Search Console's Generative AI Report: 3 Things It Hides

19,700 impressions. That is what Google’s new Generative AI report says chudi.dev earned inside AI Overviews and AI Mode since May 18. It will not name a single question that produced them.

The report appeared in every Search Console property at once on August 11, 2026. Google never announced it. Most site owners logged in, saw a pop-up pointing at a new section, opened it, and found a number they cannot act on: impressions, with no query attached, no click attached, and no way out of the dashboard except by hand.

I probed the API to confirm that last part. Then I checked what the traffic behind that number actually did on my own site over 90 days.

What Is the Search Console Generative AI Report?

The Generative AI performance report is a Search Console section that isolates your impressions inside Google’s generative AI surfaces, AI Overviews and AI Mode, from your standard organic search impressions. Google launched it on June 3, 2026 to a limited subset of properties, mostly UK-based, with data reaching back to May 18, 2026. On August 11 it went live for everyone, with no formal announcement, which Search Engine Watch confirmed independently the same day.

It breaks impressions down four ways: by URL, by country, by device, and over time. That is the whole report.

Here is mine, still carrying the Beta label, covering the full window from May 18 to August 12, 2026:

That is 19,700 impressions across three months, on a site with a domain rating in the teens. The curve matters more than the total. Near zero through late May. First real movement in early June. A step change from early July that has held since. Whatever Google changed in its source selection over the summer, small sites got pulled into the answer layer, and none of us could see it until yesterday.

The screenshot has no clicks metric, no CTR, and no average position. There is no queries tab either. The tabs are PAGES, COUNTRIES, DEVICES, and DAYS.

What the Generative AI Report Hides

Three omissions, in ascending order of how much they cost you.

1. No Queries

The report will tell you that /blog/your-post picked up 4,000 impressions in AI surfaces. It will not tell you what anybody asked. You get the page, never the prompt.

This matters more in AI surfaces than in blue links, because AI Mode queries are conversational and long. A page can be cited for a question you never targeted and would never have written for. Without the query, you are optimizing a page whose actual job you cannot see.

2. No Clicks

Impressions only. No clicks, no CTR, no position. Search Engine Journal’s read on the launch put it plainly: the report shows where your content appears, but not why it was used or whether the visibility mattered.

3. No API

There is an EXPORT button in the top right of that screenshot, so you can pull a CSV by hand. What you cannot do is pull it programmatically, and that distinction is not documented anywhere I could find, so I tested it. On August 12, 2026, I sent searchAnalytics.query requests against my own verified property with six candidate values for the type parameter:

web:            OK    rows=5  impressions=3050
aiMode:         ERROR Invalid value at 'type'
AI_MODE:        ERROR Invalid value at 'type'
generativeAi:   ERROR Invalid value at 'type'
GENERATIVE_AI:  ERROR Invalid value at 'type'
aiOverview:     ERROR Invalid value at 'type'
genAi:          ERROR Invalid value at 'type'

Every generative-AI value is rejected by the enum. The endpoint still accepts only web, image, video, news, discover, and googleNews.

The practical consequence: the Generative AI report is a manual artifact. A human has to log in and click EXPORT every time. You cannot schedule it, you cannot alert on it, you cannot join it to revenue in a warehouse, and no agency can put it in a monthly client deck without someone doing it by hand. Every AI-visibility dashboard being sold right now is still inferring this number, because Google is not shipping it to anyone’s pipeline.

What 90 Days of Prompt-Shaped Queries Actually Did

Since the report withholds queries, the only usable proxy is query shape. AI Mode and AI Overview queries skew long and conversational, so I classified every query on chudi.dev as prompt-shaped if it ran eight words or longer or ended in a question mark, then compared it to everything else.

Pulled from the Search Console API on 2026-08-12, covering 2026-05-14 to 2026-08-12, 4,856 query-page pairs:

Query shape	Pairs	Impressions	Clicks	CTR
Prompt-shaped, positions 1-10	527	6,235	4	0.06%
Keyword-shaped, positions 1-10	3,376	78,580	730	0.93%

524 of those 527 prompt-shaped pairs returned exactly zero clicks across 90 days, on 6,139 impressions. Ranking in the top 10 for a prompt-shaped query was worth roughly one fourteenth of ranking in the top 10 for a keyword.

Some individual rows are stark. One page sat at position 3.1 for “what’s the minimum viable aeo optimization?” across 125 impressions and earned nothing. Another held position 4.0 for “aeo platform that shows which urls chatgpt cites from my site” over 160 impressions, also nothing. Those are buying questions, asked at the exact moment somebody is shopping, answered on a page that ranks, generating zero visits.

Two honest caveats. Query shape is a proxy, not a label, so some of those rows are ordinary long-tail searches rather than AI surfaces. And an AI answer counts an impression differently from a blue link, so a low CTR here may mean something else entirely. The missing query dimension is what makes that unresolvable. Nobody outside Google can currently tell the two cases apart.

Zero Clicks Is Not Automatically a Loss

If your page is the source an AI answer was built from, the answer did its job and the reader did not need you. That is fine when your goal is brand presence, and fatal when your goal is a lead.

The failure case is narrower and worse than a low CTR. It is being used as an uncredited source while a competitor gets named in the same answer, which is a thing you cannot see in this report at all, because the report only knows Google’s surfaces. ChatGPT, Perplexity, and Claude are the majority of the AI-answer surface for most B2B queries, and none of them appear in Search Console. If you want the shape of the wider problem, domain authority stops predicting citations once you leave blue links, and what actually predicts an AI citation is a different, more mechanical list.

What to Do With the Report This Week

Four steps, in order, none of which take longer than an afternoon.

1. Screenshot the baseline. No API means no history. Capture the URL breakdown today, because the report’s own retention window will eventually roll past May 18, 2026 and you will have nothing to compare against.
2. Cross-reference the top AI-impression URLs against your standard report. Any URL that is heavy in AI impressions and thin in clicks is a page whose job has changed from traffic to citation. Rewrite its opening 80 words as a standalone, quotable answer.
3. Build the prompt-shaped filter yourself. Filter the standard performance report to queries of eight or more words, export it, and compute the CTR gap against your short-query baseline. That number is your real AI exposure, and it is the closest thing to the missing query dimension.
4. Measure the engines Search Console cannot see. Google’s report covers Google. Getting cited by ChatGPT and Perplexity runs on separate crawl access and separate citation mechanics, and no Google dashboard will ever report on it.

The Report Shows Where. It Never Shows Why.

All four steps end in the same place. You can see which pages appear in AI answers. You cannot see which question they answered, whether the answer named you, or what the pages that beat you did differently. Google shipped the symptom and withheld the diagnosis. The API probe above says the diagnosis is not reaching your tooling either.

Closing that gap is mechanical, not mysterious. Run the actual prompts against the actual engines. Record which URLs each one cites. Diff your page against the pages that got cited instead. That is the whole method behind the AI Citability Audit I run as a fixed-price engagement, and it is deliberately reproducible, because the useful output is not a score, it is the list of specific reasons a specific engine reached for somebody else’s page.

If you would rather build it yourself, start with the six factors answer engines use to choose sources and work backward from there. The report Google just gave you tells you where to point that work. It will not do the work.