You Can Finally Measure AI Search. Sort Of.

For most of the past two years the honest answer to “how visible are we in AI results” was that nobody could tell you from first-party data. As of 31 August 2026 that changed for every site in the world, and it changed less than the announcements suggest. Google now reports AI impressions in Search Console. It reports nothing else, and the gap between an impression and a decision is where most of the work still sits.

What is now measurable that was not

The generative AI performance report carries one metric. Impressions are defined as “how many times links to your site were shown to a user in a generative AI feature on Google Search”. That metric is available across four dimensions: pages, which groups “by the final URL linked by a generative AI feature after any redirects”; countries, grouped by where the search originated; dates, grouped by day, week or month; and devices, split across desktop, tablet and mobile.

Google’s announcement of the generative AI performance reports went out on 3 June 2026 and initially reached a subset of site owners in the UK. The update note on that announcement now reads: “As of August 31, 2026, we’ve rolled out these features to all websites worldwide.” The Search generative AI control in Search Console settings reached global availability on the same date, which means the reporting and the opt-out arrived together — deliberately, since one is useless without the other.

The three things still missing

There are no clicks. There is no click-through rate. There is no average position. Those absences are not a rollout gap, they are the shape of the report, and every analysis built on it has to survive them.

There is also no split between AI Overviews and AI Mode. The report combines both Search-side generative surfaces into a single view, with Discover reported separately. So a site whose impressions climb cannot tell whether it is being surfaced more often in the summary above ordinary results or inside a conversational session, and those two are different products with different user behaviour behind them.

The fourth constraint is the one people trip over. The report “includes data from the Web search type in the Performance report (Search results)”. These impressions are already counted in your ordinary performance report. Adding the two together double-counts, and a dashboard that sums them will report growth that did not happen.

Reconstructing click behaviour without click data

The method that holds up is comparison over time rather than calculation. Take total Web clicks from the standard performance report and generative AI impressions from the new one, plot both across the same date range, and look for divergence: impressions rising while clicks flatten is the pattern worth investigating, and a coincident drop in both is usually something more ordinary.

State the limits when you present it. The two series are not independent, so this is not a funnel and no ratio derived from it is a click-through rate. Seasonality, unrelated ranking movement and query-mix changes all move both lines. What the comparison supports is a direction and a date of onset, not a number. Anyone converting it into a percentage is inventing precision, and GA4 for SEO decisions covers the same failure mode on the analytics side.

What a defensible AI visibility report contains

Four things, and deliberately not more:

• Generative AI impressions by page, over a period long enough to see a trend, with the eligible page set named rather than implied.
• Country segmentation, because generative surfaces roll out unevenly and a change confined to one market is a rollout, not a performance change.
• Date of onset for any step change, cross-referenced against your own publishing and technical changes before anything is attributed to Google.
• An explicit limitations note stating that impressions are the only metric, that AI Overviews and AI Mode are not separated, and that the 1,000-row limit that applies to the standard performance report applies here too.

Referral identification outside Search Console fills some of the remaining gap and should be labelled as partial when it does. A referrer-based segment catches assistant traffic that arrives with a referrer attached; it misses traffic from clients that send none, and it cannot distinguish a citation nobody clicked from one that was never shown. Treat it as a floor on assistant referrals, never a total.

The strategy this measurement was supposed to validate is the one set out in building a strategy when you couldn’t measure it, and impressions data does not overturn its conclusions so much as let you check them. The signal question underneath — what actually gets a site cited — is still better addressed by the correlation work in backlinks versus AI citations than by anything in this report. Google has said further metrics may follow based on feedback, and until they do, an impressions count is what there is.