Search Console's AI report is impressions only

Google's generative AI performance report in Search Console is an impressions-only view of how often links to your site were shown inside generative AI features such as AI Overviews and AI Mode. It confirms that Google's AI surfaces used your pages. It does not tell you what that visibility was worth, and treating it as a performance metric will produce reports that quietly overstate what you know.

The report was announced on June 3, 2026 and is documented in Google's Search Central blog post. It is a real improvement. Before it existed, AI Mode activity was folded into the ordinary Search totals with no way to separate it: Google confirmed in June 2025 that AI Mode clicks and impressions count toward the same Performance report as blue links, and that you could not filter to see them on their own. Now there is a dedicated view. The question is what that view supports.

#What the report contains

Google's help documentation defines the metric plainly: impressions are how many times links to your site were shown to a user in a generative AI feature on Google Search. The report breaks that number down by four dimensions: pages, countries, devices, and dates.

That is the whole surface. There is no click column, no click-through rate, no average position, and no query dimension. Search Engine Journal's read of the launch listed the same gaps and added the ones people notice later: no citation placement, no indication of which passage was used, and no conversion data.

Three consequences follow, and each one changes how you should write the report up.

#An impression is not a visit, and the clicks are elsewhere

The absence of clicks in this report does not mean AI Overview and AI Mode clicks go uncounted. They are counted, they are just counted somewhere else: in the ordinary Web search type total, blended with everything else, exactly as they have been since Google started including AI Mode in Performance data in 2025.

So you now have two numbers that cannot be divided by each other. The generative AI report gives you AI impressions. The Search report gives you total clicks. There is no honest CTR you can compute across them, because the denominator and the numerator come from different populations. Any dashboard that puts "AI CTR" in a cell is doing arithmetic on two unrelated measurements.

The practical version: use the generative AI report to answer "did our pages appear," and use the standard Performance report, plus your analytics, to answer "did anyone arrive." Keep the two questions in separate rows.

#Property-level aggregation compresses your best pages

The chart aggregates by property by default. Google's documentation is explicit that if two results from the same site appear in one generative AI feature, they count as a single impression in the chart total. Filter by URL and the chart aggregates by that URL instead. The table follows a similar rule: grouped by country, device, or date it aggregates by property, and grouped by pages it aggregates by page.

That default is reasonable, and it is also a trap for anyone comparing periods. A month where one page appeared in a thousand responses and a month where four pages appeared together in a thousand responses can produce a similar chart. Depth of coverage, which is the thing a content team actually influences, shows up in the Pages table rather than in the headline line graph. Read the Pages view first, and treat the chart as context.

#The data exists in one place, and it can be wrong

This is the limit that turned concrete this month. Google's data anomalies page records that a logging error decreased impressions on the generative AI performance report in Search for data from August 13 to August 17, 2026, and that a separate logging error decreased clicks and impressions on the Discover performance report for August 13. Search Engine Land covered the Search-side bug while it was still open, with Google's own framing that the issue affected data logging only. John Mueller's public note was that this was a logging issue and not a change in visibility.

Reporting bugs happen in every analytics product. What makes this one instructive is that there was nowhere else to look. Search Engine Journal's tracking guide reports a re-verification on a live property on 11 August 2026 that neither the Search Analytics API nor the BigQuery bulk export exposes the generative AI data, and that there is no official way to track AI Mode traffic at query level. If the interface is the only copy, then a five-day logging gap becomes a five-day hole in your record, and you cannot reconstruct it later or compare it against an independent pull.

There is a cheap defense. Export the report to a sheet on a fixed schedule, keep the export dated, and annotate known anomaly windows against it. That does not fix the underlying counts, but it preserves what you saw and when, which is what you need when someone asks in November why August looks soft.

#How to use it in a GEO report

The generative AI report answers one question well: are Google's AI surfaces using our pages at all, and which ones. That is a coverage signal, and coverage is a legitimate leading indicator. A page that has never registered an AI impression is not in the candidate set. A page that registers them steadily is being retrieved and shown, whatever the downstream click behavior looks like.

What it cannot do is stand in for a citation metric. It is limited to Google's own surfaces, so it says nothing about ChatGPT, Perplexity, or Claude. It reports appearances rather than influence, so it cannot distinguish a link that shaped an answer from one listed underneath it. It also has no query dimension, so you cannot tie an impression to the prompt that produced it.

Our own approach at Crawlmind is to keep prompt-level citation tracking as the primary measurement, because that is where the query, the engine, and the competitive set are all visible at once, and to read the Search Console generative AI report alongside it as Google-specific corroboration. When the two disagree, the disagreement is usually informative: it points at a surface, a market, or a page template that one method sees and the other does not.

Use the report. Just label it accurately. It is a count of appearances on one vendor's AI surfaces, held in one interface, and it is occasionally wrong for a few days at a time.