# Search Console counts your AI impressions and cannot connect one of them to a visit > Google's generative AI report shipped worldwide with impressions and no clicks. Three separate changes hit AI citations in the same two weeks, and the report cannot price any of them. Canonical: https://brandonlazovic.dev/articles/ai-impressions-without-clicks/ Author: Brandon Lazovic Published: 2026-09-15 ## The short version - Google's generative AI performance report shows impressions only, with no clicks, no click-through rate, no average position and no prompt, and Google has committed to no date for adding them. - Three separate changes touched AI citation visibility in a two-week window, one confirmed and two only observed, and none of them arrived with a way to measure the effect. - I pulled my own property to test this and it could not answer, at 43 impressions across 70 days, which is a limit of my site's size and not of the report. - An impressions-only number can tell you that you appeared and cannot tell you that anything followed, so treat a rise in it as presence, never as traffic. Google finished rolling out its generative AI performance report to every property worldwide, and the report counts impressions. It does not count clicks. There is no click-through rate in it, no average position, and no record of what the person asked. In the same two weeks that the measuring instrument reached everyone, three separate things changed how often sites get cited inside Google's AI answers, and the new report cannot price any of them. > Google's own documentation offers impressions and nothing downstream of them, so a site can watch its AI presence rise all quarter without learning whether one reader arrived. ## What does the generative AI performance report actually show? The generative AI performance report shows how often your pages appeared inside Google's AI features, broken out by page, country, device and date. Google announced it on June 3, 2026 and describes it as covering both AI Overviews, the summary block above the blue links, and AI Mode, the full conversational search experience. [1] Its single metric is the impression, meaning your page was present in a generated answer. Google intends to add more, saying it plans to include click data over time based on feedback from site owners. No timeline came with that intention, and the report still carries a Beta label. **Confirmed, from Google's own documentation:** clicks, click-through rate, average position and the user's prompt are all absent today. ## What moved in the same two weeks? Three separate changes touched AI citation visibility between September 1 and September 8, 2026, and they differ sharply in how well established each one is. One is confirmed by Google itself. The other two rest on practitioner sightings, which is a much weaker footing. Grouping all three as equally solid would be the easy error here, so each one below carries its evidence and its status. On September 3, Google put Gemini 3.8 Flash into AI Mode on the model's release day. [2] **That launch is confirmed** by Google's own announcement. Testers then reported seeing fewer citations in AI Mode answers afterward, which **is observed and not confirmed**, resting on a small number of practitioners posting screenshots. [3] Two people's screenshots are a lead worth following, never a controlled before-and-after. On September 8, Google was spotted testing a narrower citations panel in AI Overviews, which would give cited publishers less visible space. [4] **Also observed**, from a spotted interface test, and interface tests reverse all the time without announcement. And on September 1, the report itself finished reaching every property. ## Why can't the report answer whether those changes cost you traffic? A change in citation visibility is a change in what happens AFTER the impression, and the impression is the only event this report records. Suppose the narrower panel ships and your citation gets less space. Your page still appeared, so your impression count is unchanged. Fewer readers notice the link and click it, and that difference lands entirely in the column Google has not published. The same holds for the model swap. If a newer model in AI Mode cites fewer sources, some sites lose impressions outright and will see it. For the sites still being cited, any change in how compelling that citation looks is invisible here. My read is that this is a timing accident and not a strategy, since the report was announced in June and these changes landed in September. The consequence is the same either way. **The instrument arrived unable to measure the thing that moved next to it.** ## What did my own property show? This property showed almost nothing, and I would rather report that than stretch it. I exported its generative AI performance data on September 8 using the report's own last-three-months filter, which returned 70 daily rows running from June 29 to September 6. Across those **70 days the property logged 43 impressions**, with only 26 of those days carrying a single impression and a busiest day of six. Desktop accounted for 38 and mobile for five. Across the window that would matter for the changes above, September's first week against late August, I count 9 impressions spread over 13 days. Three fell before September 1 and six after. Reading a citation-visibility change out of a three-to-six move at that sample size would be arithmetic dressed up as evidence, so I am not going to do it. One more limitation worth naming. My export stops on September 6 even though I pulled it on the 8th, because the report lags a couple of days behind. ## Is that the report's fault or my site's? My site is the reason, not the report's design, and keeping those two apart is the whole discipline here. This property is small, publishes on a narrow technical beat, and generates few AI impressions in absolute terms. A retailer with a hundred thousand products would pull a series with real statistical power out of the same report tomorrow. What that larger site would still not get is a click. Sparsity is a property of my site and it goes away as a site grows. The missing click column belongs to the report's design and applies identically to every property on earth, including the largest. Blaming the instrument for my small sample would be the sloppy version of this argument, and it would also let the real gap off the hook. I wrote in [most AI-driven site visits carry no referral tag at all](/articles/ai-referral-tag-undercount/) about how badly AI traffic is already undercounted before it reaches your analytics. That piece owns the measurement gap itself. What is new here is that the official instrument has now shipped everywhere, which changes the practitioner question from whether you will get data to what you should conclude from the data you got. ## What can you actually do with an impressions-only report? Treat the number as presence and stop there. An impression says a machine surfaced your page while answering somebody. That is genuinely useful, and it is worth having, because until this report existed most sites were inferring even that much from server logs and referrer strings. For one thing, you can see which of your pages get pulled into AI answers at all, which is a content signal you can act on without needing a click. For another, you can watch the page-level list change over time and notice when something that used to get cited stops. And where a page appears in AI answers but earns nothing in ordinary search, that mismatch is worth a look on its own terms. What you cannot do is put a value on any of it. A brief that says AI impressions rose 40% this quarter is reporting that Google's systems read you more often. It is not reporting traffic, and anybody who converts that number into a revenue estimate has invented the conversion rate that Google specifically declined to publish. ## What should you record now, before the click column exists? Start exporting the report on a fixed schedule, because when Google does add clicks you will want history to read them against. The report holds three months. Once a day passes out of that window it is gone from the interface, and no later feature will bring it back. Two habits cost almost nothing. Export the page-level table monthly and keep the files somewhere you will still have them next year, since the page list is the part most likely to reward a look backward. Write down the date of anything you change that could plausibly affect how machines read your pages, because a dated note beside a dated series is the closest thing to a controlled comparison you will get here. There is also a trap worth avoiding while you wait. Do not add these impressions to the impressions in your main Search performance report and present a combined total, because you would be summing two counts whose definitions and surfaces are not the same, and the result would be a number that describes neither. Keep the two series apart and labeled. **My honest read on the wait:** Google said it plans to add click data and attached no date, and I would plan on a year rather than a quarter. The company has published unfalsifiable traffic reassurances for eleven months, which I covered in [Google's billions-of-clicks claim](/articles/google-ai-clicks-open-web/), and a click column would make several of those claims checkable by anyone with a property. That is a reason for caution about the timeline, not a prediction. ## Sources 1. Google Search Central Blog: Introducing Search Generative AI performance reports in Search Console — https://developers.google.com/search/blog/2026/06/gen-ai-performance-reports 2. Google: Introducing Gemini 3.8 Flash and 3.8 Flash Cyber — https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/ 3. Search Engine Roundtable: Google AI Mode Gemini 3.8 Flash Not Link / Citation Friendly? — https://www.seroundtable.com/google-ai-mode-gemini-38-no-links-42011.html 4. Search Engine Roundtable: Google Tests Thinner AI Overview Citations Panel — https://www.seroundtable.com/google-thinner-ai-overview-citations-panel-42040.html