Published
Decide which of the four numbers you need
AI Overview visibility is not one measurement. Four different things are sold under the name, they come from different instruments, and they move independently of each other. Presence asks whether a query triggers an AI Overview at all. Inclusion asks whether your URL is among the sources cited inside it. Impression asks whether Google counted your link as shown to a user. Referral asks whether anyone clicked and arrived.
Nothing forces those four to agree. Your page can be cited in a block that no reader ever scrolled into view, which produces inclusion with no impression. It can be shown and never clicked, which is the ordinary case rather than the exception. Choosing an instrument before deciding which of the four you are measuring is how tracking programs end up with a chart that nobody can interpret and a monthly argument about whether the number went down.
Two of these you can obtain as a census of your own property. Two you can only sample. Knowing which is which is most of the discipline.
Start with the report Google gives you free
Since 3 June 2026, Google Search Console has carried a generative AI performance report, with data beginning 18 May 2026. It reports how many times links to your site were shown to a user in a generative AI feature on Google Search, broken down by page, country, date and device. That is the entire report: four dimensions and one metric.
Read the boundaries carefully, because most of the reporting mistakes in this subject are made in the first ten minutes with this screen. There is no click metric, no click-through rate, no query dimension, and no dimension separating AI Overviews from AI Mode. Whatever figure you take from it is AI Overviews and AI Mode combined. The report inherits Search Console's usual limits, including the thousand-row cap; recent data is marked preliminary and can change; and chart totals and table totals can differ because one aggregates by property and the other by page. Rollout is still incomplete, and John Mueller confirmed as much on 11 August 2026: "Correct, it's not yet every domain."
Within those limits it is the most valuable number available to you, for one structural reason. Every third-party AI Overview tracker samples. Google counts. If you are going to build a trend line, build it on the census and use sampled data to explain the census rather than to replace it.
What one of those impressions actually represents
Google says standard impression rules apply to AI Overviews. Brodie Clark's crowd-sourced test, first published 18 June 2024 and updated 8 June 2026, established the practical shape of that: AI Overview impressions register only when the citation is scrolled or expanded into view, not automatically when the page loads. The mechanism finding has not been contradicted since. The click-through rate from the same test — 13.5%, derived from 362 valid US impressions when AI Overviews were largely limited to logged-in users — does not generalize to 2026 and should never be quoted as a benchmark. Keep the mechanism, discard the rate.
The consequence is specific and it disposes of a whole genre of dashboard. A scraper that expands the block to enumerate citations records sources that no human ever revealed, and Google will not count those as impressions. A citation buried behind a "show more" control under-registers in Search Console by design. Your tracker's citation count and your impression count are therefore measuring different events, and the gap between them will be widest exactly where your citations sit deepest. Reconciling the two is not diligence; it is a category error, and any vendor offering to reconcile them for you should be asked how.
One check is one draw from a distribution
The measurement problem peculiar to this subject is that the thing being measured is generated rather than retrieved from a fixed index, and the size of that effect has been quantified twice.
Ahrefs, analyzing 540,000 query pairs of September 2025 US data and publishing on 15 December 2025, found that 45% of AI Overview citations change between generations. Regenerate the same overview and close to half the source list is different. SE Ranking parsed the same 10,000 keywords three times on a single day in AI Mode, published 29 August 2025, and found that 9.2% of cited URLs appeared in all three runs. Both studies were funded by the vendors that ran them, and both sell software to the people who cite them; neither result has been contradicted by anyone.
Taken together they mean that a daily inclusion check on one keyword is close to a coin-flip-grade instrument. A domain can appear on Monday, vanish on Tuesday and return on Wednesday with nothing having changed on the site or in the index. "We lost our AI Overview citation" from a single observation reports a draw, not an event.
What survives this is a rate. Fix a keyword set and freeze it. Sample each keyword several times inside each collection window rather than once. Report the share of samples in which you were cited, and print the number of samples beside it. A tool that cannot tell you its runs per keyword cannot tell you whether last week's drop happened, and neither can you.
The number your tracker reports and Google does not recognize
Position is the one measurement in this exercise that does not exist. Google's documentation is unambiguous: "An AI Overview occupies a single position in search results, and all links in the AI Overview are assigned that same position" (Search Console Help). Every citation in the block shares one position number.
So a tracker reporting that you moved from citation four to citation one is describing the visual order of a rendered block, which is real enough on screen and corresponds to nothing in Google's own reporting. Microsoft has said the same thing from the other side about its own data, stating that Bing's citation counts do not indicate ranking, authority, or the role of any page within an individual answer. Two operators, independently, have told the market that there is no meaningful ordering of sources inside an answer.
The practical instruction is short. Any score that weights your citations by where they appear in the block is inventing an ordering that both platforms deny, and the column should come out of the report. Citation is binary within a single generation. It becomes a rate only when you sample repeatedly, which is the honest version of the metric that vendors were reaching for when they built the position field.
Your rank tracker stopped covering the citation pool
For roughly two years the working assumption was that AI Overview citations came mostly from the classic top ten, which would have made ordinary rank tracking an adequate proxy and this whole exercise unnecessary. That assumption was true and is no longer.
Ahrefs measured 76.10% of AI Overview citations coming from pages ranking in the organic top ten on 21 July 2025, with the median organic position of the first citation at 2. The same publisher re-ran the study on 863,000 keyword SERPs and about four million AI Overview URLs and reported 37.9% on 2 March 2026 (Ahrefs). Two caveats travel badly and matter here: Ahrefs states the second study used improved parsing, so an unquantified part of the movement is measurement change rather than Google change, and both studies are the work of one vendor with a commercial interest in the urgency of the finding. Treat the direction as well evidenced and the exact multiple as approximate.
Either way the operational conclusion holds. Roughly a third of citations now come from pages that do not rank in the first hundred results for the visible query — pages your rank tracker is not watching at all. That is also what query fan-out predicts. When a system decomposes one question into several sub-queries before an answer is written, and selects sources against sub-queries the user never typed, there is no reason for the citations to come from the visible query's top ten. Tracking organic position as a proxy for citation eligibility was defensible in mid-2025. It is not defensible now.
Location, device, and a citation set the reader configured
Country and device are dimensions in Google's own report precisely because results differ across them. A tracker sampling from one datacenter in one country is measuring one cell of a large table, and a report built on it should say which cell rather than implying a global reading.
Personalization makes this sharper. Since 27 May 2026, a user's Preferred Sources selections in Google Search are reflected inside AI Overviews and AI Mode, which makes part of the citation set user-configured. No logged-out scraper's view is any particular person's view, and no amount of sampling closes that gap, because the variation is by design rather than by noise.
The same caution applies to borrowed benchmarks. Semrush's panel of more than ten million keywords recorded AI Overview presence at 6.49% in January 2025, 24.61% in July 2025 and 15.69% in November 2025 — one panel, one year, three numbers that have each been used to argue that AI Overviews are everywhere or that they are retreating. Presence is a product decision that moves in both directions. Quote a presence rate with its month, its panel and its country, or do not quote it.
Separating a retrieval failure from a selection failure
When a section of a site is never cited, there are two very different explanations. Either the pages were never fetched, or they were fetched and not chosen. Only server and CDN logs distinguish those, and the distinction is worth more than another visibility chart, because the remedies have nothing in common.
Logs will tell you which directories AI crawlers reach and which they never touch, what status codes they receive, and whether crawl volume is being spent on parameter noise. A page that was never fetched cannot be cited, and that is a diagnosis you can act on this week.
What logs will not tell you about AI Overviews specifically is which crawler was working on them, because there isn't one. Google documents Googlebot as affecting Google Search including Discover and all Google Search features, and AI Overviews and AI Mode are Search features. Google-Extended is a separate control governing whether content is used to train Gemini models and ground the Gemini apps and Vertex AI generative APIs, and Google states it does not affect a site's inclusion in Google Search and is not a ranking signal there. There is no user agent that isolates AI Overview crawling from ordinary Google crawling. Any tool showing you "Google's AI crawler" in your logs is showing you Googlebot with a new label. Beyond that, crawling is not citing, and no operator or study has published any ratio between the two.
Frequently asked questions
Can I calculate an AI Overview click-through rate?
No, not from Google's data as of 29 August 2026. The generative AI performance report has impressions and no clicks, and the clicks that do exist are pooled into the undifferentiated Web search type with no filter. The division has no denominator on one side and no numerator on the other, so any AI Overview CTR you are shown was constructed from an inference the vendor should be asked to explain.
Does the Search Console generative AI report show AI Overviews only?
No. It combines AI Overviews and AI Mode with no dimension separating them. Trade coverage has said otherwise — Search Engine Journal's 4 August 2026 explainer described the report as treating the two features as distinct — but Google's help page lists four dimensions, pages, countries, dates and devices, and no surface dimension (Search Console Help). Where the trade press and the documentation disagree, the documentation is checkable.
How many times should I sample each keyword?
No platform publishes variance figures for its own output, so there is no authoritative answer to copy. The workable standard is empirical: sample often enough that your reported citation rate is stable across consecutive collection windows, and publish the run count next to the rate so a reader can judge it. One run per keyword per week is not a measurement of anything at 45% citation change between generations.
Should the report say we lost an AI Overview citation?
Not from one observation, and not from one week. Report the share of sampled runs in which the citation appeared, across a fixed keyword set, and compare like with like. A citation rate that falls across several windows on a frozen keyword set is a finding. A single absence is the expected behavior of the system.
Where do AI Overview clicks appear in Google Analytics?
They do not, and they cannot. Clicks from AI Overviews and AI Mode carry a google.com referrer that is identical to organic Google search, so no signal exists to segment on. Any dashboard presenting AI Overview traffic from analytics data is presenting an inference rather than a measurement, and the inference method is the thing to ask for.