From Anker SOLIX to xTool — the brands above already see how ChatGPT, Gemini and Perplexity mention, cite and recommend them. Your brand is being talked about in AI right now. See it.
What Are Grounding Queries? Definition & GEO Guide | GEOly | GEO Data Platform for DTC Brands
What Are Grounding Queries? The Hidden Searches Behind AI Answers
Summary
Grounding queries are the search queries an AI model generates and runs behind the scenes to retrieve fresh web content before answering — rank for them and you can be cited; miss them and the AI writes its answer without you.
2026/07/05
6 min read
Grounding queries are the search queries an AI model generates and runs against a live search index — Bing for Microsoft Copilot, Google for Gemini — to retrieve current information before it composes an answer. They are not what the user typed; they are the machine's own reformulations, and they "ground" the response in retrievable web content instead of the model's frozen training data. That makes them the closest thing AI search has to a keyword list: surface for a grounding query and you can be cited, miss it and the answer gets written without you.
Key takeaways
Grounding queries are machine-generated searches that AI engines run behind the scenes to verify facts and fetch fresh content. They usually differ from the user's original prompt in wording, specificity, and count.
One prompt typically fans out into several grounding queries, so a single AI answer creates multiple retrieval opportunities — and multiple ways to be skipped.
Bing Webmaster Tools' AI Performance report, in public preview since February 2026, is the first major webmaster console to expose real grounding queries alongside the URLs they caused to be cited.
Retrieval is the gate in front of every downstream GEO metric: no grounding-query visibility means no mentions, no citations, no product cards.
How grounding queries work
Large language models are frozen at their training cutoff. To answer anything time-sensitive or brand-specific, they use retrieval-augmented generation (RAG), and grounding queries are the retrieval half of that loop:
A user asks, "Is GEOly better than Ahrefs for tracking AI search visibility?"
The model decomposes the prompt into its own searches: "GEOly AI platform features", "GEOly vs Ahrefs comparison", "AI visibility tracking tools 2026 reviews".
Each query runs against a search index — Copilot queries Bing, Gemini and Google AI Mode query Google, Perplexity queries its own index.
The model reads the top passages, selects a handful as AI citations, and synthesizes the answer.
Steps 2 and 3 are grounding. Notice the translation that happens along the way: the user speaks in casual comparison language, while the model searches in document language — more specific, more fact-shaped, often with a year attached. A shopper typing "best budget projector for a bright living room" may trigger grounding queries like "projector lumens needed for daylight viewing" and "high brightness projectors under $500 2026".
Google calls this expansion query fan-out in AI Mode, and AI Overviews work the same way. Gemini developers can even see the exact list: the API returns every executed search in the webSearchQueries field of its grounding metadata, per Google's grounding documentation. Microsoft uses the term grounding queries directly in its Copilot data. Same mechanism, different labels.
Query fan-out tracking: how ChatGPT expands buyer questions into web search queries, with popular searches and demand themes — Source: GEOly AI (app.geoly.ai)
Why grounding queries matter in 2026
Grounding queries reveal the intent an AI is actually trying to satisfy, which is frequently not the intent your keyword research describes. They also mark the point where GEO diverges from SEO: classic search optimizes for what humans type, while generative engine optimization optimizes for what machines search. And because the user never sees these queries, the whole competition happens inside a zero-click black box — observable only from the outputs unless a platform exposes the data.
Every metric in the AI visibility stack — mention rate, citation rate, AIGVR — sits downstream of retrieval. Content that never surfaces for grounding queries has a ceiling of zero on all of them.
Where to find grounding query data
Three practical sources exist today.
Bing Webmaster Tools. Microsoft launched the AI Performance report in public preview in February 2026. It shows the grounding queries Copilot and Bing's AI summaries generated when your pages were retrieved, which URLs got cited, and how citation activity trends over time. A June 2026 update added Intents (grounding queries classified as informational, commercial, navigational, and more), Topics (queries clustered into themes), Citation Share, and Compare. If you run a website, this is the closest thing yet to a Search Console for AI answers.
The Gemini API. Developers can run priority prompts through Gemini with Google Search grounding enabled and log the webSearchQueries returned with each response. It is a lab method rather than a console, but it shows real fan-out behavior on Google's side.
Prompt-level monitoring. Each webmaster console covers one engine. GEOly AI approaches the problem from the output side: it tracks your priority prompts across seven engines (ChatGPT, Gemini, Perplexity, Copilot, Grok, Google AI Mode, AI Overview), shows the query fan-out behind each prompt, and connects every answer to the domains that were retrieved and cited. Pairing Bing's grounding-query data with cross-engine monitoring tells you both what machines search and whether that search finds you.
Citation source analysis: source type distribution and the domains AI engines cite most — Source: GEOly AI (app.geoly.ai)
How to optimize for grounding queries
Fill data voids. Grounding queries are disproportionately fact-shaped: "pricing tiers", "vs [competitor]", "specs", "return policy". If no page answers them cleanly, the AI settles for whatever exists — often a third party's guess. Running a content gap analysis against the queries in your AI Performance report is the fastest way to find these holes.
Write retrievable passages. Retrieval happens at the passage level. Lead each section with a direct answer, keep one claim per paragraph, and use headings that mirror query phrasing — "How much does X cost" beats "Investment options".
Mark up your facts. Structured data makes prices, specs, and availability machine-extractable, which raises the odds a retrieved page actually gets used in the answer.
Work the third-party layer. Queries like "GEOly reviews" retrieve pages you don't own. Citation analysis shows which review sites and communities each engine pulls from in your category — that is your earned-media target list.
Stay crawlable. None of this matters if AI crawlers can't fetch your pages. Check robots.txt before you check anything else.
Common mistakes
The most common error is treating grounding queries like a 2019 keyword list and optimizing only head terms; machine queries skew long, specific, and multi-variant. The second is watching only Bing because it is the only console with data — each engine reformulates differently, and ChatGPT or Perplexity may search your category in ways Copilot never does. The third is expecting on-site content to win queries that structurally resolve to third-party sources, like reviews and "best of" comparisons.
FAQ
Are grounding queries the same as keywords in Search Console?
No. Search Console keywords are what humans typed. Grounding queries are what an AI generated after reading a human prompt — usually longer, more specific, and several per prompt. There is overlap, but treating them as identical hides exactly the reformulation layer you need to optimize for.
How can I see the grounding queries ChatGPT uses?
OpenAI doesn't offer a webmaster console, so there is no direct feed. You can watch the "Searching the web" status lines in the ChatGPT interface for hints, use Bing Webmaster Tools as a proxy since ChatGPT's browsing has historically drawn on Bing's index, and track prompt-level outcomes across engines with a monitoring platform like GEOly AI.
What's the difference between grounding queries and query fan-out?
Query fan-out is Google's name for the expansion step — one prompt becoming many searches in AI Mode. Grounding queries are the searches themselves. In practice the terms get used interchangeably: Bing says grounding queries, Google says fan-out queries.
Can I control which grounding queries an AI generates?
Not directly — query generation happens inside the model. You influence it indirectly: clear entity naming, consistent terminology across your site and profiles, and comprehensive coverage of your category's fact-space all raise the odds that whatever queries get generated resolve to you.
