Grounding Queries
- Grounding queries are written by the AI system, not typed by the buyer.
- They shape which pages are retrieved, and so which can be cited.
- Bing reports a sample; several provider APIs return the searches for your own test prompts, usually but not always completely; Google names no fan-out report for site owners.
- Report grounding-query figures with counting unit, access path and completeness: disclosed queries are an observable slice of the search.
Consensus definition
A grounding query is a search the AI system writes and runs to fetch evidence for an answer. The term reached site owners on 10 February 2026, when Bing Webmaster Tools added an AI Performance report that defines grounding queries as "the key phrases the AI used when retrieving content that was referenced in AI-generated answers"1; the report covers Microsoft Copilot, AI-generated summaries in Bing and selected partner integrations. The mechanism is older than the term: in February 2023 Microsoft described how its Prometheus model generates a set of internal queries through the Bing Orchestrator and grounds the answer on the results2. Other providers document the same step under other names; the table below sets out what each discloses. A grounding query is the unit of retrieval, not the unit of demand: one prompt can produce several of them, and none has to repeat the prompt's wording.
rhinegold operator refinement
Rhinegold's refinement: a page is retrieved for the query the system wrote, so retrieval is audited along the chain from prompt to grounding queries to sources and answer — provider by provider, with the disclosure limit stated next to every number. Grounding queries add a level of analysis; they do not replace the prompt. For most site owners, Bing's report is the only first-party record we know of that links the search phrases behind AI answers to the site's cited pages, and it lists a phrase only when content retrieved for it was cited1; Google's guidance for site owners counts AI feature traffic in Search Console but names no report of the fan-out queries3. Everything else has to be generated: run a fixed prompt set through the APIs that disclose their searches and label the result as a measurement of the API path, not of what buyers see. Read Bing's list as a sample of retrieval language — useful for vocabulary, intent mix and passage gaps — and never as search volume.
| Provider and access path | What is disclosed | Completeness |
|---|---|---|
| Bing Webmaster Tools, AI Performance report | Key phrases behind content that was cited, with citations and citation share per phrase; grouped by intent and topic since 16 June 202614 | A sample of overall citation activity; sampling method not published1 |
| Gemini API, grounding with Google Search | Search queries in the webSearchQueries field of the grounding metadata5 | The field description does not say whether every search is listed5 |
| OpenAI API, web search tool | Search actions in a web_search_call item; the sources list is often longer than the visible citations6 | Queries usually, but not always, included6 |
| Anthropic API, web search tool | Each query in a server_tool_use block; simple factual questions typically 1 to 3 searches, comparative or multi-entity research can use 10 or more7 | Returned in full by default; searches run from inside code execution can be dropped with response_inclusion set to "excluded"7 |
| Google AI Mode and AI Overviews | Query fan-out described: the question is broken into subtopics and many searches are issued on the user's behalf83 | No report of fan-out queries for site owners3 |
Operational use
A workable sequence has three steps. First, export the grounding queries from Bing Webmaster Tools per period and group them with the intent and topic labels Microsoft provides4; compare their wording with the pages cited for them. Second, run a fixed set of buyer prompts through the APIs that return their searches — Gemini5, OpenAI6, Anthropic7 — and record per prompt how many grounding queries were issued and, where the response lists sources, whether the own domain was among them. Third, check each recurring grounding query against the site: is there a self-contained passage that answers that query in its own terms (passage-level retrieval)? Gaps become content briefs; matched queries become the panel for re-measurement. Run each test prompt several times and store model, date and access path with every observation, because API and interface runs of the same provider can differ (API vs UI measurement).
Measurement boundary
The grounding queries behind a real buyer's prompt in a consumer chat interface are not exposed to the site owner; what can be observed is a provider's sample or one's own test runs. Microsoft describes Bing's grounding queries as a sample without stating how it is drawn1, so totals, shares of all retrieval events and the prompts that triggered a query cannot be reconstructed from them. Bing's Average Cited Pages, a daily site-level average, "does not indicate ranking, authority, or the role of any page within an individual answer"1, and Microsoft calls Citation Share an observational metric, not a ranking system; both features roll out in preview, and citation patterns also shift with models, freshness signals and partner refresh cycles4. API runs also show how a system decomposes a prompt the tester chose, not what buyers ask, and they need not match the interface: one practitioner comparison of ChatGPT (GPT 5.6; 555 prompts, 47 runs each) reports at most 7 fan-out searches through the API against up to 20 in the interface; the author states that "the result has been reproduced many times" without naming the reproductions, and does not describe how the fan-out counts were read9. Disclosed grounding queries show an observable slice of the search, not all of it. Report every grounding-query figure with its counting unit, access path and completeness: distinct phrases and executed searches are different counts, and a list can merge repeats. Bing shows a sample, from which no estimate of total volume follows1; OpenAI includes the queries usually but not always6; Anthropic returns them by default7; the Gemini field description does not say whether every search is listed5. Where disclosure is incomplete, a count is a lower bound for that unit and path; shares such as the intent mix are sample shares.
Distinct from
From Grounding: grounding is the mechanism by which a model retrieves sources and conditions its answer on them; a grounding query is the single search string inside that mechanism, and the part of it that can be read off. From Query Fan-Out: fan-out is the decomposition of one question into many searches; grounding queries are the individual searches it produces. An answer that triggered a search rests on at least one, but not every answer triggers a search7. From the discovery prompt: the prompt is what a buyer or a tester types; the grounding query is what the system writes from it. From a classic keyword: a keyword is typed by a person and reported with demand; a grounding query is machine-written and, where reported at all, comes as a sample without search or prompt volume. From Citation Rate: citation is the outcome; the grounding query is what led to retrieval. Through the APIs retrieval is wider than citation — OpenAI's sources list is often longer than the visible citations6 — while Bing lists a phrase only when content retrieved for it was cited1. From Retrieval-Augmented Generation: RAG is the general architecture; grounding queries are its web-search input in commercial AI products.
Common mistakes
- Reading Bing's grounding queries as search demand or prompt volume — they are a sample of retrieval phrases behind citations, without search or prompt volume.
- Optimising a page for the prompt wording used in tracking, while the system retrieves for differently worded grounding queries.
- Generalising one provider's disclosure to all — Bing reports a sample, OpenAI returns queries usually but not always, and Google names no fan-out query report for site owners.
- Comparing grounding-query or fan-out counts across API and interface runs without stating the access path.
- Treating Citation Share or Average Cited Pages as a rank — Microsoft calls Citation Share observational and says Average Cited Pages does not indicate ranking.
Where consensus is missing
There is no shared vocabulary: Microsoft says grounding queries, Google says query fan-out and webSearchQueries, OpenAI search actions, Anthropic server tool use. There is no published sampling method for Bing's report and no standard for counting fan-out per prompt. We have not found a peer-reviewed study that treats grounding queries as a measurement unit for AI visibility.
Sources & deeper reading
- 1Bing Webmaster Blog — "Introducing AI Performance in Bing Webmaster Tools (Public Preview)" (10 February 2026; grounding queries defined, sample of citation activity, Copilot, Bing summaries and partners)
- 4Bing Search Blog — "New AI Visibility Insights in Bing Webmaster Tools: Intents, Topics, Citation Share, Compare" (16 June 2026; Citation Share as observational metric)
- 5Google AI for Developers — Gemini API reference, generate-content (GroundingMetadata with webSearchQueries; last updated 13 October 2026)
- 6OpenAI API docs — Web search tool guide (web_search_call actions usually include the queries; sources list vs citations)
- 7Claude Platform docs — Web search tool (server_tool_use with input.query, returned in full by default; response_inclusion; several searches per request; answers without searching on stable knowledge)
- 3Google Search Central — "AI features and your website" (query fan-out in AI Overviews and AI Mode; AI feature traffic in the "Web" search type; last updated 10 December 2025)
- 9Malte Landwehr — "API vs UI: How to track LLM visibility?" (LinkedIn, 29 August 2026; ChatGPT, GPT 5.6; 555 prompts, 47 runs each; fan-out up to 7 via API, up to 20 in the interface; single practitioner study)
- 2Jordi Ribas, Microsoft Bing Blog — "Building the New Bing" (21 February 2023; Prometheus generates internal queries through the Bing Orchestrator; grounding)
