# Grounding Queries: Definition, Examples and How to Find Them

- Canonical: https://www.geoly.ai/blog/grounding-queries
- Published: 2026-07-05
- Updated: 2026-11-14
- Author: [GEOly AI](https://www.geoly.ai/blog/author/geoly-ai)
- Tags: [Glossary](https://www.geoly.ai/blog/tag/glossary), [AI Search](https://www.geoly.ai/blog/tag/ai-search), [GEO](https://www.geoly.ai/blog/tag/geo)

> **Summary:** What are grounding queries? See a worked example, how they relate to query fan-out, and where Bing, Clarity and API traces expose retrieval evidence.

A grounding query is a search an AI system issues to retrieve evidence for an answer. It may reuse a user’s words or reformulate them. Query fan-out describes expanding a request into multiple searches; those searches can be used for grounding. Retrieval and citation behavior varies by engine and request.

## The one-sentence definition

When you ask a large language model a question, one of the first things it does is decide: *can I answer this from memory, or do I need to check the web?* If it needs to check, it silently rewrites your question into one or more precise, fact-shaped searches, runs them against a live index (Google for Gemini and AI Mode, Bing for Copilot), reads the results, and grounds its answer in what it found. **Those machine-generated searches are grounding queries.**

Grounding can connect an answer to external evidence, but it does not guarantee factual accuracy or a citation. A brand can also be mentioned from model knowledge. Distinguish mentions, retrieved pages, visible citations and website visits when measuring outcomes.

## The three-layer query stack: typed query → fan-out → grounding

The following stack is a useful research model, not a mandatory three-stage architecture. A system can fan out into several grounding queries, run one search, or answer without web retrieval.

![Figure 1 — How one prompt becomes many machine searches: fan-out for breadth, grounding for depth.](https://gslkfhamqhkktynnjqng.supabase.co/storage/v1/object/public/cms-media/grounding-fig1-en.png)

**Layer 1 — The query you typed.** What the human enters: *"Which engagement ring has the best diamond quality for the price?"* Conversational, messy, full of intent.

**Layer 2 — Fan-out queries.** Semrush defines [query fan-out](https://www.geoly.ai/blog/query-fan-out) as "an AI search system process that splits a user query into multiple sub-queries, collects information for each sub-query, then merges relevant information into a single response." The AI decomposes your one prompt into a *breadth* of parallel sub-searches to build a rich picture: *"best value engagement ring brands,"* *"lab-grown vs natural diamond quality,"* *"engagement ring buying guide 2026,"* *"DJI vs GoPro price."* Google's Head of Search frames AI Mode as "bringing a whole new level of intelligence to search" — in one documented example, **Google AI Mode fired eight simultaneous searches** to answer a single vacation-planning prompt. Fan-out is about **coverage**: casting a wide net across many source types (official pages, reviews, forums, expert posts).

Layer 3 — Grounding evidence. The retrieved pages provide material for the answer. Grounding queries can be broad comparisons or narrow fact checks; they are not always a separate step after fan-out.

The practical takeaway: cover the distinct questions a buyer needs answered, and support important claims with clear evidence. Breadth and factual detail are complementary editorial goals, not guarantees of inclusion.

## How grounding actually works (under the hood)

Grounding is an application of **Retrieval-Augmented Generation (RAG)**: the model's fluent-but-frozen language ability is tethered to an external, current source of truth. Google's own Gemini API documents the pipeline explicitly, and it generalizes to every grounded AI system.

![Figure 2 — The grounding decision (stable vs. volatile) and the retrieval pipeline.](https://gslkfhamqhkktynnjqng.supabase.co/storage/v1/object/public/cms-media/grounding-fig2-en.png)

1. **The model reads your prompt and decides whether a search would improve the answer.** This is a routing decision, not a reflex — stable facts skip it, volatile facts trigger it (see the next section).
2. **If grounding is needed, the model generates one or more search queries automatically** — reformulating your natural-language question into terse, retrieval-optimized strings.
3. **It executes those searches, reads the returned passages, and synthesizes them** into a single answer.
4. **It returns the answer plus grounding metadata** that names its work.

In the Gemini API, that metadata is inspectable, and it is the clearest window into how the black box actually behaves:

| **Field** | **What it exposes** |
| --- | --- |
| `webSearchQueries` | **The exact grounding queries the model ran** — the reformulated searches, not your prompt |
| `groundingChunks` | The web sources (URIs + titles) it retrieved and pulled from |
| `groundingSupports` | Which *sentences* in the answer map to which sources — passage-level attribution |
| `searchEntryPoint` | The rendered "Search Suggestions" chips shown to users |

Two things follow from this that most SEOs miss:

- **The unit of retrieval is a passage, not a page.** `groundingSupports` ties individual claims to individual chunks. A page can be grounded for one sentence and ignored for the rest. You optimize passages, not documents.
- **Grounding has a price, so the machine is selective.** Google bills Grounding with Google Search at roughly **$35 per 1,000 grounded queries** (and, on Gemini 3, per *each* search the model chooses to run). Because grounding costs money and latency, models only ground when the payoff is real — which is exactly why *which* queries get grounded is predictable and strategic.

## Grounding vs. fan-out, side by side

![Figure 3 — Two retrieval layers, two different ways to win.](https://gslkfhamqhkktynnjqng.supabase.co/storage/v1/object/public/cms-media/grounding-fig3-en.png)

| **Dimension** | **Grounding query** | **Fan-out query** |
| --- | --- | --- |
| Purpose | Verify one specific claim | Map the whole information landscape |
| Breadth | Narrow — 1–2 authoritative sources | Wide — many parallel sub-queries |
| Best-fit content | Primary research, statistics, specs, prices | Comparisons, thought leadership, FAQs |
| Preferred sources | Structured, institutional, "of record" | Diverse — expert, social, review, official |
| Citation behavior | Often *uncited* but shapes answer confidence | More often *visibly cited* |
| How to win | Be *the* citable source of a fact | Be *present everywhere* the topic is discussed |

## Which queries actually get grounded? (The most useful insight in the field)

Not every query triggers grounding — and knowing which ones do is where the strategy lives. The sharpest framing here comes from Mark Williams-Cook: **AI systems ground queries whose answers are** ***volatile*****, and skip queries whose answers are** ***stable*****.**

- *"What do red blood cells do?"* → **No grounding.** The answer is settled knowledge; the model's token predictions are highly consistent. To change how AI answers this, someone would have to retrain the model.
- *"What happened in the news today?"* / *"best AI visibility tool 2026"* / *"is X in stock"* → **Grounding.** Answers shift over time, the model's internal confidence is low/variable, so it reaches for the live web.

This variability is even measurable. ChatGPT exposes an internal `search_prob` signal; a score around **0.65+** indicates the model is likely to run a web search. Google's Gemini API lets you inspect grounding directly. Community ML models estimate grounding likelihood too.

Why this matters so much:

- **Grounded queries can be influenced in** ***days*****.** Publish or update the right fact-shaped content, get it crawled, and you can move the AI's answer almost immediately — no model retraining required.
- **Non-grounded ("solved knowledge") queries are near-immovable** in the short term. Chasing them with content is wasted effort.
- **This lets you** ***triage*** **a keyword list.** Of, say, 500 related keywords, 200+ may be stable-knowledge queries not worth near-term investment. Concentrate your GEO budget on the grounded, convertible subset.

## Why grounding queries are the foundation of GEO

Traditional SEO optimizes for **what humans type**. Generative Engine Optimization (GEO) optimizes for **what machines search**. That is the whole shift in one line.

![Figure 4 — Retrieval is the master gate: without it, every AI-visibility metric is zero.](https://gslkfhamqhkktynnjqng.supabase.co/storage/v1/object/public/cms-media/grounding-fig4-en.png)

Because the user never sees the grounding query, the entire competition happens inside a **zero-click black box** — observable only from the outputs unless a platform hands you the data. And the stakes are counterintuitive:

- **Retrieval is the gatekeeper.** Every downstream metric — mention rate, citation rate, product-card inclusion, AI share of voice — sits *below* retrieval. If you don't surface for the grounding query, none of them can be non-zero.
- **AI citations don't follow classic rankings.** Roughly **67% of AI citations do not rank in the top 10** for the corresponding query. Pages have accumulated **1,000+ AI grounding citations** while remaining near-invisible in blue-link search. The old scoreboard doesn't measure the new game.
- **The source mix is different, too.** Community sources punch above their weight — Reddit alone accounts for roughly **one in five** off-site AI citations, and its share has been climbing. Grounding rewards being genuinely present where a topic is discussed, not just owning a polished landing page.

## Case study: DJI in the Cameras category

Everything above is abstract until you watch GEOly analyze a real brand. The three screenshots below are current-version GEOly interface captures from a DJI workspace, created in September 2026. The dated metrics come from the latest structured repository snapshot available in this checkout: Cameras, US/en, ChatGPT, August 1–31, 2026. The interface captures and repository snapshot cover related but distinct scopes: the screenshots show what can be inspected, while the snapshot supplies the category metrics.

### Section 1 — DJI in the latest repository category snapshot

![GEOly DJI 趋势、竞品与 AI Search Keywords 桌面端高清截图（当前版本，US/en）](https://gslkfhamqhkktynnjqng.supabase.co/storage/v1/object/public/cms-media/codex-clipboard-35ad85c7-6c88-44a8-952e-978aadd3ff76.png)

The latest structured camera-category dataset available in this repository covers August 1–31, 2026, US/en, ChatGPT, across four product spaces: action cameras, 360 cameras, pocket cameras, and camcorders. It contains 100 brands, 56 topics, and 6,029 aggregate mentions. DJI ranks #1 with 899 mentions across 42 topics, and its repository GEO Index is 98.7. These are dated category observations, not current DJI account totals, traffic, revenue, or proof of causality. The repository traffic and revenue fields are modeled monthly values with API windowApplied=false, so they are intentionally not used here.

### Section 2 — DJI prompt-level records

![DJI 品牌 Prompt 详情与执行记录（当前版本）](https://gslkfhamqhkktynnjqng.supabase.co/storage/v1/object/public/cms-media/dji-prompts-detail-latest-20260904.png)

The current-version DJI prompt screens show the next layer down: individual prompts, their execution records, and the query-level evidence GEOly exposes for inspection. Use this view to examine how a particular answer was generated and which searches or sources were attached to that record. It should be read as a record-level diagnostic, not as a replacement for the dated category aggregate above.

- *Prompt wording and intent: what the user asked and how the record is scoped.*
- *Execution details: which query-level steps or retrieved material GEOly exposes for that record.*
- *Citation evidence: which domains or URLs are attached to the answer, with the engine and time range kept in view.*
- *Workspace context: which DJI workspace, locale, and platform were used for the record.*
- *Interpretation boundary: one record explains one answer path; it does not establish category-wide visibility.*

The important distinction is between an individual DJI record and the aggregate category snapshot: a prompt-level observation can explain one answer path, but it does not establish DJI's overall category visibility. Keep engine, locale, time range, and source coverage attached to every interpretation.

### Section 3 — Citation sources and AI answers in the DJI workspace

![GEOly DJI Top Citation Domains 与 AI 回答桌面端高清截图（当前版本，US/en）](https://gslkfhamqhkktynnjqng.supabase.co/storage/v1/object/public/cms-media/codex-clipboard-ea090032-dd61-4696-84c7-26160226b002.png)

The citation-source and AI-answer view connects the monitored response to the sources GEOly recorded for that response. It is useful for asking which domains or URLs are being cited, whether the source is first-party or third-party, and what evidence is available for follow-up. Treat the screenshot as an interface snapshot from September 2026; do not turn one answer's citations into a category-wide ranking claim.

- *Source identity: the domain or URL attached to the monitored answer.*
- *Answer context: the exact answer record in which the source appeared.*
- *Scope check: the engine, locale, time range, and available source coverage before drawing a conclusion.*
- *Follow-up path: inspect the underlying prompt and execution record before recommending an optimization.*

This view may expose multiple query variations for one prompt, but the presence or absence of a brand in that answer is record-level evidence only; it should not be generalized to the full Cameras category.

> What the case suggests: the latest repository snapshot gives DJI a strong dated position in the Cameras category, while the current-version UI screenshots show how GEOly can be used to inspect the prompt, execution record, and citations behind individual answers. These are different levels of evidence: the category snapshot describes the monitored window; the UI records help investigate why a particular answer looked the way it did. Neither alone proves future performance or causality.

## How to actually *see* grounding queries

You can't optimize what you can't observe. Four windows exist today, in rough order of directness:

1. **Bing Webmaster Tools → AI Performance report.** Rolled out in public preview February 2026 and expanded in June 2026 with **Intents, Topics, Citation Share, and a Compare view**. This is the first mainstream console to expose grounding-query-level exposure and cited URLs for Copilot/Bing-powered answers.
2. **The Gemini API** **`groundingMetadata`****.** If you build against Gemini, `webSearchQueries` literally returns the searches the model ran — the ground truth for how your prompts get grounded.
3. **Prompt-level monitoring across engines.** Because each engine reformulates differently, the durable method is to probe your target questions at scale and inspect *what each engine cites*. [GEOly](https://www.geoly.ai/overview), for example, tracks prompt-level outcomes across **seven engines** — ChatGPT, Gemini, Perplexity, Copilot, Grok, Google AI Mode, and AI Overviews — and reverse-engineers the grounding and fan-out queries behind the answers.
4. **`search_prob`** **/ grounding-likelihood signals** to classify which of your keywords are grounded vs. solved *before* you spend a dollar of content budget.

### Microsoft Clarity: the first free, first-party window

Microsoft Clarity’s Citations dashboard provides grounding-query and cited-page views for supported experiences. Verify your project and inspect the report’s scope and available data. It is a useful observation source, not a universal log of every AI assistant’s searches.

One boundary has to be stated plainly: **Clarity’s data covers Microsoft’s ecosystem only** — Copilot and Bing generative search. It cannot show you citations in ChatGPT, Gemini, Perplexity or Claude, because those platforms do not share their internal retrieval logs with Microsoft. That makes Clarity a high-value but narrow window: ideal for seeing what grounding queries genuinely look like, not a cross-engine picture.

How to use it: export the real Copilot grounding queries from Clarity and hold them against your keyword list. Most of them will not exist in a conventional keyword tool at all — they are too long, too specific, comparative and attribute-driven. That gap is exactly where grounding happens, and it is why a keyword report cannot tell you whether you are winning it.

Further reading: [the Clarity Citations availability announcement](https://clarity.microsoft.com/blog/citations-now-generally-available/).

## How to win grounding queries: an eight-point playbook

Grounding rewards content that is **precise, atomic, fresh, and structured for extraction.** Do these:

1. **Fill fact-shaped gaps.** Grounding queries hunt data points: prices, specs, comparisons, dates, percentages, "how many," "as of when." If the citable fact about your category doesn't exist on your site in clean form, someone else's does.
2. **Publish primary research.** Original surveys, proprietary statistics, and regularly-updated benchmarks are the *most undervalued* GEO asset. They make you the terminal source of a fact — the thing a grounding query is built to find.
3. **Write in atomic passages ("chunks").** Because the retrieval unit is the passage (`groundingSupports`), lead each section with a direct, self-contained answer, then support it. Every heading should be able to stand alone as an answer to a sub-question — Semrush calls this "writing for NLP."
4. **Use tables and lists.** Structured data is trivially machine-extractable and disproportionately grounded. Comparison tables, spec sheets, and numbered steps are grounding magnets.
5. **Build topic clusters for the fan-out layer.** A pillar page for the broad topic, linked to granular cluster pages for each sub-query, builds the topical authority AI systems prioritize — and covers the breadth of sub-queries a single prompt fans out into.
6. **Implement structured markup (JSON-LD).** Schema (Product, Offer, FAQ, etc.) makes claims unambiguous to machines and meaningfully lifts extraction and snippet inclusion.
7. **Keep it fresh and dated.** Volatile-answer queries are the *only* ones that get grounded — so freshness is the signal. Put visible "last updated" dates on data. Stale statistics silently lose the citation.
8. **Designate one canonical URL per fact, and work the third-party layer.** Don't fragment your best statistic across five posts — concentrate authority on a single, linkable, updatable source of record. Then reinforce presence off-domain: reviews, analyst mentions, and authentic community participation (Reddit, niche forums) win the fan-out layer.

And the table stakes: **stay crawlable.** If your `robots.txt` blocks the AI crawlers, or your facts are locked behind JavaScript the retriever can't render, you've lost the grounding query before it started.

> **Case study — Stripe.** Semrush highlights Stripe as a model of two-track optimization: solution pages tailored to different business stages and use cases (fan-out breadth), detailed subsections answering specific sub-topics with direct explanations (grounding depth), content distributed across blog, guides, and support center, all cleanly structured. The result: Stripe out-visibilities competitors across Google AI Mode, ChatGPT/SearchGPT, Perplexity, and Gemini.

## A practical five-step diagnostic

1. **Map intent to retrieval type.** For each target query, classify: grounded or solved? Grounding-heavy or fan-out-heavy?
2. **Audit your grounding candidacy.** Do you own citable, precise, current facts for your category — or only prose?
3. **Audit your fan-out coverage.** Are you present across the *breadth* of source types the AI samples, or just your own site?
4. **Find the gaps.** Where does the AI answer your category question *without* you? Those are your retrieval holes.
5. **Build the two-track content strategy.** Deep, canonical, fact-of-record pages for grounding; distributed, granular, independently-useful content for fan-out.

## Common mistakes

- **Optimizing the typed keyword, not the grounding query.** You're writing for a search the machine will never run in that form — treating grounding queries like 2019 head-term keywords.
- **Chasing solved-knowledge queries.** No amount of content moves an answer the model already "knows." Triage first.
- **Watching only one engine.** Each engine reformulates differently; Bing's report is a proxy, not the whole picture.
- **One giant pillar page.** Grounding wants atomic, extractable passages, not a 5,000-word wall. Fan-out wants breadth, not a single monolith.
- **Expecting on-site content to win queries that resolve to third-party sources.** If Reddit and review sites define your category to the AI and you're absent there, you lose the fan-out layer entirely.
- **Assuming top-10 rankings = AI visibility.** Two-thirds of AI citations come from outside the top 10. Different game, different scoreboard.

## A worked grounding-query example

Consider the prompt “Which project management tool supports SSO for a 20-person team?” Possible retrieval searches include “project management SSO plan limits” and a vendor-specific security documentation query. These are illustrative examples, not a captured engine trace. One search establishes the shortlist; another checks a purchasing constraint.

Map each question to evidence: a comparison page for the shortlist, current plan limits for availability, and security documentation for SSO configuration. Avoid creating a separate thin page for every wording variation.

## Where to find grounding queries: Clarity, Bing and API traces

- **Bing Webmaster Tools:** inspect AI Performance for your verified site. Its grounding-query phrases and cited pages describe supported AI experiences and sampled citation activity, not a complete cross-engine query log. See [Microsoft’s report documentation](https://blogs.bing.com/webmaster/2026/2/Introducing-AI-Performance-in-Bing-Webmaster-Tools-Public-Preview/).
- **Microsoft Clarity:** use the Citations dashboard when available for your verified project. Inspect cited pages and associated queries; keep citation measurements separate from referral visits. See the [Clarity Citations announcement](https://clarity.microsoft.com/blog/citations-now-generally-available/).
- **Your own API requests:** inspect search and grounding details returned by the API you used. [Google’s grounding documentation](https://ai.google.dev/gemini-api/docs/google-search) describes its supported interface. An API observation is evidence about that request, not necessarily about a consumer search session.
- **Google Search Console:** use it for search performance and page-query mapping. It is not a complete log of the searches an AI generates internally.

For the practical research workflow, continue to [how to find grounding queries](https://www.geoly.ai/blog/how-to-find-grounding-queries). To distinguish retrieval counts from human demand, read [prompt volume](https://www.geoly.ai/blog/prompt-volume).

## FAQ

### Are grounding queries the same as the keywords I track in Search Console?

No. Your tracked keywords are what *humans* type. Grounding queries are the reformulated, fact-shaped searches the *AI* generates *after* reading your prompt. They frequently target an intent your keyword research never surfaced.

### What's the difference between a grounding query and query fan-out?

Fan-out is the expansion of a request into multiple retrieval queries. A grounding query is a search used to obtain evidence. These concepts overlap: the searches generated through fan-out may be grounding queries. Grounding can also use one search or a non-web source.

### How can I see ChatGPT's grounding queries?

Available traces depend on the product and interface. Inspect the retrieval or source details provided for that specific request. Bing and Clarity observations describe their supported experiences; they are not a feed of ChatGPT’s internal searches. A cited URL alone does not reveal every query that led to it.

### Can I control which grounding queries an AI generates?

Not directly — the model generates them. But you *can* control whether you're the best answer *when* it runs one (via entity naming, consistent terminology, and category coverage), and you can predict which of your queries get grounded (volatile vs. stable) to focus effort.

### How fast can I influence a grounded answer?

There is no guaranteed timeline. A published correction must be accessible and retrieved before it can influence a grounded answer; it may never be selected. Track crawl access, sampled answers and cited URLs separately rather than promising a change within days.

## Where GEOly fits

The DJI case above uses a dated repository snapshot plus current-version interface captures. GEOly can connect category-level visibility observations with prompt, citation, and answer records across supported engines; exact coverage depends on workspace configuration, engine, locale, and time range. Use the aggregate report to see the monitored position, then inspect the underlying records before drawing an optimization conclusion.

[**Run the same grounding-query analysis for your own brand — free — at app.geoly.ai.**](https://www.geoly.ai/auth/register) In minutes you'll see the real phrases the AI runs in your category, where you're present, where rivals win instead, and the exact gaps to close.

*Data source:* [*www.geoly.ai*](https://www.geoly.ai)*.*

## Sources & further reading

- [What Is Query Fan-Out? How It Works & How to Optimize — Semrush](https://www.semrush.com/blog/query-fan-out)
- [Grounding Queries vs Fan-Out Queries: The Two Hidden Layers of AI Search Visibility — Adapt Worldwide](https://www.adaptworldwide.com/2026/06/29/grounding-queries-vs-fan-out-queries-the-two-hidden-layers-of-ai-search-visibility/)
- [Identify searches that need grounding to focus your AI strategy — Mark Williams-Cook, Majestic "SEO in 2026"](https://majestic.com/seo-in-2026/mark-williams-cook)
- [Grounding Queries: The Definitive Guide to Verifiable AI [2026] — AI Search Rankings](https://www.aisearchrankings.com/content/grounding-queries/grounding-queries-the-definitive-guide-to-verifiable-ai-2026.php)
- [Grounding Queries vs Query Fan-Out: An SEO Guide to AI Search — O Positive](https://opositive.io/blog/grounding-queries-vs-query-fan-out)
- [Grounding with Google Search — Gemini API, Google AI for Developers](https://ai.google.dev/gemini-api/docs/google-search)
- [Google's Guide to Optimizing for Generative AI Features on Search — Google Search Central](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide)
- [The Complete Guide to Generative Engine Optimization (GEO) — Peec AI](https://peec.ai/blog/the-complete-guide-to-generative-engine-optimization-geo)
- [Generative Engine Optimization — Wikipedia](https://en.wikipedia.org/wiki/Generative_engine_optimization)

*Updated October 2026 with the latest repository-backed DJI camera-category snapshot available in this checkout (August 1–31, 2026, US/en, ChatGPT). The foundational explanation remains based on the top-ranking results for "grounding queries," Semrush's query fan-out guide, and Google's official Gemini grounding documentation.*

Related reading: [how to find your own grounding queries](https://www.geoly.ai/blog/how-to-find-grounding-queries).

## Connect query evidence to a content action

Interpret figures within the platform, sample and observation window described in this article; they are not universal rules for AI products. Fan-out describes expansion across related needs, while grounding describes supporting an answer with external information. These can overlap. Keep observed searches, generated suggestions, retrieved pages and final citations distinct.

Start with the [free planner](https://www.geoly.ai/free-tools/query-fanout-analyzer), then use the [content optimization workflow](https://www.geoly.ai/blog/query-fanout-content-optimization) to review existing pages. To evaluate the evidence, read the [platform comparison](https://www.geoly.ai/blog/chatgpt-google-query-fanout) and [tool selection guide](https://www.geoly.ai/blog/query-fanout-tools). Organize team actions with the [content gap solution](https://www.geoly.ai/solutions/query-fanout-content-gaps), then investigate subsequent observations with [query analysis](https://www.geoly.ai/features/query-fanout-analysis).
