01Notes · 25 August 2026
Grounding Queries: What Bing Exposed
Quick answer
A grounding query is the search an AI system runs against an index before it writes an answer. It is frequently not the question the user typed: the system reformulates, and often runs several.
Bing has published these since February 2026. It is the closest thing available to watching an AI system show its working, and in our experience almost nobody has opened it.
Key takeaways
- You compete to be retrieved by the grounding query, not by the question the user typed.
- The two are often different. A conversational question becomes several reformulated searches.
- Bing publishes them free in AI Performance, public preview since February 2026.
- No other platform we know of exposes this. Everywhere else it has to be inferred from cited sources.
- The useful move is comparison: grounding queries against the pages that were actually cited.
- It adds a layer to keyword work rather than replacing it. Buyer language on the page, machine phrasing checked.
A grounding query, defined
When an assistant answers a question, it rarely searches for the question verbatim. It interprets what is being asked, decides what information it needs, and runs one or more searches against an index to gather source material. Those searches are the grounding queries.
The answer is then written from what came back. So the retrieval step, not the user’s phrasing — decides which pages are even eligible to be used.
The gap between what people ask and what machines search
This is where it gets practically useful, because the two are often not the same at all.
| What the user types | What the system is likely to search for | What that means for your page |
|---|---|---|
| "which of these works with our existing stack and what does it cost" | Two or three separate searches: integrations, pricing, comparison | One page answering all three beats three answering one each |
| "is [vendor] any good" | Reviews, complaints, alternatives | Your own page is not the source being sought |
| "cheapest way to do X" | Pricing comparisons, free alternatives, X cost | Published pricing becomes a retrieval advantage |
| "how do I fix [problem]" | Problem cause, problem solution, sometimes the error text | Diagnostic content outperforms solution-only content |
The middle column is where the competition actually happens. Optimising only for the left column is optimising for a query the machine may never run.
How to read the data
In Bing Webmaster Tools, AI Performance shows citations, the pages cited, and the grounding queries behind them. Three comparisons are worth making:
- Grounding query against cited page. Is the page that got used the one you would have chosen? Often it is not, and that tells you which page is actually carrying the topic.
- Grounding query against your content. Are there queries where nothing of yours is eligible? Those are gaps expressed in the machine’s own words rather than a keyword tool’s.
- Grounding query against user language. Where they diverge sharply, your page may need to serve both phrasings — one for the reader, one for retrieval.
None of this requires a tool subscription. The account probably already exists in your organisation, verified once and never opened. We made the same point about Copilot.
The limits of this number
Three honest limits, because this is a Bing-specific window rather than a general one.
It covers Bing-grounded surfaces. ChatGPT, Gemini and Perplexity run their own retrieval and none of them publish this. You cannot generalise from it.
It is a preview. Coverage and definitions may change, so treat trend readings across long periods with caution.
It shows retrieval, not selection. Being retrieved by a grounding query is necessary, not sufficient: the system still chooses which retrieved sources to use. That second step remains invisible everywhere.
What this means for you
- Open AI Performance in Bing Webmaster Tools this week. It is free and the account almost certainly exists already.
- Compare grounding queries against your cited pages. The mismatch is where the work is.
- Do not generalise it to other platforms. This is a Bing window, not a view of AI search.
02Common questions
Questions this raises
What is a grounding query?
The query an AI system runs against a search index to gather source material before writing an answer. It is often not the question the user typed: the system reformulates, and sometimes runs several.
Why does that matter?
Because you are competing to be retrieved by the grounding query, not by the user's question. Optimising for the phrasing a person types can miss the phrasing the machine searches.
Where can I see grounding queries?
Bing Webmaster Tools has published them since February 2026 in the AI Performance section, in public preview. It is the only place we know of where they are shown directly.
Do other platforms publish this?
Not that we have found. For every other assistant, grounding behaviour has to be inferred from the sources that appear, which is much weaker evidence.
What do I do with the data?
Compare the grounding queries against the pages that got cited. Where a query is being served by a weak page of yours, or a competitor's page, that is a specific and actionable gap.
Does this change my keyword research?
It adds a layer rather than replacing one. You still target buyer language on the page. You also check whether the reformulated machine query is something your page answers.
Sources
Every figure on this page comes from one of the following. Where a claim is ours rather than a source’s, it carries an evidence label instead.
- Bing Webmaster Tools — AI Performance reporting, public preview announced February 2026. Bing — Elevating the Role of Grounding on the AI Web, February 2026. Primary
- Rankability — State of AI Search, 1,645 citation observations across 31 topics, 2026. Source for the 24.1% cross-platform overlap and the 73.5% ChatGPT exclusive citation rate. Independent study
- Searlixa sampling across client engagements, labelled Observed with sample and date. Our own
Ask an assistant
Hand this page to an assistant and ask it to summarise or challenge the argument. If it gets something wrong about us, that is itself a useful finding.
→Next step