Grounding Queries: How AI Search Engines Look Things Up

Most advice about AI search assumes the model reads your customer’s prompt and goes looking for pages that match it. It doesn’t. Before an answer engine writes a word, it often rewrites the prompt into its own short searches, sends them to a search index, and builds the answer from whatever comes back. Those machine-written searches are grounding queries, and they, not the prompt, decide which pages get a shot at being cited. Optimise for prompts alone and you miss the step where the shortlist is made.

This guide covers what they are, how they differ from fan-out and retrieval, the few places you can see them, and how to turn them into content decisions.

What a grounding query actually is

“Grounding” means tying a model’s answer to sources fetched at answer time rather than training data alone. A grounding query is the search string the model writes to do that fetching. Microsoft’s own definition, from the February 2026 launch post for AI Performance in Bing Webmaster Tools, is “the key phrases the AI used when retrieving content that was referenced in AI-generated answers.”

Two things in that definition matter. The AI writes the phrase, not the user: a rambling question typed into Copilot becomes a compressed search sent to the index. And it is tied to retrieval: a grounding query only matters to you if your page came back from it and the model used what it fetched.

Grounding vs query fan-out vs retrieval

Treat these three terms as stages of one pipeline:

Grounding is the goal: anchor the answer in fetched, checkable sources.
Query fan-out is the planning step: split one prompt into several sub-questions. Google’s Search Central documentation says AI Overviews and AI Mode “may use” a query fan-out technique, “issuing multiple related searches across subtopics and data sources.” Our query fan-out guide goes deeper on that expansion.
Retrieval is the fetch: run each search, pull back ranked pages, and extract the passages that fit the model’s context window.

The model’s searches are the output of fan-out and the input to retrieval. Some vendor posts frame grounding as a separate “verification” layer; no engine documents it that way. Google, OpenAI and Anthropic all describe one loop: decide to search, write queries, read results, maybe search again, then answer with citations. Fan-out just means the loop produced more than one query.

When an AI engine decides to ground

Grounding is not automatic. Every major API makes search a choice the model takes per prompt. Google’s Gemini documentation says the model “analyzes the prompt and determines if a Google Search can improve the answer.” OpenAI’s web search guide says the model “can choose to search the web or not based on the content of the input prompt.” Anthropic’s documentation for Claude is the most explicit about triggers. Claude searches for:

Recent events, news or announcements
Current prices, rates, scores or statistics
Specific organisations, people or products whose details may have changed
Explicit requests to look something up

It answers without searching for stable facts, maths, creative work and text already in the conversation. If your category’s questions can be answered from training data (“what is a CRM?”), there may be no grounding step to win. If they involve current prices, comparisons, specs, locations or anything dated, the engine is far more likely to go looking, and a well-structured page can win a citation this week rather than at the next training run.

What the model’s searches look like in practice

The best public dataset comes from Seer Interactive. In November 2025, R&D lead Nick Haigler ran 501 prompts through the Gemini 3 API with grounding forced on every prompt and logged the searches the model returned. His findings:

An average of 10.7 searches per prompt, ranging from 3 to 28, up from 6.01 in his Gemini 2.5 runs.
An average length of 6.7 words per query.
95% of the queries had zero global monthly search volume.
Only 1% overlapped across the whole dataset.
21.3% contained a year, and 26.4% contained a brand name.

The study’s own caveats: grounding was forced, so it says nothing about how often Gemini searches unprompted, and it covers one model. These searches are longer than human keywords, rarely show up in a keyword tool, rarely repeat and often carry a year. A year in the search is a request for fresh pages, which puts an undated or stale page at a disadvantage. Our piece on content freshness and AI search covers what recency actually changes.

Where you can actually see grounding queries

You cannot pull these searches out of ChatGPT’s consumer app or Google’s AI Mode; Google folds AI feature traffic into the “Web” search type in Search Console, with no breakdown of the searches its models ran. Four places do expose real search strings:

Bing Webmaster Tools, AI Performance. Launched in public preview on 10 February 2026, it covers citations across Microsoft Copilot, AI-generated summaries in Bing and select partner integrations, and lists sampled grounding queries that led to your pages being cited. In March 2026 Microsoft rolled out a mapping view that links each query to the pages cited for it, per coverage from Search Engine Roundtable and PPC Land. It shows citations, not clicks.
The Gemini API. With the Google Search tool enabled, responses include groundingMetadata with a webSearchQueries array, the sources used (groundingChunks), and groundingSupports linking each sentence to its sources.
The OpenAI Responses API. A web_search_call item records the action taken. OpenAI says a search action “usually (but not always)” includes the queries run, and a sources field lists every URL consulted, which is often more than the number cited.
The Claude API. Each search appears as a server_tool_use block with the query as input, followed by results and citations carrying up to 150 characters of cited text.

Only the Bing report reflects real user traffic. The APIs show how a related model behaves on prompts you choose: a proxy, not a log of what consumer apps did.

A repeatable workflow for capturing them

The process we use for a client category:

Start from real questions. Don’t invent prompts. Pull the questions people actually ask in the space. SEO Rocket’s AI Visibility feature, built on Ahrefs Brand Radar data, shows how many AI responses across ChatGPT, Google AI Overviews, Gemini and Perplexity cite your site, along with the real questions where you are cited, or the questions people ask in your space if you are not cited yet. That list is your prompt set. Our prompt tracking guide covers how to build and maintain one.
Run each prompt through two or three APIs with search enabled and log every query, every source URL and every cited URL. Run each prompt more than once; outputs vary per run.
Pull the Bing AI Performance export for the same period and merge it with the API logs.
Normalise and cluster. Strip years and brand names, then group by topic. Hundreds of unique strings collapse into a handful of themes.
Map each cluster to one URL on your site, or mark it as a gap.

Budget for cost and noise. Google bills Gemini 3 and newer per search the model executes, so one prompt can trigger several billable searches. Anthropic notes single-fact questions typically use one to three searches, comparative research ten or more.

How to read a grounding query

Decode the strings rather than chase them. Almost every search the model writes is some combination of four parts: an entity (a brand, product or category), an attribute (price, limits, integrations, side effects), a qualifier (a year, a location, a team size, a budget), and an implied format (comparison, list, definition, steps).

An illustrative example, not real data: a prompt asking which bookkeeping software suits a ten-person accounting firm might produce searches like “bookkeeping software small accounting firm 2026”, “[brand] multi-user pricing” and “[brand] vs [brand] accountant features”. The decision rule: for each entity-attribute pair that recurs, does one passage on your site state that attribute, for that entity, with the qualifier, in a sentence that survives being lifted out of the page? If the answer lives in a table image, a PDF or a sentence that leans on the paragraph above, the engine may fetch your page and still quote someone else.

How to get your pages retrieved and used

Winning a grounding query is classic SEO, then passage work:

Be eligible. Google says a page must be indexed and eligible to be shown with a snippet to appear as a supporting link in AI Overviews or AI Mode, and that no special optimisation beyond that is required. Bing has its own index, so verify the site in Bing Webmaster Tools too.
Let the fetchers in. An old robots.txt rule can block the bots answer engines use. Check with the free AI Crawler Robots.txt Checker; SEO Rocket’s site audit runs a real crawler and flags AI-crawler blocks as part of the same pass.
Rank for the short version. The model’s searches return ranked results, so a page invisible for the underlying keyword rarely gets fetched. SEO Rocket’s keyword research on real Ahrefs data covers that head-term side.
Write self-contained passages. State the entity, the attribute, the number and the condition in one sentence, near the top of its section. Our guide to content chunking for LLMs covers structure in detail.
Date your facts. When roughly a fifth of queries in Seer’s sample carried a year, a visible “updated” date and current figures stop being cosmetic.

Don’t spin up a thin page per captured string. They barely repeat, and mass-producing near-duplicate pages to match them is what Google’s spam policies call scaled content abuse. Cluster first, then build or improve one strong page per cluster. That is the same consolidation logic behind a playbook proven across 1,000,000+ ranking pages.

What you can’t measure, and how not to fool yourself

Microsoft says its query data is a sample of citation activity, not a complete record. Consumer apps add personalisation, location and conversation history that API runs cannot replicate. The same prompt can produce different searches on consecutive runs. And retrieved is not cited: OpenAI notes that sources consulted often outnumber citations shown.

Grounding doesn’t guarantee accuracy either. A model can fetch the right page and still misstate it, which is why the AI hallucination problem matters even for brands that are cited. Report grounding work as directional: which themes you show up for, which pages get cited, and whether that is moving. For the citation side of the picture, see our sibling guide on LLM citations, and for a measurement setup that holds up month to month, AI visibility tracking.

Frequently asked questions

Are grounding queries the same as the prompts users type?

No. The user writes the prompt; the AI system writes the search. Microsoft describes them as the phrases the AI used when retrieving content, and they are usually shorter, more specific and often include qualifiers such as a year.

Can I see the searches Google AI Mode runs for my keywords?

Not directly. Search Console reports AI feature traffic inside the overall “Web” search type. The closest proxy is the Gemini API, which returns the searches it ran in its grounding metadata.

Should I add grounding queries to my keyword list?

Use them as topic and attribute signals, not as keywords to target one by one. In Seer Interactive’s sample, 95% had zero search volume and almost none repeated, so cluster them and answer the cluster on one strong page.

Does Bing’s report cover ChatGPT?

Microsoft scopes AI Performance to Copilot, AI summaries in Bing and unnamed partner integrations. Don’t read it as a ChatGPT report.

Where to start this week

Verify your site in Bing Webmaster Tools and open AI Performance, the only place grounding data from live traffic is published. Pull twenty real questions from your space, run them through one search-enabled API, and cluster what comes back. Expect searches about attributes your pages never state plainly; fix those passages first, one strong page per cluster. Then track whether citations follow, with AI Visibility citation counts alongside your Bing data, and report the trend rather than any single run.