Search Console Now Shows AI Impressions: Reading the Generative AI Report

For two years the honest answer to “how much of our traffic comes from AI answers” was that nobody could tell you. On 3 June 2026 Google closed part of that gap by adding generative AI performance reports to Search Console, covering AI Overviews and AI Mode in Search and generative features in Discover.

It is a real improvement and a partial one. The reports show impressions. They do not yet show clicks, click-through rate or the queries behind them. Knowing exactly what that permits you to conclude — and what it does not — is the difference between a useful quarterly review and a confident wrong decision.

What the report contains

What launched is a dedicated view of data that also remains inside the overall performance report, where Google says it continues to be tracked. That distinction matters as soon as you start comparing figures, and we return to it below.

Availability is phased: Google began with a subset of site owners while testing. If the report is not in your account yet, that is the rollout rather than a problem with your site, and there is nothing to configure to accelerate it.

What an impression here actually means

The word is borrowed from the traditional report and behaves differently, which is where most of the misreading starts.

In classic search, an impression means a link to your page appeared in a results list a person saw. In a generative answer, it means your page was surfaced as a source alongside an answer that was written for the user. The person may have read the answer and left satisfied. They may never have looked at the citation. Some of these impressions represent influence with no possibility of a visit.

Three consequences follow directly.

• Impressions cannot be converted into an expected traffic figure. There is no reliable multiplier, because click behaviour in AI answers varies enormously by question type.
• Growth in impressions is not automatically good news. It can indicate that your pages are being used to answer questions people previously came to your site to resolve.
• Zero impressions is not evidence of absence. The report covers Google’s surfaces. Other assistants do not report anything at all.

Four questions the report answers well

Within its limits, the data is genuinely decision-grade for a small number of questions. These are worth building into a quarterly review.

1. Which of our pages are being used as sources? The page breakdown is the most valuable column. It usually surprises people: the pages that get cited are frequently not the pages the marketing plan was built around.
2. Are we present at all in our category? A site with meaningful traditional visibility and near-zero AI impressions has a specific problem — its content is structured for browsing rather than for answering, which is the pattern examined in our note on which citation factors the evidence actually supports.
3. Where geographically? The country breakdown matters for a business selling in several markets, because AI features have rolled out unevenly and a gap in one country may be a product of availability rather than of content quality.
4. Did something change on a date we can name? The date breakdown lets you line up a drop with a deployment, a content change or a Google announcement. That is the same diagnostic discipline as in our note on finding the cause of a traffic drop.

The opt-out toggle, and why it is rarely the answer

Alongside the reports, Google added a control that lets a site be excluded from AI Overviews, AI Mode and generative features in Discover while remaining in ordinary search results. It sits at property level in Search Console settings, and it exists because of a binding order from the United Kingdom’s competition regulator rather than because publishers asked politely. It arrived in the UK first, with Google honouring changes from 17 June 2026, and began appearing outside the UK during July.

Before anyone reaches for it, four things are worth knowing.

It removes visibility, not just extraction. Opting out does not mean your content is protected while your competitors’ is used. It means an answer will still be generated, from other sources, and you will not be among them.

It does not cover everything with the same brand on it. The control addresses generative features in Search and Discover. It does not govern Gemini apps, and it does not govern model training — a separate control handles that — so “we opted out of AI” is a considerable overstatement of what the toggle achieves.

It is currently a whole-property decision. Page-level controls are required by March 2027 under the same order, but today the choice applies to the entire site, not to the ten pages you actually care about.

It is reversible but not instant. A change takes a few days to process before content drops out of the features, and coming back is equally unhurried. Treat it as a considered policy change with a review date, not as an experiment you run for a fortnight. It is not, at least, a ranking signal: the setting does not affect your position in ordinary results.

The businesses for which opting out is defensible are narrow: publishers whose product is the answer itself, and sites under a licensing arrangement that requires it. For a company that sells goods or services, being absent from the summary that a buyer reads before choosing a supplier is a strange thing to arrange deliberately. The related question of which crawlers you allow at all is a separate decision, covered in our note on what to check in robots.txt after the AI bot defaults changed.

How to use this data this quarter

Four steps, and none of them requires new tooling.

Start by exporting the page-level report and comparing it with your list of commercially important pages. Where an important page has no AI impressions and a competitor’s equivalent presumably does, that is a content-structure task: a direct answer near the top, a clear definition, specific numbers, a scannable structure.

Next, look at what is being cited that you did not plan for. An old explainer pulling steady AI impressions is telling you something about demand that your keyword work may have missed — the input the page map in our note on turning a query list into a page map is designed to capture.

Then set a baseline and stop looking daily. This is quarterly data. Hourly granularity is available and, for almost every business, actively harmful to sound judgement.

Finally, keep measuring the thing that pays. Referral sessions from AI assistants, enquiries, and orders are still the outcomes; impressions are an early indicator, not a result. Where your analytics can separate assistant referrals, that segment tends to be small in volume and unusually strong in intent, and it deserves its own line in the report you actually read.

Reading it next to the traditional Performance report

One line in Google’s announcement is easy to skim past and it governs how these numbers may legitimately be used: this data is included in the overall performance report, where it continues to be tracked. What launched in June is a dedicated view of it, not a separate accounting of it. That rules out the first thing many people try.

Do not add the two together. AI impressions plus web search impressions is not your total visibility. Some of the same impressions sit in both figures, so a combined number is inflated by an amount you have no way to calculate. Present them side by side, or present the overall figure — never the sum.

The new report does not explain a fall in the old one. Because the generative data was already inside the overall totals, splitting it into its own view changed nothing about them. If your overall impressions dropped in June, the launch is not the cause, and looking for the explanation there costs you a diagnostic cycle you could have spent on the real one.

Align the date ranges before comparing anything. The dedicated view’s history begins when Google started surfacing this data, which is not where your main report’s history begins. A year-on-year window in one panel against a two-month window in the other produces a difference that is entirely an artefact of the axis.

Ignore the hourly granularity. It is offered down to the hour, alongside daily, weekly and monthly. On a signal this thin, hourly data manufactures the appearance of events out of ordinary variance — someone will bring you a chart of a Tuesday afternoon and ask what happened, and the answer will be nothing. Monthly is the right unit; quarterly is the right review.

The most valuable thing to do in the first month is dull: write down a baseline. One impression figure per month for each commercially important page, kept wherever you keep the rest of your reporting. Google has said additional metrics may be added over time, and on the day clicks appear, their worth will depend entirely on having a before to compare them with. That is not something anyone can reconstruct afterwards.

What it does not replace

Two gaps remain, and both are worth covering with something other than Search Console.

The first is everything that is not Google. Other assistants provide no equivalent reporting, so your visibility there is measurable only through referral traffic and, imperfectly, through server logs showing their crawlers. If a meaningful share of your audience uses one, the Search Console report describes part of the picture and presents it as the whole.

The second is the answer itself. The report tells you that you were a source; it does not tell you what was said about you. Periodically asking the major assistants the questions your customers ask, and reading the result, remains the only way to catch a summary that is confidently wrong about your prices, your coverage or your services.

Key takeaways

• The reports launched on 3 June 2026 and cover AI Overviews and AI Mode in Search plus generative features in Discover, rolling out to a subset of site owners first.
• Impressions only. No clicks, no click-through rate, no query data yet; Google has said only that additional metrics may be added over time.
• It is a view, not a separate bucket. The same data remains inside the overall performance report, so never add AI impressions to web search impressions.
• An AI impression is not a near-visit. It records that your page was a source, including in answers that ended the user’s journey.
• Use the page breakdown first. It reliably reveals which content is doing work you did not plan for.
• The opt-out toggle removes you from the answer, not the answer from the internet, applies to the whole property, and does not cover every product carrying the same brand.
• Treat it as quarterly evidence, and keep judging outcomes by enquiries and orders rather than by impressions.