Google will report your AI impressions. Not one click.
Since 3 June, Search Console has counted how often a link to your site turned up inside an AI Overview or an AI Mode answer. It will not tell you what happened next. That sounds like a defect — until you notice that impressions and position are the two things capable of moving first when SEO work starts, and clicks are the last.
What it is. On 3 June 2026 Google added a report to Search Console for its AI surfaces: AI Overviews, AI Mode, and generative AI in Discover.
What is in it. Impressions. Split by page, country, device and date. No clicks, no click-through rate, no average position, no queries.
Why that is not a scandal. An impression tells you something early. A click tells you something late. In the first weeks of any SEO or GEO programme, impressions and position are the only numbers physically able to move. Reporting the report as traffic is the mistake — not the report.
What you get from the next thirteen minutes:
- plain definitions of the five words this report quietly assumes you know;
- Google’s own wording on what counts as an impression, which makes the number smaller than it sounds;
- the outside figure that explains the missing click column;
- our own site opened up week by week — what seven days of work moved, what it did not, and the number we would rather not print;
- a way to read it, and a list of what it will never tell you.
What Google actually shipped
On 3 June 2026, Google gave Search Console a report of its own for generative AI: a view of how often your pages showed up inside AI Overviews, AI Mode and generative AI in Discover.
It arrived first for a slice of properties in the United Kingdom. By 23 June, Search Engine Land had it appearing for sites in the United States, India, Switzerland and elsewhere. So if it is not in your account yet, that is the rollout. It is not something you configured wrong.
Nothing about your traffic changed the day it landed. Google’s line has been that these impressions were already inside the main Performance report all along; what is new is that you can look at them on their own. Nothing was added. Something was separated out.
| In the report | Not in the report | |
|---|---|---|
| Metrics | Impressions | Clicks, click-through rate, average position |
| Dimensions | Page, country, device, date | Queries |
| Surfaces | AI Overviews, AI Mode, generative AI in Discover | ChatGPT, Perplexity, Copilot — anything outside Google |
| Comparison | Period over period, on impressions | Against a competitor, or against any benchmark |
This is a visibility report, not a traffic report. It answers one question — was my page shown inside an AI answer? — and politely declines every question that follows.
Five words the report assumes you know
Most of the confusion around it is vocabulary rather than measurement. Here are the definitions we use with clients, and the ones the rest of this article runs on.
| Term | What it means, in one sentence |
|---|---|
| AI impression | Your link was visibly shown inside one of Google’s AI answers — and if it was tucked behind a control, only once somebody opened that control. |
| AI Overview | The generated answer box sitting above the ordinary blue links on a normal results page. |
| AI Mode | The separate conversational surface, where the answer is the page rather than a box on top of one. |
| Coverage | How many different pages of yours get shown at all. A count of URLs, not a count of impressions — and the more useful of the two. |
| Average position | Where your links sit on average in the classic results. It is a mean, so it moves when new pages arrive, not only when old pages rise or fall. Remember that one; it comes back. |
What counts as an impression
Google’s Search Advocate John Mueller answered this on 25 June 2026, quoted by Search Engine Journal:
“The impressions are based on links to your site being shown in AI Overviews / AI Mode.”
And on the links that hide behind a chevron, folded away until a reader asks to see the sources:
“If something needs to be ‘activated’ to see the link, it would only count when users do that.”
Put those two sentences together and the metric shrinks in your hands. An impression means your link was on a screen. Sometimes that is a proper citation with your name on it. Sometimes it is a favicon the size of a fingernail, in a row of six other favicons. Sometimes it exists only because one curious reader tapped to see where the answer came from.
In practice: a rising impression count means Google is willing to show you more often. It does not mean more people saw you, and it certainly does not mean more people wanted you.
An AI impression is a willingness signal from Google, not an attention signal from a human. It is the earliest evidence that your page is eligible. That is all it is, and it is still worth having.
The click column is empty. Here is why.
Similarweb went and measured what AI assistants actually do with links. Greg Jarboe reported the figures in Search Engine Journal on 30 July 2026, and one line does most of the work:
“Only 6.8% of ChatGPT answers in the U.S. included a link to an external source as of May 2026.”
A year earlier that number was about 1%. So it is climbing, and anyone still saying assistants never cite anybody is a year out of date.
The level is the problem, not the direction. More than nine answers in ten give the reader nothing to click. There is a layer that answers, and beneath it a layer that links, and the top layer is winning.
The two are not separate audiences, either. Between March and May 2026, the same dataset found that 461 million of ChatGPT’s 494 million users — 95% of them — also used Google. One audience. Two surfaces. Shown your name on one, clicking through on the other.
Google’s AI surfaces are not ChatGPT and the rates will differ. But a report that counts only impressions is not hiding a number Google finds awkward. It is describing a place where the second number is genuinely small — and, by accident or on purpose, measuring the half of the funnel that moves first.
What seven days of SEO and GEO look like in the data
We are an agency writing about our own numbers, so here is the whole context, including the parts that do us no favours. For years our SEO work went into client sites and never into ours. We have taken on few SEO projects — but the ones we took reached the first page, and several the top three. Our own domain was the one nobody was paying for, so nobody worked on it.
GEO is a different story and we are not going to pretend otherwise. Nobody has years of it. The surfaces are eighteen months old. We started working it properly a week ago, on ourselves first, which is the honest order to do it in.
The rebuilt cittago.com went live on 6 July 2026. Everything until the last week of that month was foundation rather than ranking work: structure, three languages with hreflang, structured data, a redirect map that retired 127 dead URLs inherited from the old WordPress site, generated share cards, and a publish gate that refuses to deploy a page missing its title, canonical, hreflang or schema.
None of that is glamorous and none of it moves a ranking on its own. It is the part you do so that the next part has something to stand on.
What actually changed on 25 July
- Topics came from query data instead of a calendar. We pull the Search Console API every week and pick subjects where we already have impressions and no position worth having.
- One article a day, in three languages, each written natively. Not translated — rebuilt from an outline, with its own opening, its own order and its own questions. People type differently in Italian than they do in English, and a translation carries the English phrasing across with it.
- Question-shaped headings and a real FAQ on every article. An assistant can lift a block. It cannot lift an essay.
- Internal links from every new piece to a service page and two older articles. Cheap, dull, and still skipped by most sites.
- Structured data generated at build and checked before deploy — organisation, author, breadcrumbs, FAQ, ratings.
- Weekly reading of the API rather than the interface. The reason why is a few paragraphs down, and it changed what we concluded about our own performance.
Seven days is not an SEO result. It is long enough for the two early indicators to move, and they did.
| Search Console, 7 days | 18–24 July | 25–31 July | Change |
|---|---|---|---|
| Impressions | 208 | 746 | ×3.6 |
| Average position | 43.6 | 12.3 | −31.3 places |
The generative AI report contains impressions only. It has no position and no queries. Every position and every query below comes from the classic Performance report for the same property over the same days. We use the AI report for what it gives and the classic one for the rest. Anybody quoting an “AI average position” is quoting something Google does not publish.
Coverage moved first, exactly as it should
Impressions are the leading half of the chain — the first thing capable of moving when work starts. Here is ours, on both windows.
Thirty times more of the site is being shown than three months ago. Coverage is the first thing that moves, and it moved.
cittago.com · Search Console API, 28-day and 7-day windows, May to July 2026
| Window | Period | Impressions |
|---|---|---|
| 28 days | 4–31 May 2026 | 61 |
| 28 days | 6 June – 3 July 2026 | 742 |
| 28 days | 4–31 July 2026 | 1,842 |
| 7 days | 18–24 July 2026 | 208 |
| 7 days | 25–31 July 2026 | 746 |
The same story on a longer window — and the trap inside it
Comparing 28-day windows on the same property: 61 impressions in the window ending 31 May, on the old site. 742 in the window ending 3 July. 1,842 in the window ending 31 July. Thirty times more of the site is being shown than three months ago.
The window you choose decides the answer
Same property, same summer. Switch the reporting window and the line changes direction — which is why the interval you read is a decision, not a setting.
Read in 28-day windows, average position went from 7.5 to 24.9. On this view the site looks like it fell apart in July.
cittago.com · Search Console API, classic Performance report, May to July 2026
| Window | Period | Average position |
|---|---|---|
| 28 days | 4–31 May 2026 | 14.8 |
| 28 days | 6 June – 3 July 2026 | 7.5 |
| 28 days | 4–31 July 2026 | 24.9 |
| 7 days | 18–24 July 2026 | 43.6 |
| 7 days | 25–31 July 2026 | 12.3 |
Because average position over those same windows went the other way: 7.5 to 24.9. Worse, on paper. If a client sent us that chart we would expect them to be worried.
It is not a decline. Roughly a hundred new pages entered the index down at the bottom and dragged the mean with them. The old site ranked acceptably for a handful of URLs. The new one appears for many, most of them published days ago and nowhere near page one. A mean cannot tell those two situations apart. That is the whole of it.
Bottom line: while you are actively publishing, read weekly windows. On the monthly view alone we would have reported the exact opposite of what happened.
Two queries from that week, both landing on articles published days earlier: “chatgpt ads europe” at average position 4.2, and “google ads 17 august” at 5.7. Neither is a big term. Both are the narrow, dated, specific kind of question a small site can actually win, which is the entire strategy in two examples.
What analytics says about the same weeks
Search Console tells you what Google showed. Analytics tells you what people did. May against July, same property:
| GA4 | May 2026 | July 2026 | Change |
|---|---|---|---|
| Organic search sessions | 23 | 51 | +122% |
| Engaged sessions | 68 | 137 | +101% |
| Pages viewed | 453 | 822 | +81% |
| Avg. engagement per user | 13.6s | 45.8s | ×3.4 |
Total users are down across that comparison, and we are not going to bury it in a footnote. Two reasons, both structural. The old site’s direct traffic did not survive the rebuild — people who used to arrive out of habit lost their bookmarks and their old URLs. And there is a hole in the data between 2 June and 6 July, while the old site was coming down and the new one was not up yet.
Any comparison spanning that gap is comparing two different websites. We would rather say so than pick a window that quietly hides it.
What the engagement figures do say is that the people arriving now stay three and a half times longer. On a site whose entire job is to make an argument, that is the number we would defend in front of a client.
Clicks are still negligible. Not modest — negligible. That is what a first month looks like, and anyone telling you otherwise is selling something. What we tell clients before they sign is that SEO and GEO pay somewhere between six and twelve months in. Coverage moves in week one. Position moves in month one. Clicks move once the positions stop being new.
The order things move in
This is why a report full of impressions and nothing else is less useless than it first looks. It measures the early half of a four-stage chain.
It behaves rather like hiring. Within a week you can tell whether the pile of applications got better. You cannot tell within a week whether you hired well. Both are real signals; they just arrive months apart, and confusing one for the other is how people end up disappointed.
What moves, and when
Tap a stage. Each one moves on its own clock, and only the first is visible in Google’s AI report — which is the whole reason it is so easy to misread.
Coverage is how many of your pages appear at all. It moves within days of publishing, and the AI report shows it directly — this is the stage it was built for.
cittago.com · timings are Cittago’s working assumption from client programmes, not a Google figure
| Stage | Moves in | In the AI report? | Driven by |
|---|---|---|---|
| Coverage | Days | Yes | Indexing |
| Position | Weeks | No | Content and links |
| Clicks | Months | No | Titles and intent |
| Enquiries | Later still | No | Your sales record |
- Coverage. How many of your pages appear at all. Moves in days, once indexing catches up.
- Position. Where they appear. Moves in weeks, and moves loudly at first, because there is an enormous amount of room below rank thirty.
- Clicks. Move once positions hold still and the title earns the tap. Months.
- Enquiries. The only one that pays a salary. Later than that.
The important point: read the AI report against stages one and two and it is a genuinely useful instrument. Read it against stage three or four and it will lie to you by omission, every month, in a way that feels exactly like progress.
What to put next to the report
- Read it as coverage, not traffic. The useful question is which pages appear at all, and whether that set is growing. The impression count is the least interesting number on the screen.
- Pair it with referral data. Watch the hostnames in your analytics: chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com. That is where a click actually lands, it belongs to you, and it covers the assistants Google will never report on.
- Use weekly windows while you are publishing. A 28-day average is watered down by every page that arrived this week. We misread our own data on the monthly view and only caught it on the weekly one.
- Watch the page mix, not just the total. If the pages picking up AI impressions drift away from your service pages towards your oldest posts, that is worth knowing — and it is one of the few things this report is genuinely good at showing.
- Keep one business number in the same file. Enquiries, bookings, calls. Everything above is an input. Only that one tells you whether any of it paid.
What it will not tell you
- Efficiency. No clicks and no click-through rate means there is no ratio to improve against.
- Which question you answered. There is no query dimension, so a buying question and an idle one look identical.
- Anything outside Google. ChatGPT, Perplexity and Copilot do not report to Search Console, and nothing suggests they ever will.
- Whether a rival appeared beside you. The report is a mirror, not a market view.
- Whether the impression was worth having. A favicon someone had to expand and a full citation with your brand name count exactly the same.
Questions we get
Does Search Console show clicks from AI Overviews and AI Mode?
No. The generative AI report holds impressions and nothing else. Clicks that happen to come from an AI surface are folded into your overall Performance totals, where you cannot pull them back out.
What exactly counts as an AI impression?
Definition: an AI impression is counted when a visible link to your page appears inside one of Google’s AI features — and, where the link sits behind a control, only once somebody opens it.
That is Mueller’s wording, not our reading of it.
Why does my average position get worse while impressions go up?
Nearly always because new pages entered the index low down and pulled the mean with them. It is arithmetic, not decline.
Check whether the number of distinct pages picking up impressions grew over the same period. If it did, you have your answer. Ours did, which is why there are four paragraphs about it further up.
How long before SEO and GEO work shows up?
Coverage in days. Position in weeks. Clicks in months. Enquiries later than that.
On our own site, one week moved average position from 43.6 to 12.3 while clicks stayed where they were. Plan for six to twelve months before the work pays, and treat anything earlier as luck rather than forecast.
Does ChatGPT or Perplexity traffic appear in Search Console?
No. Search Console covers Google surfaces and only Google surfaces.
If you want to know whether the assistants send you anybody, the referral hostnames in your analytics are the only first-party measurement you have: chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com.
Should I block my content from AI Overviews and AI Mode?
Google shipped a control for exactly that, reported by Search Engine Land in June 2026, and for most small companies the arithmetic is grim. You lose the impressions. Nothing in the data says the clicks reappear somewhere else.
Remember the 95% overlap: it is the same people meeting you on both surfaces. We would reach for it on price lists, paid content, and documents you did not quite mean to publish. That is a short list.
Will opting out hurt my ordinary Google rankings?
No, and Google has been unusually direct about it: the opt-out is not used as a ranking signal for search results outside the generative AI features.
One practical catch — the control works at domain level today. Page-level control does not exist yet, and the UK competition regulator has given Google until March 2027 to build it. So for now it is all or nothing.
Is the report available in every country?
It follows the property, not the country your site is about, and the rollout has been gradual: the United Kingdom first on 3 June 2026, then the United States, India, Switzerland and other markets by 23 June.
The only reliable check is to open Search Console and look. If it is not there yet, your referral data still works today and does not wait for Google.
Can I compare my AI impressions with a competitor's?
Not in Search Console, which has only ever shown you your own property.
A competitive picture has to be built by hand: run the questions your customers ask across the assistants and write down which names come back. Tedious, and still the cheapest useful thing most companies are not doing.
Where we would start
If you are looking at that report for the first time and want one thing to do this week, do this: open it, sort by page, and write down which of your service pages appear. Not blog posts. Service pages.
If the answer is none, your problem is not AI visibility. It is that the pages describing what you sell do not answer a question anybody asks out loud.
That is the first thing we check, and it was the first thing we fixed on our own site. If you would like it done properly — the questions your customers actually type, run across the assistants, with the names that come back and the pages to fix — that is what our SEO & AI search work is. It is the same method we are running on ourselves, in public, a week at a time. Book half an hour and we will read your report with you. If there is nothing to do yet, we will say so.
Related reading: why zero-click search is not the end of anything, and why AI describes your brand correctly and still recommends somebody else.
Last updated: 1 August 2026. Google has not announced click reporting for this view. Our own figures come from the Search Console and GA4 APIs for cittago.com; we refresh them here monthly for as long as the programme runs, including the months where nothing moves.
Impressions rising. Anything else?
Half an hour, your Search Console and your referral data open side by side. We tell you which stage you are actually at, and what moves next. No pitch, just the read.
