Measuring AI search visibility

Ask a marketer how much traffic AI is sending them and you will usually get one figure. There is no honest way to produce one, because Google files the two halves of AI search in two different places, and the place a visit lands has nothing to do with how the customer experienced it.

In Google Analytics, the default channel group now contains a channel called AI Assistant. Google defines it as the channel by which users arrive from sources like ChatGPT, Gemini, Deepseek, Copilot or Grok, and adds one sentence that changes everything downstream: it excludes Google AI Overviews and AI Mode. Those are defined into Organic Search instead, which Google describes as arrivals via non-ad links in organic search results, including AI Overviews and AI Mode.

So the same person, asking the same question, on the same afternoon, lands in a different channel depending on which assistant answered. Two of those rows are named after where the click came from. The third is named after nothing, because the click arrived with no name on it.

The practical consequence is that anyone quoting you a single percentage for AI traffic has either ignored the Google half or estimated it. Both are defensible if they are said out loud. Neither is defensible presented as a reading off a report.

Since 31 August 2026, Google has rolled out its generative AI performance report to all sites worldwide. It is the first official window into being shown inside an AI answer, and it is worth opening. It is also much thinner than most people expect.

The report covers impressions in AI Overviews and AI Mode. You can group those impressions by page, by country, by date and by device. That is the whole list. There is no clicks column, no query dimension and no average position, so the three numbers every SEO reaches for first are all absent. Google draws the data from the same Web search type as the ordinary performance report, so these impressions were always inside your totals; what changed is that you can now look at them on their own.

That shape is worth sitting with, because it decides what you can honestly claim. Impressions by page will tell you which of your pages Google is willing to put inside an answer, which is genuinely useful and is the closest thing to a scoreboard anyone has. It will not tell you what was asked, whether anyone clicked, or where you sat among the sources named beside you.

The same 1,000-row limit and the same date handling as the ordinary performance report apply, and the newest days are marked preliminary while they settle. One more setting matters before you read any of it: there is a control that excludes a site from Google generative AI features altogether, and a site that has been excluded will show an empty report for reasons that have nothing to do with its content.

Referral data has a structural hole in it, and it is bigger than the numbers built on top of it usually admit. When someone taps a link inside a phone app rather than a browser, the visit often arrives with no referrer header, so there is nothing for Analytics to read. It is filed as Direct, sitting beside people who typed your address from memory.

Cloudflare, which sees a very large share of the world web, said this plainly when it published its crawl-to-refer ratios in July 2025. Its own caveat was that traffic referred by a native app does not include a referrer header, that it believed the same held for other native apps, and that because its referral counts only included web-based tools, the calculations may overstate the ratios by an unclear amount. That is the company with the best view of the problem telling you its own figure is a ceiling, not a reading.

This is why we treat measurement and testing as two different jobs. Analytics tells you what happened to the people who clicked. Asking the assistants the questions your buyers ask, on a schedule, in Australia, and recording who gets named, is the only way to see the rest. It is slower and less tidy than a dashboard, and it is the only method that touches the outcome that decides the sale.

That testing loop is the core of what we do in generative engine optimisation, and the reasoning behind it is set out in how to get cited by AI in Australia and what AI assistants cite in Australia.

Measure these in sequence, because each one only means something once the one above it is in place. Nearly every AI reporting failure we see is a business measuring the fourth thing while the first is still wrong.

Two habits make the difference between a number you can act on and a number you can only recite. The first is writing down the date and the wording of every test, because answers change week to week and an undated result is worthless six weeks later. The second is comparing like with like: an assistant answering in Australia and the same assistant answering in the United States will name different businesses, and testing from the wrong place is the most common reason a result looks better than the market does.

None of this needs a large budget. It needs somebody to run the same questions on the same schedule and keep the record, which is exactly the sort of task that gets started once and abandoned. If you would rather it simply happened, that discipline is the service we sell, and our generative engine optimisation work is built around this loop rather than around a report.