---
title: "Grounding Queries vs. Impressions: The Two AI Search Metrics Teams Keep Confusing"
canonical: https://snezzi.com/blog/grounding-queries-vs-impressions/
source: https://snezzi.com/blog/grounding-queries-vs-impressions/
published: 2026-10-20
modified: 2026-11-19
author: "Team Snezzi"
category: "AI Visibility"
---

> Canonical page: https://snezzi.com/blog/grounding-queries-vs-impressions/

# Grounding Queries vs. Impressions: The Two AI Search Metrics Teams Keep Confusing

**Short answer:** Grounding queries and impressions measure different things. A grounding query, shown in Bing's AI Performance report, is a grouped phrase tied to content an AI answer cited. An impression only counts how often your link was shown. Use grounding queries to decide which pages to improve, and impressions to measure reach.

That distinction matters because a B2B marketing team can see impressions rising, citations appearing, or both, then assume it knows which content to improve. It usually does not. An impression tells you that a link appeared to a searcher. A grounding query can point you toward the topic and page associated with an AI citation, giving you a starting point for a content or page-quality decision.

Neither metric proves that a page generated a lead or influenced pipeline. The practical difference is that grounding-query data can help you decide what to inspect, while impression data mainly tells you how often your site had an opportunity to be seen. [Microsoft’s AI Performance documentation](https://www.bing.com/webmasters/help/ai-performance-9f8e7d6c) and [Google’s Search Console documentation](https://support.google.com/webmasters/answer/7042828) define these reporting concepts within different systems.

## Grounding queries and impressions are not equivalent metrics

These metrics do not represent two versions of the same measurement.

In Bing Webmaster Tools’ AI Performance report, grounding queries are grouped phrases associated with content cited in AI-generated answers. The report maps those phrases to cited pages. They are not original user prompts, complete questions, or a record of every answer in which a page appeared.

In Google Search Console, an impression means that a user saw, or potentially saw, a link to a site. Google’s generative AI performance reports show how often URLs appeared in generative AI features, but the published report fields do not include an equivalent grounding-query report. Those reports cover items such as impressions, pages, countries, devices, and dates. [Google’s announcement about generative AI performance reports](https://developers.google.com/search/blog/2026/06/gen-ai-performance-reports) describes the current reporting scope. For a side-by-side of the two reports, see [Search Console generative AI reports vs Bing AI Performance](/blog/search-console-generative-ai-reports-vs-bing-ai-performance/).

The common reporting error is treating these signals as interchangeable:

- A rise in impressions does not explain why a page appeared or whether an AI system used it as supporting evidence.
- A rise in citations does not show that the cited page ranked highly, drove traffic, or caused a business result.
- A grounding query does not reveal the exact prompt a person entered or the model’s reasoning.
- A citation change cannot normally be attributed to one page edit, content update, or other isolated event.

Bing describes its AI Performance data as aggregated citation activity. It is intended for trend and comparison analysis, not as a complete audit trail of every AI answer, prompt, or retrieval event.

## What is a grounding query?

Grounding is the process of using retrieved information to support an AI-generated response. Google describes retrieval-augmented generation, or RAG, as a technique that retrieves relevant and current pages from its Search index so an AI system can use information from those pages when generating a response. Google also distinguishes grounding from query fan-out, where a model creates related searches to gather information for a broader user request. [Google’s guide to generative AI search](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide) explains both concepts.

Bing uses the term in a more specific reporting context. Its AI Performance dashboard shows short, grouped phrases associated with content the AI retrieved and cited in supported experiences, including Microsoft Copilot, AI-generated summaries in Bing, and selected partner integrations.

A useful way to separate the terms is:

- **User prompt:** What a person asks an AI interface.
- **Grounding query in Bing’s report:** A short, grouped phrase associated with content the AI retrieved and cited.
- **Citation:** A visible reference to a page in an AI-generated answer. ([How AI citation tracking works](/features/ai-citation-tracking/).)
- **Grounding-query count:** How often content from the site was referenced in answers associated with that grouped phrase during the selected date range.

For example, a buyer might ask an AI system a detailed question about choosing software with single sign-on, audit logs, and compliance documentation. The report might group related retrieval activity under a shorter phrase such as “SOC 2 SSO audit logs software procurement.” That phrase is a diagnostic theme, not necessarily the words the buyer used.

A single grounding query can map to several pages. A single page can also appear under several grounding queries. That mapping can reveal which pages are being used around a topic, but it does not show which page was most important in an individual answer, why it was selected, or how the answer was composed.

The phrases can also vary across supported AI experiences and partner integrations. Bing says the data is grouped and summarized, and that low-volume activity may not surface. Treat the result as a useful clue about recurring themes, not as a keyword-volume dataset.

## What is an impression?

An impression is an exposure count.

In Google Search Console, an impression means that a user saw, or potentially saw, a link to a site in Google Search, Discover, or News. The exact counting rule varies by result type. For example, a result may count when it appears in the current set of results, while items in some carousels or expandable areas may need to be scrolled into view or expanded.

Bing’s Search Performance reporting similarly counts instances in which a link appears across surfaces such as web search, chat responses, news, images, and knowledge panels. The scope of an impression therefore depends on the platform and surface being reported. [Bing’s Search Performance documentation](https://www.bing.com/webmasters/help/search-performance-c680da36) provides the Bing-specific definitions.

An impression can answer a useful question: **Did our site have an opportunity to be seen?**

It cannot answer several content-quality questions:

- Why did the page appear?
- Did an AI system use the page as evidence?
- Did the searcher notice or value the result?
- Was the page the best answer for the underlying need?
- What specific information should the team add or clarify?

Here is the practical comparison:

| Question | Grounding-query data | Impression data |
|---|---|---|
| What does it represent? | A grouped phrase associated with AI citation activity | An instance in which a site link appeared to a searcher |
| Main question it answers | “What topic and page combination is associated with our AI citations?” | “How often did people have an opportunity to see us?” |
| Best use | Form a page or content hypothesis | Track visibility trends and assess reach with clicks or sessions |
| What it cannot prove | The exact prompt, ranking, business impact, or causality | Why the page appeared, whether AI used it as evidence, or content quality |

Impressions are not useless. They are useful for directional reach measurement and become more informative when reviewed with clicks, referral sessions, qualified leads, and pipeline. The mistake is using impressions alone to decide which content deserves an edit.

## The five-step query-to-page prioritization method

Use grounding queries as a starting point for investigation, then use other metrics to evaluate reach and business results.

### 1. Separate measurement sources before comparing numbers

Label every export by:

- Platform
- Surface
- Date range
- Metric
- URL or page grouping

For example, keep these as separate data sets:

- Bing AI citations
- Bing grounding queries
- Bing search impressions
- Google generative AI impressions
- Google Search impressions
- AI referral sessions
- Qualified leads and pipeline

Do not add all of these values together into one “AI visibility” number. A Bing citation count and a Google generative AI impression count are produced by different systems and describe different events. The same problem applies to blended scores such as [share of model](/blog/what-is-share-of-model-metric-llm-presence-explained/): keep the per-platform numbers underneath any summary figure.

Microsoft introduced AI Performance in Bing Webmaster Tools as a public preview. Microsoft says its data is aggregated and sampled, may omit low-volume activity, and can differ between views such as pages, grounding queries, and time-series charts. [Microsoft’s public-preview announcement](https://blogs.bing.com/webmaster/February-2026/Introducing-AI-Performance-in-Bing-Webmaster-Tools-Public-Preview) describes the report as an early set of insights rather than a complete record.

### 2. Select a grounding-query theme with meaningful citation activity

Start with non-branded B2B buyer-research themes, such as:

- Product comparisons
- Implementation requirements
- Security and compliance
- Integrations
- Pricing evaluation
- Migration or replacement decisions

Do not automatically select the phrase with the highest count. A high-volume theme may be broad and difficult to connect to a specific page decision. A smaller theme may reveal a more important buyer concern, especially if it relates to a high-value use case.

The goal is to find a recurring topic where your site is already being cited but the associated page may not fully support the buyer’s decision.

### 3. Inspect the query-to-page mapping

For each selected phrase, record:

- The cited URLs
- Each page’s current purpose
- The date each page was last reviewed
- The buyer question each page answers
- The information each page does not answer
- Any relevant conversion or referral data

Look for one of three gaps:

1. **Missing evidence:** The page makes a claim but does not provide enough documentation, examples, conditions, or limitations for a buyer to verify it.
2. **Unclear structure:** The information exists, but important details are buried, vague, or difficult to compare.
3. **Missing buyer-use-case page:** The existing pages are accurate but do not address a distinct decision, such as implementation, security review, or integration fit.

Remember that the mapping is not a ranking report. A page with more citations is not automatically more authoritative or more important. It may simply be associated with more reported activity during the selected period.

### 4. Make one evidence-led improvement

Choose one meaningful change instead of rewriting everything at once.

Depending on the gap, that might mean adding:

- Supported integrations
- Implementation prerequisites
- Product limits
- Security controls
- Audit-log scope
- Pricing conditions
- Configuration requirements
- A clear comparison of included and excluded capabilities
- Links to relevant trust, security, or implementation documentation

Use descriptive headings and a comparison table when they make a decision easier to understand. Do not create a thin page for every reported phrase or repeat the grounding-query wording mechanically in headings and copy.

Google specifically says there is no need to create separate content for every query variation. Its guidance favors useful, people-first content that addresses a subject clearly rather than a collection of pages designed around minor wording differences.

Record the exact change, the page version, and the date it was published. Without that record, a later citation change can look like proof of an edit’s impact even when other factors changed at the same time.

### 5. Judge results in layers

Review the outcome in three stages.

**First, check eligibility and access.** Confirm that the page is accessible, crawlable, and indexed where relevant. A page cannot provide a useful signal if the systems reporting it cannot access or process it.

**Next, review visibility signals.** Compare citation activity, grounding-query mappings, and AI-feature impressions over similar periods. Check whether the relevant page remains associated with the topic. Avoid drawing conclusions from a single day or a small change in a sampled report.

**Finally, review observable business activity.** Use analytics and correctly configured conversion events to assess:

- AI referral sessions (see [how to track AI referral traffic](/features/ai-referral-traffic/))
- Qualified demo requests
- Lead quality
- Pipeline contribution
- Sales outcomes

Citations and impressions alone do not demonstrate leads or revenue. A citation is not traffic, and an impression is not a conversion.

A monthly prioritization review is a reasonable operating cadence for deciding which pages to improve. A weekly anomaly check can help identify sudden reporting changes, missing data, or technical problems. These are practical working rhythms, not universal requirements.

A broader [AI visibility measurement](/features/ai-visibility-tracking/) process can help consolidate source-specific visibility signals with observable referral and conversion data, as long as the reporting keeps each platform’s definitions separate.

## Hypothetical example: turning a grounding query into a content decision

> **Hypothetical example**
>
> A B2B SaaS company sees repeated Bing AI citation activity for the grouped phrase “SOC 2 SSO audit logs software procurement.” The query-to-page mapping shows that its general product page is cited, but its security page is outdated and does not clearly state audit-log scope, SSO requirements, or where buyers can verify compliance documentation.

The team should not publish a keyword-stuffed page titled with that exact phrase. The phrase is a clue about a buyer-research theme, not a publishing brief.

Instead, the team could refresh the security page with:

- Date-checked factual statements
- A clear definition of the relevant security capabilities
- A table showing what is included and what requires configuration
- SSO requirements and implementation conditions
- The scope of available audit logs
- Links to applicable trust or implementation documentation
- A clear explanation of any product or plan limitations

The team should then monitor whether the page continues to appear in the grounding-query to page mapping and whether observable referral activity produces qualified demo requests.

That sequence does not prove that the page refresh caused a citation change. It creates a better page for the buyer and a cleaner basis for evaluating later signals.

A sensible success check would ask:

1. Is the page technically accessible and eligible to be indexed?
2. Are its factual claims verifiable on the page or through linked documentation?
3. Does the page answer the procurement concern more clearly than before?
4. Is the relevant citation pattern stable or improving over comparable periods?
5. Are referral sessions and conversion events being interpreted separately from citation counts?

## Summary: the decision rule

Use impressions to ask, **“Were we seen?”**

Use grounding queries to ask, **“What information are we being selected to support?”**

Use neither alone to claim pipeline impact.

For B2B marketing teams, impressions remain useful for directional reach. Grounding queries are more useful for content prioritization because they connect an AI citation signal to a topic and a set of pages that can be inspected. The strongest measurement approach keeps those diagnostic signals separate from clicks, referral sessions, qualified leads, and pipeline, then records page changes carefully enough to avoid treating coincidence as causation.

Your next step: export last month's grounding queries from Bing Webmaster Tools, pick one non-branded theme that maps to a page you can improve, and make one documented change. Review the same report and your referral data a month later.

## Frequently asked questions

### What is a grounding query in Bing Webmaster Tools?

It is a short, grouped phrase that Bing's AI Performance report links to content an AI answer retrieved and cited. It is not the exact prompt a person typed, and low-volume activity may not appear.

### Does Google Search Console show grounding queries?

No. Google's generative AI performance reports show impressions, pages, countries, devices and dates, but no grounding-query view like Bing's.

### Are AI impressions the same as AI citations?

No. An impression means your link was shown. A citation means an AI answer referenced your page as a source. One can rise while the other stays flat.

### Which metric should I report to leadership?

Report reach (impressions) and business results (AI referral sessions, qualified leads) side by side. Use grounding queries internally to choose which pages to improve, not as a headline number.
