On 3 June 2026 Google launched a new report in Search Console: generative AI features. It shows the number of impressions in AI Overviews – how many times Google cited a given page as a source in the answer it generates above the search results. The report is in Beta and is not in the API, so the data has to be read by hand from the panel.
It is worth pausing on what „impression” means here, because it is not the same word as in the ordinary performance report. There, an impression means the page was somewhere on the list of results – possibly in thirtieth place, which nobody scrolled to. Here it means that the model picked it out of every page on the internet and named it as a source for an answer. An AI answer carries a handful of citations, not ten or a hundred. That is a far narrower sieve.
I looked into it on two sites I run. Both have data, both are growing. Here is what comes out of them.
The figures
Data read on 27 August 2026, covering the period since the report launched:
The toolroom (metal machining to order, Krakow):
- 470 impressions in AI answers, out of 9,080 impressions in total
- AI share: 5.2%
- AI cites 25 different pages of the site
The air conditioning company (local services, Krakow and the surrounding area):
- 199 impressions in AI answers, out of 5,290 impressions in total
- AI share: 3.8%
- AI cites 20 different pages of the site
One caveat before we go further: this is not extra traffic. Those 470 impressions sit inside the 9,080, they are not added to them. The report separates out data that was already being counted.
The trend is rising in both cases. For the toolroom the chart starts at zero to two impressions a day in early June and reaches a dozen or so a day in the second half of August.
The narrow service nobody describes wins
The most interesting figure in the whole exercise belongs to the air conditioning company. Out of 199 impressions in AI, 108 go to a single page: the warranty inspection. That is 54 per cent of the site’s entire visibility in AI answers, from one narrow service.
Why that one? Because almost nobody describes it separately. Air conditioning companies usually have a single „Services” page where the warranty inspection is one bullet in a list, or is missing altogether. When the model gets a question along the lines of „what does an air conditioning warranty inspection cover”, it has very few sources to choose from. A page that gives the service its own address and a few paragraphs of specifics lands in that narrow pool.
It is the same principle that has worked in classic SEO for years: it is easier to be first for a narrow phrase than twentieth for a broad one. In AI answers it works harder, because the model does not show ten results, it cites a handful of sources.
The practical conclusion: the services your competitors treat as an afterthought are the cheapest way into AI answers. A warranty inspection, an anti-fungal treatment, one specific technical operation, an unusual variant of a service. The narrower and more specific it is, the smaller the pool of competitors.
AI cites what already ranks
At the toolroom the distribution looks different. The most frequently cited page is EDM machining, with 115 impressions. I checked how the same page does in ordinary results: the phrase „EDM machining Krakow” sits at position 5.7 and is one of the few service phrases that genuinely collects clicks.
That is no accident. The next cited pages are the home page, stamping and blanking dies, and CNC milling and turning. All from the upper end of what ranks on that site at all.
The conclusion is a comfortable one for anybody already investing in SEO: visibility in AI is not a separate channel next to search optimization, it is a by-product of it. You do not need to buy a separate „AI optimization” service. You need rankings.
The foreign-language version gets cited too
The English version of the toolroom site collected 54 impressions in AI, making it the fourth most cited page across the whole site.
There is confirmation of that in the ordinary data: Search Console carries the query injection molds near=kraków, with shopping sites excluded, at position 2.0. Somebody, or something, is looking for a supplier from Krakow in English, and the site is in the answer.
For a manufacturer that serves clients abroad, that is a concrete argument for keeping a foreign-language version, even when the Polish statistics make it look barely visited.
What the report does not show
Here it pays to be honest, because it is easy to read more into these numbers than they hold.
- Impressions, but no clicks. There is no CTR and no position, so there is no way to say how many people reached the site from an AI answer. The number of citations is a strong signal in itself, though: it says how many times Google judged the page a credible enough source to name it.
- No queries. There is no query dimension, so there is no telling which questions the company is cited for. You can guess from which pages get cited.
- No history. The data starts in May 2026 and is not backfilled, so there is no before-and-after comparison to show.
- Google only. The report covers AI Overviews and AI Mode. It does not cover ChatGPT, Perplexity or Copilot. Traffic from those arrives as referrals, measurable only in your own analytics or in the server logs.
That last one is the biggest limitation in practice. Neither of the sites described here has analytics installed today, so traffic from ChatGPT or Perplexity is entirely invisible.
It is worth lowering expectations of analytics right away: it will not tell an AI answer visit apart from an ordinary click on a Google result. In both cases it is the same organic traffic from the same domain, and Google does not pass on the information that the click landed on a citation in an AI Overview. That applies to every tool, not one in particular. So analytics closes only one side of this puzzle: referrals from assistants outside Google.
What to do about it
Three things, cheapest first:
- Open the report. Search Console, Performance, Search results, filter for Generative AI. See which pages get cited. The distribution alone is information: it shows what Google considers the most substantial content on your site.
- Describe separately the services your competitors skip. Not another text about what air conditioning is, but a page about one narrow thing, with specifics: what it covers, how long it takes, what happens afterwards.
- Check whether you have analytics at all. Without it you will not see traffic from ChatGPT, Perplexity or Copilot, because those are ordinary referrals from other sites rather than traffic from Google. In Google Analytics they show in the acquisition report as the source
chatgpt.comorperplexity.ai. It is worth splitting them into their own channel group so they do not get lost among the other referrals.
One separate thing worth knowing: the robots.txt file can block AI bots, sometimes without the site owner knowing, because some hosting panels set it by default. Blocking GPTBot or ClaudeBot is a legitimate decision, but it should be a decision rather than an accident. It is worth checking what is in that file today.
In summary
Visibility in AI answers is a few per cent of impressions for now, but the share grows month on month and it can already be measured. The two things that come out of this data most strongly: narrow, specific service pages get cited disproportionately often, and the rest of AI visibility follows the positions in ordinary results.
Both lead to the same conclusion as classic SEO. The only new part is that the effect is now visible in a place that did not exist a year ago.