Google Search Console vs. Bing Webmaster Tools: What Their AI Reports Actually Measure

Google Search Console and Bing Webmaster Tools do not publish two versions of one universal AI-visibility number. They observe events on different surfaces, counted under different rules, with different denominators - and neither provider publishes a mapping between them. Google's generative AI performance reports count impressions: how often links to your site were shown in its generative AI features. Bing's AI Performance report counts displayed citations: how often your content was visibly referenced or shown as a source in an AI-generated answer across Microsoft's supported surfaces, plus grouped grounding-query phrases, page-level citation counts, classified intents and topics, a per-query Citation Share, and a period-over-period overlay.

Both are worth having. Neither is a scorecard you can put next to the other one.

The short answer

Use each report for the event it observes, and keep the two totals in separate columns. If the question is which of our pages appeared in Google's generative AI features, and how that broke down by country, by device or by date, Search Console answers it directly. If the question is which of our pages were displayed as sources in Copilot and Bing's AI answers, under which grouped query themes, and how that has moved, Bing Webmaster Tools answers it directly and Search Console does not address it.

What neither answers is the question most B2B SaaS teams actually bring: are we being recommended, and is it producing pipeline? Both reports stop several joints short of that, and the gap is not a reporting oversight - it is what these reports were built to observe.

One terminological note, because the rest of the page depends on it. Throughout CoreAEX's work, a citation is a page or URL presented to the user as visible source attribution for an answer or claim. That is distinct from a mention (the brand is named), a recommendation (the brand is advanced as a suggested option), retrieval (a system accessed the page while generating), an impression (a link was shown, as the platform counts it), a referral (a session arrived), and influence on generated wording (the answer's text was shaped by the page). A citation does not by itself establish that a particular passage was used, or that the cited page caused anything in the answer. The fuller treatment of those distinctions is in AEO Fundamentals: How to Get Cited, Not Just Ranked; this page assumes them.

We also avoid "visibility" as a free-floating outcome here. Where the word appears, it is either a provider's product name - Microsoft's "AI Visibility Insights" - or it is immediately replaced by the observable event: a URL impression, a displayed citation, a cited page, a grounding query, a classified intent, a topic, a Citation Share figure, or a referral.

Two labels mark evidence boundaries on this page. Documented marks a platform definition or behaviour directly supported by the linked provider help documentation; surrounding interpretation remains CoreAEX analysis. Recommendation means a workflow or reporting rule CoreAEX prescribes and no source specifies. Product announcements on Google's and Microsoft's blogs are first-party but are not help documentation, so they are attributed in the prose with their dates rather than tagged.

Side by side

Every cell below is written to stand on its own, because these are exactly the rows that get copied into a slide and lose their qualifiers.

What Search Console measures

Google announced the dedicated Search Generative AI performance reports on its Search Central blog on Wednesday, June 3, 2026, in a post by Hillel Maoz and Moshe Samet. The announcement described dedicated views for Search and for Discover, listed impressions, pages, countries, devices and dates as the information the reports show, noted that devices are available for Search results, and said the reports were rolling out to a subset of websites. The detail below is from the two help pages the announcement points to: the generative AI performance report for Search and the generative AI performance report for Discover.

Availability, which the documentation leaves unresolved

Google says the insights were rolled out to all websites worldwide on August 31, 2026 - the sentence appears as a note on the June announcement and on both help pages. However, both current help pages still list phased property access among the reasons a report may not appear: "Not all properties have access to the report, as we're rolling out over time." Insufficient impressions in the relevant generative AI features is listed as a second possible reason. Documented

That may be stale help copy, an access qualification or a remaining rollout condition; the public documentation does not resolve the inconsistency, and this page does not resolve it on Google's behalf. The practical reading is narrower than either sentence alone: if the report is present in your property, use it; if it is absent, the documentation gives you two candidate explanations and no way to distinguish them. Do not tell a stakeholder the report is guaranteed available, and do not tell them its absence means your site has no impressions.

This also matters for older guidance. Measurement pages written before September 2026 - including earlier CoreAEX pages - commonly quote the original subset-rollout wording alone. Check the live help pages before repeating any availability statement.

The impression, exactly as documented

The two reports define the event differently, and the differences are not cosmetic. Documented

• Search. Impressions are how many times links to your site were shown to a user in a generative AI feature on Google Search. The help documentation names AI Overviews and AI Mode.
• Discover. Impressions are how many links to your site a user saw in generative AI features in Discover. Two conditions apply that the Search report does not carry: the link must be scrolled into view, and only one impression is counted per result per session - if a user scrolls past a card and then scrolls back, only one impression is recorded.

The two Pages dimensions are also described differently, and this is the part of Google's documentation that bears most directly on what an impression represents. For Search, the Pages dimension groups data by the final URL linked by a generative AI feature after any redirects. For Discover, it is "the page that served as the source of the information shown to the user" - the canonical page URL, not the page the user lands on when they click. Documented

Google's June announcement used a broader phrasing for both reports - how often URLs from your site appeared in generative AI features in Search and Discover. The help documentation is narrower and more useful, because "shown to a user" and "scrolled into view" are observable display conditions rather than an unspecified appearance. Where the two wordings differ, the help documentation is the one to quote.

Impressions and citations are different metrics, not opposite events

It is tempting to summarise the two platforms as "Google shows links, Bing cites sources." That is too clean, and the Discover wording above is why. A page described as the source of the information shown to the user, presented to that user as a link, may well satisfy CoreAEX's definition of a citation: a page or URL presented to the user as visible source attribution for an answer or claim. Some Google appearances plausibly are citations in that sense.

What follows is narrower, and it is what the documentation supports:

• Google does not provide a separately defined citation metric. Impressions are the only metric named in either report's current documentation.
• Google documents no link-to-claim attribution relationship - nothing that would let you say which sentence, if any, a shown link was attributed to.
• Microsoft defines its metric around content visibly referenced or shown as a source, and counts those references.
• So an impression count cannot be converted into a claim-level citation count, in either direction, and the two totals are incompatible.

Incompatible is not the same as mutually exclusive. The same interface event can be an impression to Google's counter and would be a citation under CoreAEX's definition. What you cannot do is treat Google's total as a citation total, or assume the difference between the two platforms' numbers reflects a difference in how each system attributes sources.

Aggregation: two different rule sets

There is no single "Google aggregation rule." The Search and Discover reports count and group differently, and reading one report's rule onto the other will produce a wrong cohort figure with no error message. Documented

Search report. Data on the chart is aggregated by property: if two results from the same site appeared in one generative AI search results feature, they count as a single impression in the chart total. Adding a URL filter switches the chart to aggregation by URL. In the table, data grouped by country, device or date is aggregated by property, while data grouped by page is aggregated by page. Google notes that chart totals can differ from table totals, usually because of this property-versus-page difference.

Discover report. All data is aggregated by page. If two generative AI results from the same property appear in the same Discover list, each impression is counted separately - the opposite of the Search chart's deduplication. Google also states that total impressions are accurate for the selected time span and are not truncated, and that daily values may be omitted from the chart if they do not reach a threshold while still being included in the totals.

Both reports inherit the usual data limitations of their parent performance report, including the 1,000-row limitation, and Search Console excludes data from Search Labs experiments because those are still in active development.

Two operational consequences. State which report and which aggregation produced any figure you publish - a Search chart figure and a Search page-table figure for the same cohort are different measurements, and a Discover figure is on a third basis again. And never compare a Search chart total to a Discover total as though the difference were behavioural. Recommendation

Four absence and zero cases

These are the most avoidable reporting errors in either platform, and Google's export behaviour makes one of them easy to commit. Three separate documented behaviours are involved, and they are frequently collapsed into one. Both help pages state that values shown as either ~ or - in the report - not available, or not a number - will be zeros in the downloaded data. The Discover page separately states that daily values may be omitted from the chart if they do not reach a threshold while still being included in the totals. And both reports carry the usual 1,000-row limitation on their tables. Documented

So four different things can present as an absence or a zero, and they are not interchangeable:

1. A row absent from a table subject to the 1,000-row limit. Nothing was measured as zero; the row is simply outside what the table holds.
2. A daily Discover value omitted from the chart because it did not meet a threshold. Google states it remains included in the total, so the value exists and is counted - it is the chart display that drops it. The help pages do not state whether such a value appears as a row in an export, so do not assume either way.
3. A ~ or - shown in the interface and converted to zero in an export. The displayed value means not available or not a number; the zero is created by the download, not by an observation.
4. A value that can independently be established as a measured zero, where something other than the export's own conversion supports reading it as no impressions.

Only the fourth case supports the statement that the measured value was zero. An absent table row may reflect the table limit; a missing daily Discover chart point may reflect thresholding even though it remains in the total; and an exported zero may be a converted ~ or -. These three are checked differently, so check the relevant interface view and aggregation level before interpreting any of them. An overall chart total cannot validate a zero for an individual page. Recommendation

What the data has already been wrong about

Google publishes a data-anomalies page, and the generative AI reports are on it. Read on September 7, 2026, it records a logging error that decreased impressions on the Generative AI performance report in Search for data from August 13 to August 17, 2026, with a subsequent note stating that the missing data for that period has been restored; it also records logging errors affecting Discover and generative AI in Discover on June 24, 2026 and August 13, 2026. Google classifies these as affecting data logging rather than actual performance. Documented

The durable practice, not the specific incident: check the anomalies page for your analysis window before interpreting a movement, and re-check it afterwards, because entries are amended. The August entry is the worked example - a report written in the third week of August would have described five days of depressed impressions that were subsequently restored. Recommendation

What the Search Console reports can and cannot establish

Can establish, directly: that links to specified URLs from your property were shown in Google's generative AI features during a date range, with the breakdowns each report documents - pages, countries, devices and dates for Search; pages, countries and dates for Discover - subject to that report's aggregation rules and the row limit.

Cannot establish: which query or prompt produced the appearance, whether the appearance was in AI Overviews or AI Mode, whether a shown link was attributed to a specific claim, or whether the user read, clicked or acted. The current public documentation describes no click, click-through-rate or position metric for these reports, and no query dimension. Google's separate AI-features guidance states that traffic from AI features is included in the overall Search Console figures and reported within the Web search type of the Performance report, so the generative AI reports are a dedicated view of impressions rather than a separate traffic stream to be added to your Search totals.

One forward-looking note, attributed rather than assumed: Google's June announcement said it is continuing to work with website owners on what insights would be most helpful, "such as adding additional metrics over time." That is a statement of intent, not a roadmap, and nothing should be planned against it.

What Bing Webmaster Tools measures

Microsoft announced the AI Performance report on the Bing Webmaster blog on February 10, 2026, as a public preview. The current help documentation describes the report as showing how your site's content is used in AI-generated answers across Microsoft Copilot and partner experiences, and covers three supported surfaces: Microsoft Copilot, AI-generated summaries in Bing, and select partner AI integrations. That is the boundary of the report's scope. It says nothing about ChatGPT, Perplexity, Claude, Gemini or any other system, and a number from it cannot be generalised to them.

The metrics, as documented

The help documentation defines the counted events as follows. Documented

• Total Citations - the total number of times your content was visibly referenced or shown as a source in AI-generated answers during the selected date range. The February announcement's wording was the total number of citations displayed as sources in AI-generated answers during the selected time frame.
• Cited pages - the number of unique pages from your site cited on a given day.
• Average Cited Pages - the average number of unique pages cited per day over the selected range.
• Page-level citation activity - citation counts by specific URL, showing which pages are cited most frequently. The documentation attaches its own qualifier: this view reflects how often pages are cited, not their importance, ranking, or role within a response.
• Grounding queries - the key phrases the AI used when retrieving content that was cited in its answer, shown with the number of times content from your site was referenced for that phrase.
• Timeline - citation volume over time, across 7-day, 30-day, 3-month or custom ranges, refreshed daily after a short processing delay.

The visible-reference condition in the Total Citations definition is what makes this metric line up cleanly with CoreAEX's definition of a citation: Microsoft is counting something shown to the user as a source. It still does not establish which passage was used. The documentation is explicit that the data does not show why a specific page was referenced.

Grounding queries are not prompts, and the documentation says so

This is the point most likely to be misused in a B2B SaaS reporting deck, because the rows look like keywords. Microsoft's help documentation states that each row displays a short phrase, that these are grouped representations, and that they are not full user questions or prompts. It adds that grounding queries reflect aggregated activity across AI-generated answers rather than individual user questions or specific answers; that a single grounding query can be associated with multiple pages and a single page with multiple grounding queries; that phrasing may be determined differently across supported AI experiences and partners; and that the data is summary-level and does not show individual AI answers or exact prompts. Its own worked example is a site cited in answers about solar energy appearing under phrases such as "solar energy" or "solar panels." Documented

So grounding queries are useful as a read on the thematic areas where your content is being pulled into answers. They are not demand data, not keyword volume, and not a record of what buyers asked. A grounding query with a high count tells you your content was referenced often in answers associated with that phrase - it does not tell you how many people asked anything.

Sampling, and why two views can disagree

Microsoft is unusually direct about the incompleteness of the data, and these statements govern how every figure in the report should be read. The documentation states that AI Performance data is aggregated and summarised to provide a representative view, that it is not a complete log of every instance where content may have been referenced, that not all citation activity may appear, that very low or infrequent citation activity may not surface, that totals may differ across views, and that the data shown represents a sample of overall citation activity which may be refined as more data is processed. It also explains a specific artefact: because grounding queries and pages may each be sampled over slightly different windows inside your selected range, filtering by a grounding query and then viewing a page can show a different count than filtering by that page and viewing the same grounding query - which the documentation states is expected and does not indicate missing or incorrect data. Coverage is bounded in one further way: as with Bing search, AI Performance reflects only content eligible for indexing, and Bing respects robots.txt and other supported control mechanisms. Documented

Three consequences worth writing into your own reporting rules. Do not present a Bing citation total as a complete count. Do not treat an absent page or an absent grounding query as a zero - the documentation gives sparse activity as an expected reason for non-appearance, and adds that this does not indicate a penalty or exclusion. And when two views disagree, state which view produced the figure rather than reconciling them. Recommendation

What a citation count does not establish

Microsoft answers this in its own documentation more plainly than most third-party writing does. AI Performance does not measure rankings, authority, performance or importance; it shows which content was cited. Citations do not represent traffic, clicks or user engagement. Page-level counts do not indicate a page's role within an individual answer. And on the timeline specifically: trends reflect aggregated citation activity only and cannot be attributed to a specific cause or event; changes are observational and do not indicate the impact of any single update, model change or content modification. Microsoft lists shifts in the volume or type of user questions, content updates, and system or model updates across AI experiences among the things that can move the line. Documented

Preview status, stated carefully

Microsoft announced the report as a public preview on February 10, 2026. The current help documentation describes Intents, Topics, Citation Share and Compare as four preview capabilities, and the June 16, 2026 announcement said they were beginning to roll out in preview globally. The help documentation read on September 7, 2026 does not restate a preview label for the core report, and no announcement stating that the core report has left public preview was located on Microsoft's blogs. Treat the status as unresolved rather than settled in either direction, and check the interface before describing availability to a stakeholder.

Intents, Topics, Citation Share and Compare

The June 16, 2026 Bing announcement added four capabilities. Three of them operate on grounding-query data and one does not, which matters when you are deciding how much interpretation sits between you and the number:

• Intents classify the context associated with a grounding query.
• Topics group grounding queries into broader themes.
• Citation Share uses a grounding-query-specific denominator.
• Compare provides a period-over-period view of reported citation activity. The current help documentation describes it as a charting and comparison feature and does not describe it as a classification layer over grounding queries.

Intents

Intents classify grounding queries by query intent. Microsoft's documentation lists a taxonomy including Informational, Media, Navigational, Commercial, Learn and Solve, Research, Live Event, Local, Comparison, Planning, Utility, Creation, Conversational and Other, describes it as a modern intent taxonomy going beyond the traditional informational/navigational/transactional split, and states that labels are assigned by AI/ML classifiers that analyse the grounding query and its context. On accuracy, it states that the classifiers are continuously improving and that a grounding query may be assigned a label that does not perfectly match how you would characterise it, particularly for ambiguous or multi-intent queries. Documented

For a B2B SaaS site, the practical use is the shape of the distribution rather than any single label. A vendor whose citations sit almost entirely under Informational, when the commercial case rests on being present for Comparison, has learned something about where its content is being pulled in. It has not learned why, and the classifier's own uncertainty sits between the observation and the conclusion.

Topics

Topics group related grounding queries into broader thematic clusters - the documentation's examples are Solar Energy, Home Renovation and Cloud Security - and sit one level above grounding queries. They are also assigned by AI/ML classifiers. Microsoft states that a grounding query may be grouped into a topic that feels like an imperfect fit, especially for niche or highly specialised domains, and that it expects the quality and precision of Topics to improve as the classifiers mature. Filtering by a topic to see cited pages is done by exporting the data and applying filters rather than in the interface. Documented

The niche-domain caveat is not a footnote for this audience. B2B SaaS categories are frequently exactly the kind of specialised domain Microsoft names, so a thin or oddly-labelled topic view is a documented expectation for a narrow category rather than a finding about the category.

Citation Share

Citation Share is the only relative metric documented across the two reports reviewed here, and its denominator is the whole point. Microsoft defines it as the percentage of citations attributed to your site out of all citations shown across all sites for that same grounding query. The denominator is one grounding query's citation space, not a market, not a topic, and not the AI-search ecosystem. A site can hold a high total citation count and a lower Citation Share when citation activity for the same query is distributed across many sources. Documented

Microsoft is explicit about what the metric is not. The June announcement describes it as an observational metric, not a ranking system or a competitive scoreboard, and states that it does not expose competitor domains, represent traffic share, or assign quality scores to content. The help documentation repeats that it reflects citation activity and does not represent rankings, traffic or a quality score, and lists shifts in user demand, content changes across the web, model updates, freshness signals and partner refresh cycles among the drivers of a change in share. On the question every content team will ask - can this tell us whether our content changes worked - the documentation answers directly: Citation Share can help you observe whether your share changes over time after content updates, but it does not establish a causal link.

Two errors to keep out of a board deck. Citation Share is not share of voice across AI search; it is share of one grounding query's displayed citations on Microsoft's supported surfaces. And because the underlying data is sampled, a share figure carries the sampling caveat as well as the denominator caveat - report both alongside the number. Recommendation

Compare

Compare overlays a previous time period on the same chart as the current view, with the current period as a solid line and the comparison period dashed, day-aligned on the same axis. Comparison options include the previous period at various durations and a custom option for two manually chosen ranges. Microsoft states that Compare shows what changed between two periods but does not explain why the change happened, and that it helps you observe patterns rather than establish causation. Documented

That is the correct reading and it is worth holding to even when the overlay is flattering. A dashed line below a solid one after a content release is a movement that coincided with the release. It is not the release's effect, and the provider that built the feature says so.

Why the headline numbers can't be compared

The temptation is a single slide with two numbers on it. Here is the full list of reasons that slide is not a comparison, in the order they bite.

1. The counted events are defined differently. Google counts links to your site shown in its generative AI features. Microsoft counts times your content was visibly referenced or shown as a source in an AI-generated answer. Neither provider defines its event in terms of the other's, and neither publishes a conversion between them. Some individual appearances may satisfy both descriptions; the totals are still built on different definitions.
2. The surfaces do not overlap. Google's reports cover AI Overviews, AI Mode and generative AI features in Discover. Bing's report covers Copilot, Bing's AI summaries and select partner AI integrations. There is no surface in common, so the two totals are not two views of one population.
3. The aggregation rules differ - including between Google's own two reports. Google's Search chart deduplicates to property level, its Search page table aggregates by page, and its Discover report aggregates everything by page while counting each of two results from the same property in one list separately. Bing's Total Citations counts references. A Google Search chart figure is structurally lower than a count of individual appearances; a Discover figure and a Bing citation total are not built that way.
4. The denominators differ. Google's reports expose no relative metric to divide by; impressions are absolute counts. Bing's only relative figure divides by the citations shown for a single grounding query. There is no common base.
5. Only one provider characterises its data as sampled. Microsoft states plainly that its data is a sample and not a complete log. Google documents aggregation, thresholds, row limits and the exclusion of Search Labs data, but does not characterise the underlying impression data as sampled or unsampled. That is an asymmetry in what is documented, not a demonstrated difference in completeness - and it means the two figures differ in a way neither provider quantifies.
6. Bing's context data is classified; Google's is undocumented. Intents and Topics are machine-assigned labels over sampled grounding queries, with the provider's own accuracy caveat attached. Google's current public documentation describes no query-side context to classify. So any "by intent" or "by theme" comparison exists on one side only.
7. Neither platform represents the AI-search market. Google's reports observe Google's own generative features. Bing's report observes Microsoft's supported surfaces, including select partner AI integrations whose full list Microsoft does not publish. Neither report covers the entire AI-search environment. Buyers may use the Google and Microsoft surfaces included here, but they may also use ChatGPT, Perplexity, Claude, Gemini surfaces outside these reports, or other systems the two reports do not document.
8. A larger number in one report is not stronger performance than a smaller number in the other. Given all of the above, the direction of the inequality carries no information about relative performance.

A hypothetical illustration, with invented figures. Suppose a site's Search Console generative AI report for Search shows 4,000 impressions over 30 days, and its Bing AI Performance report shows 1,200 total citations for the same window. The 4,000 is a property-aggregated count of links shown across AI Overviews and AI Mode; the 1,200 is a sampled count of visible source references across Copilot, Bing AI summaries and partner surfaces. The site could be appearing more often on Microsoft's surfaces than Google's, or less, or comparably - these two figures cannot tell you, and a ratio between them has no referent. No CoreAEX or provider data is being reported here. Recommendation

What each report can answer

Two buckets are worth separating before you open either tool: what the report answers directly, and what it can only help you investigate. A third category - what neither report establishes - is treated in detail in the states table further down, and is summarised here rather than repeated.

Answered directly by the report

May help investigate, but does not settle

• Whether a content change worked. Both reports can show movement around a release date. Neither isolates it - Microsoft says so explicitly for Compare and Citation Share, and Google's reports offer no comparison design at all. Treat movement as a prompt to look, not as a result.
• Which themes we are under-covered on. Bing's Topics and Intents give a directional read, carrying the classifier caveat and the niche-domain caveat. Use them to generate hypotheses for content planning, not to size an opportunity.
• Whether a technical change restored eligibility. Bing's documentation ties coverage to indexing eligibility and robots.txt handling, so a return of citation activity after a crawl fix is consistent with the fix having worked - alongside several other explanations.
• Whether our best pages are the ones being used. Both reports list or order pages using their respective reported metrics. Whether that ordering matches commercial value is a judgement the reports cannot make.

Not established by either report

In short: whether you were mentioned, recommended, clicked, referred, or influential over the wording of an answer or a buying decision; total demand for a topic or prompt; and anything happening on an assistant outside each provider's documented scope. The states table sets out each of these with the specific provider wording that supports it.

When corroborating evidence is required: whenever the business question crosses from display to behaviour, or to a system neither provider covers. That means referral analysis and analytics for arrival, server logs for crawler access, your own defined prompt panel for engines outside these two, and CRM data for anything commercial - each with its own denominator problem, which is a separate subject and not this page's.

Using both without blending them

A seven-step routine that keeps the two data sets in their own lanes. It is deliberately short; the discipline is in steps one and three. Recommendation

1. Name the event before opening either tool. Not "our AI visibility" - "impressions for these 14 product URLs in the Search report's page view" or "displayed citations for these URLs on Microsoft's supported surfaces." If the sentence does not name an observable event, no report can answer it.
2. Pick the report that observes that event most directly. Google's generative AI features on Search: the Search report. Google Discover: the Discover report, which counts on a different basis. Microsoft's supported AI surfaces: Bing Webmaster Tools. Anything else: none of them, and say so rather than substituting the nearest available number.
3. Keep each provider's definitions intact in the output. Label every figure with its provider, its report, its metric name as the provider spells it, its aggregation or sampling condition and its date range. A column headed "AI visibility" containing both providers' numbers is the failure this whole page is about.
4. Segment where the report supports it, and only there. Page, country, device (Search report only), date, grounding query, intent, topic - treating each as a separate view rather than assuming they combine. Do not construct a segment the documentation does not describe, such as an AI Overviews / AI Mode split; Bing's topic filtering runs through an export.
5. Keep a change log beside the numbers. Site releases, page edits, crawl or robots changes, product launches, pricing changes - and reporting-side changes too: a provider metric-definition change, a preview feature moving, an entry on Google's data-anomalies page for your window.
6. Treat every movement as a reason to investigate. Write "coincided with" and "was associated with." Both providers' own documentation supports that verb choice and neither supports a stronger one.
7. Escalate to corroborating evidence when the question outgrows the reports. Display questions stop at these two reports. Behaviour, attribution and other engines need their own instruments, chosen for the event and with their denominators stated.

For a worked example of steps one to three applied to a single narrow question - whether a specific structured-data change moved anything, and what a before-and-after comparison can honestly support - see How to measure whether SaaS pricing schema affects AI visibility, which takes one measurement problem down to its denominators.

Eight states, and where each is observable

Eight distinct states get collapsed into "we're doing well in AI." The table below is the detailed treatment: what each provider's documentation supports for that state, and what remains unknown after both reports have been read. Note that the first two rows are not mutually exclusive - a single displayed link can be counted by Google as an impression and would meet CoreAEX's definition of a citation.

On absences, a wording discipline. Several cells above rest on something not appearing in the current documentation, and that is a weaker statement than a provider saying a thing does not exist. Three phrasings, in descending strength:

• The provider states it does not. Microsoft's statements that AI Performance does not measure rankings or authority, that citations do not represent clicks or traffic, and that grounding queries are not user prompts are of this kind.
• The current public documentation does not describe this field. This covers clicks, click-through rate, position and query data in Google's generative AI reports, and the AI Overviews / AI Mode breakdown. Google has not said these do not exist; it has not documented them.
• The current interface or export would need to be checked. Both products are moving, exports and interfaces sometimes carry fields the help pages have not caught up with, and Google has said additional metrics may come. Where a decision turns on a field's availability, open the product rather than citing this page.

And the rule that keeps all three honest: a missing row is not a zero, and neither is an exported zero on its own. Google's tables carry a 1,000-row limit, its Discover chart omits sub-threshold daily values that remain in the totals, and both reports convert displayed ~ and - values into zeros on download - three separate behaviours; Microsoft states that low or infrequent citation activity may not surface and that this does not indicate a penalty or exclusion. None of those is a measurement of nothing happening. Recommendation

The decision rule

• Use Google's generative AI performance reports to examine documented URL impressions in Google's own generative AI surfaces - AI Overviews and AI Mode in Search, and generative AI features in Discover - naming the report, the dimension and its aggregation basis beside every figure, since the Search and Discover reports do not count the same way.
• Use Bing's AI Performance report to examine displayed source citations across Microsoft Copilot, Bing's AI-generated summaries and its select partner integrations, with the grouped grounding-query context, the classified intents and topics, and Citation Share against that same grounding query - and with the sampling caveat stated beside every figure.
• Do not merge or benchmark the totals. Different definitions, different surfaces, different aggregation, different denominators, and an asymmetry in what each provider documents about completeness. A ratio between them has no referent.
• Choose the instrument by the event you are trying to observe - and when the event is a mention, a recommendation, a referral or a commercial outcome, accept that neither report observes it and find something that does.

Sources

Sources: Google's announcement of the reports, its impression wording, the device note and the statement about adding additional metrics over time are from Introducing Search Generative AI performance reports in Search Console (Search Central blog, Hillel Maoz and Moshe Samet, posted June 3, 2026, carrying a note dated to the August 31, 2026 rollout) - a first-party product announcement rather than help documentation. The report definitions, dimensions, aggregation rules, session and scroll conditions, threshold and truncation statements, export-to-zero behaviour, row limit, Search Labs exclusion, availability reasons and rollout sentence are from the Search Console help pages Generative AI performance report and Generative AI performance report (Discover); the logging-error entries and their restoration note are from Data anomalies in Search Console, which is by design a moving record and should be re-read for your own analysis window. Search Console Help pages publish no last-updated dates, so those three were read on September 7, 2026 and that is the only date available for them. The statement that AI-feature traffic sits inside the Performance report's Web search type is from AI features and your website (last updated December 10, 2025). Microsoft's original metric definitions, supported surfaces, sampling statement and preview status are from Introducing AI Performance in Bing Webmaster Tools Public Preview (February 10, 2026) and the four expanded capabilities from New AI Visibility Insights in Bing Webmaster Tools (June 16, 2026) - both first-party announcements, attributed above rather than treated as documentation. The metric definitions, grounding-query and sampling statements, intent taxonomy, topic and Citation Share definitions, Compare behaviour and the non-ranking statements are from the Bing Webmaster Tools AI Performance help documentation, which publishes no last-updated date and was read on September 7, 2026. Where the announcements and the help documentation word a definition differently, the help documentation is quoted above and the difference is noted in the text. No CoreAEX observation, client result or interface screenshot is reported on this page, and the two figures in the hypothetical illustration are invented for the purpose and labelled as such.

---