What AI search analytics can and cannot tell you

The report that arrived on 3 June 2026

Google Search Console gained a generative AI performance report on 3 June 2026, with data beginning 18 May 2026. Google's definition of the metric is one sentence: impressions are how many times links to your site were shown to a user in a generative AI feature on Google Search.

The report has four dimensions and no others — pages, countries, dates and devices — where pages groups data by the final URL linked after any redirects. It carries Search Console's ordinary constraints: the thousand-row limit, the standard time periods, and a warning that the newest data is preliminary and may change within hours. Chart and table totals can disagree, which Google attributes to aggregating by property in one and by page in the other, and Search Labs experiments are excluded. Rollout was deliberately partial, described by Google as a subset of website owners while it tested, and still incomplete more than two months later: asked on 11 August 2026 whether the report had reached every domain, John Mueller replied, "Correct, it's not yet every domain."

One structural point about the report is easy to miss and changes how it should be read. Google states that it draws on data already inside the Web search type in the main Performance report. It is a view onto something Search Console was already counting, not a new collection — which is exactly why its silences are so specific.

Impressions only, and what that forecloses

The report contains no clicks, no click-through rate, no query dimension, and no dimension separating AI Overviews from AI Mode. Each of those absences closes off a question that people arrive expecting to answer.

You cannot compute an AI Overview click-through rate: there are impressions on one side and nothing on the other, so any AI CTR you are shown came from somewhere else. You cannot learn which query produced an AI impression, because there is no query dimension. You cannot report a figure from it as an AI Overviews number, because it is AI Overviews and AI Mode combined. And you cannot compare generative AI impressions against Web impressions to compute an AI share of your exposure: the generative impressions are a subset already sitting inside the Web total, so the ratio is a proportion of a thing to itself rather than a share of anything you can act on.

That last point is where the trade press and the documentation part company, and it is worth naming who says what. Search Engine Journal's 4 August 2026 explainer described the report as separating AI Overviews from AI Mode and treating them as distinct features. Google's help page lists four dimensions and no surface dimension. The two features are described separately in Google's counting documentation, which is the likely origin of the confusion, but the report does not segment by them. Where secondary coverage and primary documentation disagree, the documentation is the one a reader can check.

The clicks are counted. They are pooled.

This is the most consequential measurement fact on the subject, and it is easy to state. Google documents that "Clicking a link to an external page in the AI Overview counts as a click," and says the same of AI Mode (Search Console Help). Separately, Google's documentation on AI features, last updated 10 December 2025, states that sites appearing in AI features such as AI Overviews and AI Mode are included in the overall search traffic in Search Console, inside the Web search type.

So the clicks exist, Google counts them, and there is no filter that isolates them. Search Engine Land reached the same conclusion about AI Mode on 16 June 2025, reporting that Google would not let anyone filter to see AI Mode impressions, clicks and CTR on their own. Nothing in the June 2026 report changed it; the report added impressions and left the clicks where they were.

Read the two halves together, because they are usually quoted apart. It is not that Google cannot see AI-driven clicks, or that they are lost. They are measured, held, and reported to you only inside an aggregate containing every ordinary blue-link click on your site. That asymmetry is a disclosure decision, and it is why the commercial measurement category exists at all.

One position for every link in the answer

Google's counting documentation contains a second sentence that disposes of a whole product genre: "An AI Overview occupies a single position in search results, and all links in the AI Overview are assigned that same position."

Two consequences follow. Because AI features are counted inside the Web search type, the position an AI Overview contributes to your average position is the block's position rather than a rank inside it, so an average position that moves may reflect a change in which surfaces you appeared on rather than a change in ranking. And any tool reporting that you moved from the fourth citation to the first is describing the visual order of a rendered block. Microsoft says the same from its own side: Bing's citation data "does not indicate ranking, authority, or the role of any page within an individual answer." Two operators have independently told the market there is no meaningful ordering of sources inside an answer, and a score weighted by citation position contradicts both.

AI Mode queries are already in Search Console, unlabeled

"AI Mode queries are invisible in Search Console" is repeated constantly and is wrong as stated. They are present. They are simply not where people look, and they are not marked.

John Mueller, 6 August 2026: "If a user asks a follow-up question within AI Mode, they are essentially performing a new query. All impression, position, and click data in the new response are counted as coming from this new user query." Those queries land in the main Performance report, mixed in with everything else and carrying no label. Practitioners identify them by their conversational shape — strings such as "yes" and "yes go on" that no one types into a search box — which is a heuristic rather than a filter. It will miss real ones and collect false positives, and there is no ground truth published anywhere to validate it against.

The useful version has two halves and both matter: AI Mode queries appear in the main report, unlabeled and unfilterable, and they are absent from the generative AI report, which has no query dimension. Anyone quoting only the first half overstates what you can do with them; anyone quoting only the second claims an absence that is not there.

What Microsoft reports that Google does not

Microsoft shipped platform-owned AI reporting four months before Google did, and on the questions that matter most for diagnosis it still reports more.

Bing Webmaster Tools' AI Performance report, in public preview since 10 February 2026, shows the total number of citations displayed as sources in AI-generated answers across Microsoft Copilot, AI-generated summaries in Bing and select partner integrations. It also exposes grounding queries — the retrieval queries the system actually issued, rather than the words the user typed. That is the sub-query layer Google's query fan-out produces and exposes in no product anywhere. Since 16 June 2026 Microsoft has added Citation Share, the percentage of citations attributed to your site out of all citations shown across all sites for that same grounding query.

Citation Share is the only share-of-voice metric in AI search with a genuine denominator, because Microsoft is counting citations it served rather than sampling a prompt list. It is also fenced in by its owner with unusual candor: the citation data "does not indicate ranking, authority, or the role of any page within an individual answer"; Citation Share "does not expose competitor domains" and is not a "competitive scoreboard"; and the report carries no traffic metric of any kind. Microsoft tells you why you were retrieved and what fraction of the citations you held, and nothing about what it sent you. Google tells you that links were shown and nothing about why. Neither is a performance report, and between them they cover two operators out of the field.

After the click: what your own analytics can see

Attribution is a narrower question than visibility. Visibility asks whether you appeared; attribution asks whether appearing produced anything, and it turns on two ordinary signals surviving the trip: the referrer header naming where the link was, and a UTM parameter where the operator adds one.

As of 29 August 2026 exactly one AI operator has made a publisher attribution commitment. OpenAI's publisher FAQ states that ChatGPT automatically includes utm_source=chatgpt.com in referral URLs — scoped to referral URLs in ChatGPT search results, not to every outbound link in the product. Secondary coverage of a June 2025 extension renders the parameter as utm_source=chatgpt, disagreeing with OpenAI's own documentation, so match a prefix and the referrer host rather than one exact string. Every other identification in your analytics, including Copilot's referrer observed since 29 March 2024 and Perplexity's, is a reverse-engineered convention no operator has promised and any of them can change silently.

The largest surface is the least attributable. Clicks from AI Overviews and AI Mode arrive with a google.com referrer, indistinguishable from organic Google search, so the surface with by far the most reach is the one where neither your analytics nor Google's reporting separates the traffic. Any AI segment you build excludes it by construction.

On size, the largest dataset available is SE Ranking's: 101,574 sites across 250 countries with analytics integration over sixteen months, published 18 June 2026 and funded by SE Ranking, which sells SEO software. AI referral traffic was 0.02% of all website traffic in 2024, 0.24% in 2025 and 0.32% in 2026 — sixteenfold growth on a base of about a third of one percent. Quoting the multiple without the base overstates the present; quoting the base without the multiple misses the trajectory. The same work reports AI-referred visitors spending 67.7% longer on site than organic visitors, and 60% of AI referral traffic landing on homepages against 17% from organic search — which matters for reporting, since landing-page analysis of AI traffic piles up on the homepage and says little about which page earned the recommendation.

The three things no instrument sees

Some questions have no measurement behind them at any budget, and a report that does not name them will be read as though it did.

The brand named without a link. An assistant recommends you, the reader searches your name or types your domain, and arrives. This is not attribution failure; there is no referral event to capture. It surfaces as direct traffic or branded organic search, it is invisible to every instrument on both sides of the transaction, and it is plausibly the largest single category of AI-influenced visit — it is exactly the outcome an AI answer is built to produce.

The size of the referrer gap. The mechanisms are well understood — native apps that do not send referrers, restrictive referrer policies, copied and pasted URLs, cross-device journeys. The magnitude is not: no study was found on 29 August 2026 measuring the direct-traffic leakage rate for AI referrals with a published method. The one quantified estimate located is practitioner research from 17 June 2026 putting true Claude-influenced visits at roughly two to three times measured referrals, from a dataset of more than a hundred brands whose methodology is not published. That is the strongest figure available, and it is not strong enough to forecast on. Report every AI referral number as a floor of unknown tightness.

Whether the answer changed what someone did. The strongest evidence on this comes from outside the industry. Pew Research Center — funded by The Pew Charitable Trusts, and the only non-vendor funder among the studies underpinning this page — recruited 900 US adults who shared browsing data, covering 68,879 unique Google searches of which 12,593 produced an AI summary, and published on 22 July 2025 (Pew Research Center). Users clicked a result link in 8% of visits where an AI summary appeared, against 15% where none did; 1% of visits included a click on a link inside the summary; and 26% of sessions with a summary ended browsing entirely, against 16% without. Google's position, published by Liz Reid on 6 August 2025, is that total organic click volume has been relatively stable year over year and that Google sends slightly more quality clicks, where quality clicks are defined as those where users do not quickly click back. That post contains no figures, no sample size and no method, while describing third-party studies as using flawed methodologies. One side published a specified method and numbers; the other published an assertion. Report Google's statement as its position, attributed and dated, and never as a finding.

Frequently asked questions