How to measure visibility in AI search: making sense of Search Console generative AI reporting

Executives increasingly ask a simple question: does our brand appear in AI search? Until recently, many teams tried to answer it through manual prompts, third-party visibility scores or broad changes in organic traffic. Those methods can provide clues, but they do not observe the same population or use the same definitions. A repeated prompt sample is not market share. A vendor score is not automatically first-party exposure data. A traffic trend does not identify which answer caused it.

In June 2026, Google announced dedicated Search Console generative AI performance reports. Bing Webmaster Tools introduced AI Performance in public preview earlier in the year. These developments create useful first-party evidence, but not complete attribution. The measurement task is to use what each platform explicitly reports, connect it to existing website and CRM outcomes, and state what remains unobserved.

The short answer

Begin with the current official reports and date their definitions. Google currently exposes impressions for supported generative AI features, with page, country, date and device views, to a subset of eligible properties. The dedicated view does not currently provide a complete query, click, position or answer-citation report. Bing’s public preview uses citation-oriented measures and sampled grounding phrases, which are not directly comparable with Google’s impressions. Add landing-page behaviour, qualified actions and CRM outcomes as separate layers. Use controlled prompt research qualitatively. Report a compact evidence set, limitations and change log rather than one artificial visibility score.

The boundary will change as first-party reports evolve, so date every definition.

AI-search measurement has entered a new phase

Dedicated first-party reporting matters because it moves part of the question from observation to platform data. It also creates a new risk: treating a newly visible measure as the whole outcome. An impression can show that a URL from the property appeared in a supported AI feature under the platform’s current counting rules. It does not show that the user read the citation, formed a preference, visited later through another channel or became a customer.

Build a measurement dictionary before a dashboard. Record the official report name, rollout status, included experiences, metric definition, aggregation, dimensions, time zone, data history, export behaviour and known exclusions. Add the date checked and a link to official documentation. AI-search products and reports are changing quickly; a correct screenshot can become a misleading artefact if the definition is not preserved beside it.

What Google’s generative AI report currently shows

At the research date, Google’s Search generative AI performance report was rolling out to a subset of website owners. Google says the report includes impressions from AI Overviews and AI Mode, with a separate report for generative AI features in Discover. In the Search report, an impression is counted when links to the site are shown to a user in a supported generative AI feature. The current report can group evidence by page, country, date and device.

Google assigns most page-level performance data to the canonical URL, after redirects. The chart aggregates by property unless a URL filter is applied; the table aggregation changes with the selected dimension. Google notes that chart and table totals can differ because of aggregation. Dates use Pacific Time, and the report inherits usual Search performance limitations, including the stated row limit. Export is available for chart and table data. All of these details must be rechecked immediately before publication.

What the Google report does not currently show

The dedicated report is currently an impression view, not a complete answer-level attribution system. Its documented dimensions do not include a query tab, a cited passage, an answer transcript or a traditional average-position measure. Search Labs experiments are excluded. A property may not see the report because rollout is limited, the site has insufficient impressions or eligibility controls affect inclusion. Absence of a report is therefore not proof of zero AI visibility.

The report also cannot show every later business effect. A user may see a brand, visit through a different query, type the domain directly, involve a colleague or return weeks later. Identity and consent boundaries limit connection across sessions and systems. Report unknowns explicitly: unexposed prompts, unseen answer content, complete competitor presence, model reasoning and commercial causation. A useful dashboard does not hide these gaps behind a composite score.

Relate the dedicated view to overall Search performance

Google says generative AI visibility is included in the overall Search Console performance report within the Web search type, while the dedicated report provides a separate view. Do not add dedicated generative AI impressions to total Web impressions as if they were separate traffic. Use the dedicated view to understand a component of the broader total. Record whether a dashboard extracts overlapping views and label any share calculation carefully.

Compare changes over aligned dates rather than treating the two views as independent channels. A rise in dedicated AI impressions may coincide with a fall, rise or stable pattern in overall clicks. Investigate pages, countries, devices, brand activity, content releases, seasonality and general demand. The dedicated report currently does not create a separate revenue source in analytics. Any attempt to connect it to outcomes is an inference with limits, not a deterministic attribution path.

Add Bing’s citation-level evidence carefully

Bing’s AI Performance public preview uses a different measurement language. Microsoft describes total citations, average cited pages, sampled grounding queries, page-level citation activity and trends across supported AI experiences. It states that citation counts do not indicate placement, ranking, authority or the role of a page within an individual answer. Grounding-query phrases represent a sample of overall citation activity.

Do not merge Bing citations with Google impressions into one trend line. A citation and an impression are not the same event, and platform coverage differs. Keep separate panels with their own definitions and included experiences. Bing evidence can help identify pages and themes that are being referenced, while Google evidence can show impression patterns for its supported features. Both remain platform-specific and incomplete.

Avoid competitive league tables unless every brand is observed through the same documented method. Search Console and Bing Webmaster Tools expose only the verified organisation’s first-party property data, not a complete competitor dataset. A third-party tool may sample public answers, but differences in indexed pages, prompts, countries, personalisation and citation detection remain. If a comparison is still useful, label it as an observed sample, disclose the date and universe, and avoid turning a directional gap into a claim about total category visibility. Preserve the underlying observations so another reviewer can reproduce the comparison under the same stated conditions.

Report definitions last checked 6 August 2026

Google’s dedicated generative AI reporting is being rolled out to a subset of Search Console properties, while Bing AI Performance remains in public preview. The available dimensions and definitions are not equivalent. Revalidate official documentation and the actual property interface immediately before publication and before each material reporting decision.

Use one governed dashboard schema

For each evidence source, record the metric definition, date range, filters, page cohort, onsite outcome, commercial outcome, limitation, owner and next decision. If AI impressions rise for a service-page cohort while overall organic clicks remain flat, first check rollout and definitions, then filters, canonicals, page changes, demand, onsite actions, CRM outcomes and the change log. Do not infer causation from the first movement.

Create a baseline and change log

Capture the first available data before making a large content change. Record property access, report rollout date, date range, included experiences, leading pages, country and device mix, export settings and known limitations. Preserve the raw export where governance permits. A baseline is not a benchmark against other brands; it is the organisation’s starting evidence under a documented platform definition.

Maintain a change log for content releases, internal-link changes, technical incidents, migrations, major campaigns, product launches, brand events and platform report updates. Search demand and AI-feature availability can move without a website change. Use annotations or a separate register so the team can identify plausible explanations without claiming causation. If Google changes inclusion or aggregation, begin a new comparable period rather than joining incompatible data silently.

Diagnose movement before recommending content

When impressions change, begin with measurement conditions. Confirm that property access, filters, date ranges, time granularity, canonical assignments and included surfaces are unchanged. Check whether the platform expanded rollout, altered aggregation or added an experience. Then review demand, country and device patterns, affected pages, indexing, major brand events and site releases. Only after those checks should the team infer that a content change is a likely driver. A dashboard alert should open an investigation, not automatically open a writing brief. Outdated exposed pages may need the content decay decision framework.

Classify the response by cause and confidence. A technical eligibility problem may require crawling, indexing or snippet-control work. A strong page losing visibility after becoming outdated may need factual maintenance. A new page with no first-party exposure may simply need time, internal discovery or a clearer customer purpose. An executive request to appear for one manually tested prompt may not justify any change if the prompt is unrepresentative. Record the hypothesis, evidence, action and review point so later results are interpreted against what the team actually changed.

Set escalation rules for material factual misrepresentation or harmful cited content. Preserve the prompt, answer, date, platform, account conditions and cited sources where policy permits. Check the organisation’s own page first, then identify whether the issue comes from stale content, ambiguity, another source or system behaviour outside the organisation’s control. Correct owned information through the normal editorial process and seek qualified legal or communications advice where risk warrants it. SEO reporting does not guarantee correction in an external AI answer.

Connect visibility to onsite outcomes

Start with the pages exposed by first-party reports. Review their overall organic landing sessions, engagement, navigation to service or product content, downloads, enquiries, purchases and other qualified actions. Use existing analytics definitions and consent controls. Do not create an artificial AI conversion channel unless a platform supplies a defensible source signal. A page can gain AI impressions without clicks, and a visitor can arrive through ordinary Search after earlier AI exposure. Any CRM connection must respect first-party data and consent.

Connect qualified web actions to CRM or commerce outcomes where legitimate identifiers and processes exist. Evaluate lead quality, pipeline progress, order value or repeat behaviour at an appropriate cohort level. Keep attribution language cautious: pages associated with increased AI impressions also produced certain onsite outcomes. That is different from saying the AI feature caused the outcomes. Analytics implementation, CRM integration and reporting design are separately scoped paid services.

Use prompt research as qualitative evidence

Manual prompt research remains useful for seeing answer presentation, checking factual representation and identifying which sources appear in a controlled set of decision questions. Define the market, language, device or account state where relevant, prompt set, date, platform and repetition method. Capture cited sources and material inaccuracies. Use the exercise to generate hypotheses for content and reputation work, not to estimate population-wide share.

Responses can vary by wording, context, location, model and time. Personalisation and experimentation may also affect the result. A third-party tool can provide consistent monitoring at a larger scale, but its prompt universe, geography, frequency, account state and citation detection still define the result. Review methodology, exportability and coverage before adopting a score. Never describe a vendor index as official Google or Microsoft market share.

Each layer retains its own definition, date range and limitation.

Report a range of evidence, not one score

• Google supported-feature impressions, leading pages, country and device trends under the current definition.
• Bing citations, cited pages and sampled grounding phrases under the public-preview definition.
• Organic landing-page behaviour and qualified actions for the relevant content cohort.
• CRM or commerce outcomes with identity and attribution limitations stated.
• Qualitative prompt observations, factual issues and cited-source patterns from a documented sample.
• Content releases, technical changes and platform-report changes that affect interpretation.

Use the dashboard to make bounded decisions. Improve a page when first-party evidence and customer needs show a content gap. Repair technical eligibility when indexing or snippet controls are wrong. Strengthen proof when answer observations expose ambiguity. Continue monitoring when a change is too recent to interpret. Do not publish dozens of near-duplicate pages simply because a third-party prompt list exposes variations; Google says existing SEO fundamentals remain relevant and no special AI markup is required.

Set an ownership and reporting cadence

Name the owner of each evidence layer. SEO should own platform-definition monitoring and search interpretation. Analytics should own website-event quality. CRM or commerce owners should define qualified outcomes. Content owners should maintain factual pages and source records. Communications or legal specialists may be required when observed answers create reputational or regulatory risk. One dashboard owner can coordinate the view, but that person should not silently redefine another team’s measures. Access to Search Console and exports should follow client security and governance requirements. Emote can scope search engine optimisation around a dated, governed evidence baseline.

Choose a cadence that matches data maturity. Weekly monitoring may identify report outages or abrupt page changes; monthly reviews may support content and technical prioritisation; quarterly reviews can assess broader themes against pipeline or commerce outcomes. Avoid daily executive reactions to preliminary data, small samples or a single manual prompt. Escalate only when the evidence meets an agreed threshold, and preserve a review date for inconclusive changes. Reporting, prompt research, dashboard development and ongoing SEO monitoring are paid activities requiring an explicit scope.

A trend is only interpretable when the measurement boundary and major changes are recorded.

Frequently asked questions

Does Search Console show AI Overview clicks?

The dedicated generative AI report documented at the research date focuses on impressions and page, country, date and device dimensions. AI-feature activity also sits within overall Web performance. Recheck current documentation before publication.

Why can a property not see the generative AI report?

Google currently lists limited rollout, insufficient impressions and eligibility controls as possible reasons. Missing access does not establish that the site has never appeared in an AI feature.

Can AI-search visibility be reduced to one percentage?

Not credibly across the whole market with current data. Any share or score reflects a defined platform, prompt universe, surface and method. Present the boundary beside the number.

Are Google and Bing AI reports directly comparable?

No. Google currently reports supported-feature impressions, while Bing’s preview uses citation measures and sampled grounding phrases. Keep each source in a separate panel with its own definitions.

Does appearing in an AI answer guarantee traffic?

No. Visibility may support awareness or later behaviour without a measured click, and an impression does not prove attention. Track onsite and commercial outcomes separately.

Should a business buy a third-party AI visibility tool?

Consider it when the methodology, prompt coverage, geography, frequency, exports and decision use justify the cost. It should complement first-party evidence, not be presented as platform-internal truth.

How Emote can help

Emote can help establish a practical AI-search measurement baseline that connects available search data, content visibility and meaningful business outcomes.

The smallest credible pathway uses first-party and webmaster evidence before adding speculative tools or new reporting layers. Paid Full Website Discovery is reserved for complex measurement, data or governance requirements.

If you want to define what useful AI-search visibility measurement looks like for your organisation, book a meeting with Emote.