Search Console Generative AI Report: What It Measures

There is a particular kind of meeting that happens about a week after Google ships a new report. Somebody screenshots the new chart, drops it into the channel, and writes: look, we are showing up in AI. The number is large. The line points up and to the right. Everybody feels better for roughly four days, right up until someone asks what the number means for the business and the room goes quiet.

That meeting has been happening in a lot of companies since August 31, 2026, the date Google finished rolling the Search Console generative AI report out to every property worldwide. The feature had been in limited testing with selected UK site owners since the June 3 announcement, then widened through July, and now everyone has it. For the first time there is an official Google surface that says, in numbers, how often your pages appeared inside AI Overviews and AI Mode.

That is genuinely useful. It is also much narrower than the celebration suggests, and the narrowness is not a footnote. It changes what you are allowed to claim.

AUG 31

the 2026 date Google finished rolling the report out to all properties worldwide, after a June 3 announcement and a UK-only test phase

metric in the report. Impressions. There is no clicks column, no click-through rate, and no average position

query rows. The report breaks out pages, countries, devices and dates, and it does not break out what anyone asked

1,000

row limit, inherited from the standard performance report, which caps how far into the page-level detail you can actually see

Read that band again before you build a dashboard on top of it. One metric and three usable dimensions is not a visibility system. It is a presence detector. Those are different instruments and they answer different questions.

What the Search Console generative AI report actually counts

An impression here means a link to your site was shown to a user inside a generative AI feature on Google Search. That covers AI Overviews and AI Mode. It explicitly excludes Search Labs experiments, which matters more than it sounds like it should, because Labs is where a lot of the features people are anxious about are still living. Generative AI in Discover is tracked separately, in its own report, so a Discover-heavy publisher looking only at the Search view is reading half a picture.

DIMENSION OR METRIC	IN THE GENERATIVE AI REPORT	IN THE STANDARD SEARCH PERFORMANCE REPORT
Impressions	Yes, the only metric available	Yes
Clicks and click-through rate	No	Yes
Average position	No	Yes
Queries	No	Yes, sampled and privacy filtered
Pages	Yes, grouped by canonical URL	Yes, grouped by canonical URL
Countries and devices	Yes	Yes
Dates	Yes, by day, week or month in Pacific Time	Yes
Surfaces covered	AI Overviews and AI Mode, Labs excluded	Web search results overall

The column on the right is the one people skip, and it is where the useful comparison lives. Everything the generative AI view gives you, the standard view already gave you, minus four things the standard view also gives you. The new report is not a richer dataset. It is a filtered one.

THE DISTINCTION THAT MATTERSAn impression says a link was displayed. It does not say the link was read, clicked, credited or acted on. If your reporting and analytics practice already separates exposure metrics from outcome metrics, this report slots cleanly into the exposure column and nowhere else.

Those impressions were already in your totals

Here is the part that gets missed, and it is the single most consequential line in Google's own documentation. The data behind the generative AI report is drawn from the web search type in the performance report. The same web search type you have been exporting every month for years.

Which means these are not new impressions. They are a labelled subset of impressions you were already counting. When your AI Overviews appearance registered in August, it registered in your web search total in August too. Nothing was hidden and nothing has been added. Google just told you which slice of the existing number came from a generative surface.

“The report did not find you new visibility. It put a label on visibility that was already in the file, and a label is not a lift.”

This matters in one very practical way. If you add AI impressions to your organic impressions and present the sum as total search visibility, you have double counted, and the error scales with how well your AI work is going. The better you do, the more wrong the slide gets. We walked through an almost identical failure mode in why missing analytics data needs reporting redundancy, where teams reconciled two sources that were never independent and quietly manufactured growth that did not exist.

The correct framing is a ratio, not a sum. AI impressions as a share of web search impressions tells you how much of your existing exposure is being mediated by a generative surface. That number is small for most sites today and rising for nearly all of them, and watching it move is worth far more than watching the raw count move.

Three columns that are not there

Google has said it will introduce additional metrics over time, with no timeline attached. Until then, three absences shape everything you can do with this data.

01No clicks, so no value sideImpressions without clicks give you a numerator with no denominator and no outcome. You cannot calculate a rate, you cannot compare surfaces, and you cannot tell a growing appearance count apart from a growing appearance count that sends nobody. Those two situations call for opposite responses.

02No queries, so no diagnosisPage, country, device and date tell you where an appearance happened. None of them tell you what triggered it. When an AI impression count drops 30% on a page, the report cannot tell you whether demand fell, whether a competitor displaced you, or whether the question itself stopped being asked in that shape.

03No position, so no competitive readIn classic search, position is how you know whether you moved or the results around you did. Inside a generated answer there is no position to report, and the citation order that does exist is not exposed here. Your appearance and a competitor's appearance look identical in this data.

04Labs excluded, so no early warningThe experimental surfaces are where the next behaviour change gets tested, and they are outside the report by design. A feature can be reshaping how your category gets answered for months before any of it reaches a chart you can see.

Lily Ray flagged the missing clicks column on the day of the June 3 announcement and read the omission as a signal in itself, that appearance and traffic are becoming separate things rather than two views of one thing. That reading has held up. Ilias Ism had argued a related mechanical point a week earlier, on May 28, 2026: assistants that read Google's results often pull the page itself without ever passing through the tracked link, so the citation is recorded and the visit never is. Neither of them has measured click data proving it, and both were careful to say so. The hypothesis is honest and unconfirmed, which is exactly how it should be reported.

The click studies fill the gap the report leaves

Google will not tell you what an AI appearance is worth. Three independent studies have tried, using three different methods, and they do not agree on a number. They agree on a direction, which is more useful than a false consensus would be.

Pew Research: users clicked a result on 8% of visits with an AI Overview against 15% without, a 47% relative fall47%

Ahrefs, updated research published February 2026: click-through rate for the top ranking page on AI Overview keywords, down 58%58%

Seer Interactive: 25.1 million organic impressions across 42 organizations, organic click-through rate from 1.76% to 0.61%, down 61%61%

Three separate measurements of what an AI Overview does to clicks. These are not directly comparable: each one measures a different population with a different method, and the spread between them is the honest error bar on the question.

Do not average these. Pew measured user behaviour across visits. Ahrefs measured the top ranking page on a keyword set. Seer measured aggregate organic click-through rate across a client portfolio. Three populations, three definitions, one direction. Anyone who hands you a single tidy percentage for what AI Overviews cost you has collapsed a real disagreement into a number that sounds better than it is.

What you can take from the spread is a planning range rather than a point estimate. If an AI appearance converts to a click at somewhere between a third and a half of the rate a comparable classic result would, then a rising impression count with a flat click count is the expected outcome, not a failure of your work. It is the shape of the surface. Teams that have not internalised that tend to react to a normal pattern as though it were a crisis, which is how good programmes get cancelled.

How to read the Search Console generative AI report honestly

The report is worth instrumenting properly. It is just worth instrumenting as what it is. Five things to set up this week, none of which take longer than an afternoon.

Track the ratio, never the raw countExport generative AI impressions and web search impressions on the same cadence and report the share, not the sum. The share is the only figure that survives contact with a growing site, because a raw count rises when traffic rises for reasons that have nothing to do with AI surfaces.

Join it to landing page data yourselfThe report gives you pages. Your analytics gives you sessions and outcomes by page. Google will not join those for you, and the join is where the value is. It is approximate and it is still better than an impressions chart on its own, provided you label it as an approximation on the slide.

Recover the query layer from elsewhereSince February 23, 2026 the performance report has accepted natural language filters that it converts into regular expressions behind the scenes, which makes query cohorts much faster to build. Those cohorts are not AI specific, but pairing a cohort trend with the generative impressions trend for the same pages gets you closer to intent than either view does alone.

Separate exposure reporting from ranking reportingKeep the generative impressions line on a different chart from your position data. The two move independently and stacking them invites the exact confusion we described in why rank tracking accuracy breaks on redirects. One is a presence signal, the other is a placement signal.

Record whether the opt-out is set, and whenGoogle has stated the generative AI control is not used as a ranking signal elsewhere in Search, so the setting is a genuine choice rather than a trap. It still changes what the report can show you, so the date it was last touched belongs in your annotations. The tradeoffs are the ones we laid out in the generative AI search control and what opting out costs.

Two operational cautions while you build. The most recent days are marked preliminary with a dotted line and will move, so do not let a Monday morning dip start a fire drill. And the chart totals and the table totals can legitimately differ, because the chart aggregates at property level while the page table aggregates by canonical URL. That is not a bug and it is not your export breaking. It is the same aggregation behaviour the standard report has always had, in a place people are not used to seeing it.

The number worth putting on the board

If you get one line into the executive deck about generative search this quarter, do not make it impressions. Make it the share of your search exposure now being mediated by an AI surface, with last quarter beside it and the trend between them. That single ratio answers the question executives are actually asking, which is not are we in AI, it is how much of our demand now passes through a layer we do not control.

Then add one honest sentence about measurement: appearance data exists and outcome data does not yet, so the value of an AI appearance is estimated from a range of published studies rather than measured directly. Saying that out loud costs you nothing and buys you the credibility to revise the estimate later, which you will have to do. We made the same argument about attributing AI-influenced sessions in how AI traffic distorts attribution and bidding signals, and it has aged well for one reason: the teams that stated their uncertainty early were the ones still trusted when the numbers changed.

The report is a real improvement. For two years the honest answer to are we showing up in AI Overviews was a shrug and a third-party estimate, and now it is a number from the platform itself. That is progress. It is also the first floor of a building, and a lot of people are decorating it like a penthouse.

DO THIS NEXTPull generative AI impressions and web search impressions for the same 90 days, put them in one sheet, and calculate the share for each month. If that share is climbing while your organic clicks are flat, your exposure work is functioning and your conversion path is the constraint, which is a generative engine optimization problem with a known shape rather than a mystery. Read Google's own documentation on the report before you brief anyone on it, because the stated limits are the part that changes what you are entitled to claim.

See where you are cited today

A free snapshot audit of your rankings and AI citations before we ever talk.

Schedule a free consultation

Tyler TruffiMANAGING PARTNER, SOMETHING INC.

Tyler leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.

Your AI Impressions Number Is Not Visibility