Google will tell you if AI cited you. Not if anyone came.

A long-requested report arrives half-built

Since AI Overviews began appearing at scale, site owners have asked Google a simple question: is my content being used? For eighteen months the official tools had no answer.

That changed in 2026. Search Console began reporting impressions across AI Overviews, AI Mode and AI features in Discover, broken down by page, country, device and date. Data starts on 18 May 2026, with no historical backfill, so there is no comparison available with the period before the features expanded.

Google added a control at the same time, letting owners decide whether their site appears in and helps ground responses in generative Search features — the first switch of its kind, and one with consequences most sites have not modelled.

What the report withholds

The reporting is impressions only. There is no click data, no click-through rate, no average position and no query-level breakdown. Search Console's natural-language query analysis does not yet cover AI features either.

The result is a metric that confirms appearance and nothing else. A page can be shown to have featured in an AI answer with no indication that anybody subsequently visited the site.

Set against what the independent studies show about click compression inside those same features — a 38% to 61% reduction depending on method — the missing half is precisely the half that determines whether an appearance had any commercial value.

The reporting hazard this creates

Performance marketing has been built for two decades on a chain running from impression to click to conversion. AI features break the chain at the first link, and Google's reporting does not yet repair it.

This produces a familiar failure. When a number becomes available it tends to become a target, whether or not it means anything. AI impressions are easy to report, straightforward to grow and, on their own, impossible to connect to revenue.

The predictable outcome is a season of monthly reports showing a rising line that nobody can tie to a customer — and, eventually, a difficult conversation when someone asks what the line bought.

Three honest substitutes

Workarounds exist while the reporting matures. Each should be labelled as a substitute rather than presented as a measurement.

The first is branded and direct traffic. If a company is being described accurately in AI answers, a proportion of those people will later search its name or type the address. That movement is already visible in standard analytics, and it has the advantage of measuring an outcome rather than an exposure. Watched as a trend over quarters, it is the closest available proxy for whether AI visibility is producing demand.

The second is manual auditing. Asking the AI systems the questions a buyer would actually ask, at regular intervals, and recording whether the business is named, described correctly and cited. It is slow, it does not scale, and it measures the thing that matters. Because citation patterns diverge sharply between systems — only 11% of domains cited by ChatGPT overlap with those cited by Perplexity — the audit has to cover several of them separately.

The third is asking. Adding a question to enquiry forms about how someone came across the company is imprecise, undercounts badly and is currently more informative than the dashboard.

What to do with the new control

The opt-out deserves more thought than it is receiving. Removing a site from generative features protects it from being summarised without a click, and simultaneously removes it from answers where citation would have raised both organic and paid click rates.

On the published evidence, opting out is difficult to justify for most businesses. Cited brands earn more clicks, not fewer. The exception is a site whose entire commercial model is selling access to information that an Overview can restate in full.

For everyone else the more productive response is to accept the exposure, measure it with the proxies above, and concentrate on being described accurately when the answer is generated.