Introducing Fact Check: see what AI says about your brandIntroducing Fact Check: see exactly what AI says about your brand

Grounding Queries

Grounding queries are the search queries an AI model writes and runs to find sources for an answer. They decide which pages it reads and can cite.

Updated October 15, 2026

Definition

Grounding queries are the search queries an AI model generates and runs to collect evidence for an answer. The user types a prompt; the model rewrites it into one or more search-engine queries, reads the results, and grounds its response in what comes back. Users rarely see these queries, but they decide which pages the model reads and is able to cite.

OpenAI describes the process in its help center. When ChatGPT uses a search partner, it 'typically rewrites your query into one or more targeted queries'. Its example: a question about drugs that target CCR8 for cancer becomes the search 'CCR8 immunotherapy drug development 2025', and after reading those results ChatGPT may send a more specific follow-up such as 'CHS-114 conference 2025'. ChatGPT may also add an approximate location from the user's IP address ('top restaurants San Francisco') or details from saved memories ('good vegan restaurants San Francisco') (OpenAI).

Developer APIs return the queries directly. Google's Gemini API includes a google_search_call step listing the queries the model executed; its documentation shows 'Who won the euro 2024?' turning into the search 'UEFA Euro 2024 winner' (Google AI for Developers). OpenAI's Responses API records each web_search_call with the queries it ran.

Grounding queries overlap with query fan-out. Fan-out is the specific pattern of splitting one prompt into several parallel sub-queries, the name Google uses for how AI Mode works. Grounding queries is the broader term: it also covers a single rewritten query and the sequential follow-ups a model writes after reading its first results.

The queries look different from what users type into Google. In an October 2025 study of more than 8,500 ChatGPT prompts across nine industries, the agency Nectiv found that 31% of prompts triggered at least one search (a grounding rate of 31% for that sample), those prompts averaged 2.17 searches, and the queries averaged 5.48 words. 'Reviews', the current year, 'features' and 'comparison' were among the most frequent terms. Models also use search operators. Since August 8, 2026, about one in six ChatGPT fan-out queries has been a site: query aimed at a single domain, up from about 0.4% before, according to our site: operator tracking.

For GEO, grounding queries are the closest thing AI search has to a keyword list. A model never reads your site 'for the prompt'; it reads whatever ranks for the queries it wrote. Promptwatch extracts these queries from tracked responses in ChatGPT, Perplexity, Microsoft Copilot and other engines that expose them, and our ChatGPT fan-out data shows how they are distributed. A practical routine:

Start with prompts where your visibility is low and list the grounding queries behind them.
Run each query in the search engine the AI system uses and note whether your page appears.
Where it does not, create or rework a page that targets that query, including the qualifiers models add, such as the year, 'reviews' or 'vs'.
Re-check after the next crawl, since the queries for a prompt change between runs and only recurring patterns are worth chasing.

The queries vary from run to run even for identical prompts, so single samples mislead. Patterns across many runs, such as recurring modifiers, competitor names or review-site targets, are the useful signal.

Examples of Grounding Queries

A user asks ChatGPT for the latest on cancer drugs that target CCR8. ChatGPT first searches 'CCR8 immunotherapy drug development 2025', then follows up with 'CHS-114 conference 2025' after reading the first results.
A developer calls the Gemini API with Google Search enabled for 'Who won the euro 2024?' and the response shows a single executed query, 'UEFA Euro 2024 winner', alongside the cited sources.
A B2B team sees ChatGPT search 'best expense management software reviews 2026' for its main category prompt. Its comparison page is titled 'Top expense tools' and does not rank for that query, so the team retitles and refreshes it.
A plumbing company appears in ChatGPT answers only when the grounding query includes its city, which ChatGPT adds from the user's approximate location.

Terms related to Grounding Queries

Query Fan-Out

Query fan-out is how AI search splits one question into several sub-queries, runs them at once and merges the results, so pages compete on each sub-query.

AI Grounding

AI grounding ties a model's answer to verifiable sources, such as live search results or supplied documents, so each claim can be traced and cited.

Grounding Rate

Grounding rate is the share of AI answers built from live retrieved sources rather than the model's memory. It shows which prompts content work can change.

GEO

site: Operator Queries

site: operator queries limit search results to one domain, like 'site:example.com pricing'. AI engines now run them in fan-out to read specific sites.

RAG (Retrieval-Augmented Generation)

RAG is how AI answer engines like ChatGPT search, Perplexity and AI Mode fetch web pages at question time and cite them, deciding which brands appear.

ChatGPT Search

ChatGPT Search is the web search built into ChatGPT: it rewrites prompts into search queries, reads results and answers with inline citations to sources.

GEO

Prompt Research

Prompt research is finding and prioritizing the questions buyers ask AI assistants, so you know which prompts to track and create content for.

GEO

Latent Intent

Latent intent is the set of unstated needs behind a search or prompt. AI search systems infer it and cover it through query fan-out and follow-up suggestions.

GEO

Long-tail Keywords

Long-tail keywords are specific, lower-volume search phrases with clear intent. In AI search they become detailed prompts, where topics beat exact phrases.

SEO

Frequently Asked Questions about Grounding Queries

Learn about AI visibility monitoring and how Promptwatch helps your brand succeed in AI search.

Grounding queries are the searches an AI model runs behind the scenes to find sources for its answer. The model rewrites the user's prompt into one or more search-engine queries, reads the results and cites some of them. Because the model only reads what ranks for these queries, they determine which websites can appear as sources in the final answer.

Query fan-out is one pattern of grounding queries: a single prompt split into several sub-queries that run in parallel, as Google describes for AI Mode. Grounding queries is the wider term. It includes a single rewritten query, parallel fan-out queries and the sequential follow-up searches a model writes after reading its first set of results.

Sometimes. Developer APIs expose them: Gemini's search grounding returns the executed queries, and OpenAI's Responses API records them in each web search call. In consumer products they are usually hidden, though some engines reveal them in the page data. AI visibility tools collect them from tracked responses for the engines that expose them.

Collect the queries models run for your important prompts, then check whether your pages rank for them in the relevant search engine. Models often add qualifiers such as the current year, 'reviews', 'pricing' or 'vs', so pages should answer those angles explicitly. Specific, dated, verifiable facts in self-contained sections make a page useful once a query retrieves it.

No. The exact wording changes between runs, models and dates, and the model may add location or memory details for individual users. What tends to persist is the structure: the same topics, modifiers and competitor names recur across many runs. Analyze queries in aggregate over weeks rather than reacting to a single response.

Be the brand AI recommends

Monitor your brand's visibility across ChatGPT, Claude, Perplexity, and Gemini. Get actionable insights and create content that gets cited by AI search engines.