered (both lists), missing (grounding queries, no prompts), and unvalidated (prompts, no grounding queries).
7. Write new prompts for missing topics. Phrase them the way a buyer would ask, not as a copy of the grounding query. The grounding query is the engine's rewrite. Your prompt should be the human question that could plausibly lead to it.
8. Review the unvalidated bucket. Keep prompts that matter commercially even without evidence. Cut or rewrite ones that test nothing real.
9. Record the date and export. Save both lists so the next comparison has a baseline.

Say you sell payroll software (a hypothetical). Your prompts all ask "best payroll software for small business." The grounding queries show searches about contractor payments and payroll tax deadlines. Those are two missing topics, and your tracking says nothing about them.

Run through this before you act on the coverage table:

• Your site is verified in Bing Webmaster Tools and the AI Performance report shows data.
• The date range is long enough to include at least one full cycle of your sales season.
• You compared topics, not exact strings.
• The same person or rule assigned topics to both lists.
• You noted that grounding queries are sampled and cover Microsoft surfaces only.
• Your prompt set records country and language, since rewrites differ by market (see prompt sets for multiple countries and languages).
• New prompts read like questions a buyer would type, not search strings.
• You kept brand and non-brand prompts separate.
• You saved the export and the date so the next run is comparable.

If you can't tick the first item, you have no grounding queries. Fall back to building prompts from buyer research, as covered in how to build a prompt set for AI visibility.

Good

• Most grounding query topics appear in your prompt set.
• The topics with the most grounding queries are also the ones you track most heavily.
• Grounding queries contain intents you expected, such as comparisons, pricing and setup.
• Missing topics are few and specific, so you can add prompts for them this week.

Bad

• A large share of grounding queries fall on topics you never test.
• Your prompt set is all broad "best X" questions while grounding queries are narrow and specific.
• Many of your prompts land in the unvalidated bucket, and they are the ones you report to leadership.
• The report shows almost no grounding queries. That usually means few citations on Microsoft surfaces, not good coverage. Check Bing index status and crawler access before you read anything into it.

A bad result is information, not failure. It tells you your visibility numbers describe a narrower slice of demand than you assumed. The fix is to change what you measure.

Don't over-read a good result either. High overlap means your prompts sit near real retrieval topics. It doesn't predict citations. The best available review of GEO studies (arXiv, July 2026) found topical relevance among the most reproducible factors in already-retrieved content, but no technique with a stable causal effect on discoverability.

Repeat the full comparison once a quarter. That's often enough to catch drift and rare enough that the sampled data has time to accumulate.

Run it out of cycle when something changes the questions buyers ask:

• A product launch or a new feature category.
• A pricing or packaging change.
• Entering a new country or language.
• A major content release on a topic you didn't cover before.
• A large shift in the report itself, such as new intents or topics appearing.

Between full runs, a monthly glance at the top grounding queries takes ten minutes. If a new topic appears near the top, add a prompt for it then rather than waiting for the quarter.

Be plain about the edges. Grounding queries come from one company's surfaces. They don't show how ChatGPT, Perplexity, Claude, Gemini or Google AI Overviews rewrite the same question. They also don't show the prompts people typed: Bing exposes the engine's queries, not the user's words. For the wider picture, read whether AI visibility tools can see the prompts people actually type.

The comparison also can't fix access problems. If a crawler can't reach your pages, or your content only appears after JavaScript runs, no prompt set will show you being cited. Check access first, then tune what you measure.

What you control: the topics you track, the phrasing of your prompts, and how often you check them against real retrieval data. What you don't: how each engine rewrites, what it retrieves, and what it chooses to cite.