Google AI reporting: Search Console vs Bing

The short answer

Google Search Console and Bing Webmaster Tools now expose useful AI-specific evidence, but they do not expose the same event. Google's dedicated generative AI report shows aggregate impressions for links from your site in AI Overviews and AI Mode. Bing AI Performance shows aggregated visible citation activity across Microsoft Copilot, AI-generated summaries in Bing, and select partner AI integrations. Neither report is a replay of an individual answer.^1,2,5,6

Use the consoles as aggregate first-party evidence, then keep comparable answer monitoring and web analytics in separate rows. The consoles tell you whether exposure or citation activity exists at scale on their covered surfaces; a captured answer tells you exactly what one named surface returned for one question; analytics tells you whether a measurable visit followed. The wider surface map explains why those boundaries matter.

An impression does not prove a citation. A Bing citation count is not a click. One captured answer is an observation, not a trend. Traffic is a separate outcome. Those are not disclaimers around one metric; they are four different measurements that can corroborate one another without being merged.

Four evidence layers, four different claims

Start by writing the claim each row can support before looking at whether its number rose or fell. Robot Visible's evidence methodology uses the same discipline: preserve eligibility signals, competitive context, and observed answers as separate states, and never turn an unchecked surface or missing dataset into a negative result.

What Google's generative AI report actually shows

Google launched the dedicated Search generative AI performance report in June 2026 and is rolling it out to a subset of site owners. The Search report includes AI Overviews and AI Mode, excludes experiments in Search Labs, and sits inside the Web search type of the ordinary Search results performance data. Not seeing the report can mean the property has not received access, has too little qualifying activity, or is excluded from the covered features; absence of the report is not evidence that the site never appeared.^1,2

The report's metric is impressions: how many times a link to the property was shown to a user in a supported generative AI feature. Its dimensions are Pages, Countries, Devices, and Dates. There is no query dimension, no answer transcript, and no citation-position field. Google may add metrics later, but the current report cannot tell you which question caused an impression or reproduce the surrounding answer.^1,2

The chart is aggregated by property, so two results from the same site in one generative AI feature count as one property impression. Add a URL filter and the chart becomes URL-aggregated; group the table by page and its rows are page-aggregated. That is why chart and table totals can differ without either being wrong. The usual Search performance limits, including the 1,000-row interface limit, also apply.¹

Treat the page dimension as a discovery list, not a ranking. It tells you which canonical URLs received covered impressions and gives you candidates for exact-question checks. It does not reveal whether the page supplied a factual passage, appeared as a supporting link, or was one of several links from the same property. Use an observed answer record when the claim needs that precision.

How Google counts an AI impression

AI Mode uses Google's standard impression rules. AI Overviews are narrower: a link must be scrolled or expanded into view before its impression is counted, and every link inside the overview receives the position assigned to the overview as a whole. A click is recorded only when the user follows an external link. A follow-up in AI Mode is treated as a new query and its resulting impressions and clicks are counted separately.³

That definition explains why impressions without clicks are neither success nor failure by themselves. The site earned visible exposure under Google's counting rules; the reader did not produce a recorded outbound click. The answer may have satisfied the need, the link may have been peripheral, the result may have appeared below the most useful passage, or the impression volume may simply be too small for a click yet.

Do not divide the dedicated report's impressions by clicks from a different report and label the result an AI click-through rate. The numerator and denominator do not necessarily describe the same population. Export the dedicated view as its own series, keep its scope in the column name, and only calculate a rate when Google exposes two metrics from the same filtered dataset.

The export and API boundary

The dedicated Google report has an export button for its chart and table. Values displayed as a tilde or dash become zero in the downloaded file, so preserve the raw export and document that conversion before treating zero as a measured absence. The newest points may also be preliminary and can change while Google finishes processing them.¹

Google's documented Search Analytics API exposes standard search types and dimensions such as query, page, country, device, date, and search appearance. It does not currently document a dedicated generative AI report endpoint or filter. The dedicated UI report also has no query dimension. A standard Search Console API connection can still bring valuable Web performance and URL inspection evidence into a product, but it should not be described as automatically reproducing the separate generative AI report unless Google documents that access path.⁴

For Robot Visible, that means connected Search Console evidence remains useful for demand, indexing, and page-level organic performance while the dedicated AI export stays a separately scoped input. The integrations page describes the connection that exists today; it does not imply access to an undocumented report API.

What Bing AI Performance actually shows

Bing AI Performance is a different dataset, not Microsoft's version of Google's impression report. It reports visible citation activity across Microsoft Copilot, AI-generated summaries in Bing, and select partner AI integrations. Its core views include total citations, cited pages, average cited pages, page-level citation activity, grounding queries, and trends over time.^5,6

Bing defines total citations as the number of times content was visibly referenced or shown as a source during the selected range. Cited pages counts unique pages cited on a day, while average cited pages averages that daily count across the range. Those metrics describe citation volume and coverage. Microsoft explicitly says they do not measure rankings, authority, importance, placement, traffic, or a page's role inside an individual answer.^5,6

The data is refreshed daily with a processing delay and is designed for trends rather than exact accounting. Microsoft says it is aggregated, representative, and a sample of overall citation activity: low-frequency citations may not surface, totals can differ between views, and the displayed history is limited to the ranges available in the dashboard. Exports are available in CSV and Excel for page, grounding-query, and time-series analysis.⁵

A Bing citation is stronger evidence of source use than a Google impression because Microsoft is reporting visible citation activity. It is still aggregate evidence. If you need to say what the answer claimed, how your page was framed, or which other sources appeared beside it, you need a captured answer with the provider, surface, question, URL, and time preserved.

Grounding queries are grouped phrases, not prompts

Bing's grounding queries are grouped phrases associated with retrieved content that was cited. They are not full user prompts and do not expose an individual user's question or answer. One grounding-query group can map to several pages, and one page can map to several groups. The wording can also vary across Microsoft surfaces and partners.⁵

The count beside a grounding query reflects how often the site was referenced in answers associated with that group. It does not identify the exact prompt, the reason a page was selected, or the page's role in an individual answer. Sparse activity may be withheld, and filtering from a query to a page can produce a different sampled count from filtering from the page back to the same query.⁵

Use grounding queries for topic discovery and grouping. Use them to find recurring concepts, locate pages with citation traction, and design a fixed monitoring set. Do not paste a grouped phrase into a slide and call it the question customers asked. That changes summary-level metadata into evidence Microsoft explicitly did not provide.

A reconciliation workflow that keeps the evidence honest

Reconciliation does not mean forcing the numbers to match. It means aligning dates, pages, and covered surfaces so each source can answer its own question. Start with a stable reporting window and preserve every raw export before normalising URLs or aggregating rows. The public measurement methodology uses the same rule: record the source and transformation before publishing a comparison.

• Export Google's generative AI report by date and page. Record the property, date range, export time, aggregation choice, and whether the newest data was preliminary.
• Export Bing AI Performance for the same dates. Keep total citations, cited pages, grounding-query groups, page mappings, and the dashboard's sampling limits in the source notes.
• Canonicalise page URLs cautiously. Google assigns most data to Google's selected canonical; Bing may report another public URL. Keep both originals before joining them.
• Run a fixed set of buyer questions on named answer surfaces. Preserve the exact question, captured answer, provider, model, surface, cited URL, locale where applicable, and timestamp.
• Join by page and time window, not by pretending Google's absent query dimension equals Bing's grouped grounding phrase. A page is the strongest shared key the two consoles currently provide.
• Add analytics last. Report attributable visits and conversions as outcomes after exposure or citation, while acknowledging that many impressions and citations produce no measurable referral.
• Annotate content releases, indexing changes, outages, report roll-outs, and known data anomalies. A before-and-after movement does not establish causality when several things changed.

How to read common combinations

The minimum record behind a defensible report

A chart without scope is almost impossible to audit later. Keep the fields below beside every exported series or captured answer, even if the reporting deck shows only a few of them. They are what let another person reproduce the join, distinguish a report change from a site change, and correct the record without rebuilding the analysis from memory.

What Robot Visible can combine today

Robot Visible can connect authorised Google Search Console and Bing Webmaster properties for search and indexing evidence, and it separately records comparable answer checks with their questions, providers, surfaces, models, cited URLs, and timestamps. Those records belong together in an investigation, but the product does not relabel standard Search Console metrics as Google's dedicated generative AI report or claim that a provider integration saw an undocumented dimension.

The practical next step is page-led. Use the console reports to identify pages with new exposure or citation activity, inspect those pages for the exact claims they can support, add representative buyer questions to monitoring, and compare later answers without interpreting a provider error or an unchecked surface as a loss. The crawler reference is the companion check when a page has no evidence at all: establish access before diagnosing competitiveness.

Run a free exact-URL check

How this guide is maintained

The Google and Bing report facts on this page were checked against first-party documentation on 2026-08-10. A content test expires the review after 45 days so a renamed dimension, new API, changed roll-out, or revised sampling statement cannot sit here indefinitely under a recent page date.

Freshness is not certainty. A passing date proves only that the documentation was reviewed recently. Product access can still vary by property, preview cohort, region, or account; report values can remain preliminary; and provider definitions can change between reviews. The source list below is part of the page's evidence, not decoration.