How Do You Actually Measure AI Visibility for Your Brand?

How do you actually measure AI visibility for your brand?

Start with the one first-party report that exists, which is Google's generative AI performance report in Search Console, then add manual prompt testing for the engines that publish nothing. Anything beyond those two is estimation, and it should be labelled as estimation when you put it in front of a client.

This is a much better position than we were in a year ago, because until recently there was no official data at all and the entire category ran on third party sampling. There is now a real number from the company that operates the surface, at least for Google's own features.

What follows is what that report contains, what it does not, and the honest method I use for the rest.

What did Google actually launch, and when?

On June 3, 2026, Google announced new Search Generative AI performance reports in Search Console, including dedicated reports for Search and Discover, to help site owners understand their visibility within generative AI features on Search. The announcement came from Hillel Maoz and Moshe Samet of the Search Console team.

Availability matters here and it changed recently. Google's launch post says the reports were initially rolled out to a subset of websites for testing, and carries a note stating that as of August 31, 2026, these insights have been rolled out to all websites worldwide. So if you looked for this report earlier in the year and did not have it, look again.

Google also notes the data is included in the overall performance report where it continues to be tracked, and that the new launch is a separate view dedicated to visibility from generative AI features. You are getting a filtered lens on existing data, not a parallel universe.

What does the report actually show you?

Five things, named in Google's announcement. Impressions, meaning how often URLs from your site appeared in generative AI features in Search and Discover. Pages, so you can check which URLs appeared. Countries. Devices, which Google notes is available for Search results. And dates, with hourly, daily, weekly and monthly granularity.

Read that list carefully, because what is absent is as informative as what is present. Google's launch post lists impressions rather than clicks among the metrics it describes, and closes by saying the team is continuing to work with website owners on what insights would be most helpful, such as adding additional metrics over time.

So the honest framing for a client is that you can now see whether your pages are appearing inside AI features, in which countries, on which devices, and how that changes over time. Whether that appearance produced a visit is a separate question this view was not built to answer.

Why is impression data still worth having?

Because it separates two failures that look identical from the outside. If your traffic is flat and your impressions inside AI features are high, you have a click problem. If impressions are near zero, you have a presence problem. Those need completely different work, and before this report you could not reliably tell them apart.

The pages dimension is the most actionable part for me. Knowing which specific URLs are appearing tells you what the system considers your strongest material, which is frequently not the page you would have guessed. I have found pages surfacing for questions their authors never had in mind.

Country data matters more than it sounds for anyone selling across borders. Appearing in AI features in one market and not another is a genuinely different problem from appearing nowhere, and it usually points at content fit rather than technical eligibility.

What is the eligibility check people skip?

Whether your site can appear at all. Google's optimization guidance states that to be eligible for its generative AI features a page must be indexed and eligible to be shown in Google Search with a snippet, and that a site must be included in Search generative AI features in Search Console.

That is two separate conditions and both are worth verifying before you conclude anything about your content. A page excluded from snippets by a directive, or a property that is not included in those features, will produce an empty report that says nothing about how good your writing is.

Google is also careful to say that meeting every requirement does not guarantee crawling, indexing or serving. I quote that line to clients because it prevents the conversation where a technically perfect implementation is expected to produce a guaranteed outcome.

How do you measure visibility in engines that publish nothing?

Manually, systematically, and with honest labelling. Build a list of the twenty to forty questions your buyers actually ask, run them across the engines you care about on a fixed schedule, and record whether you were mentioned, whether you were cited with a link, and which competitor was cited instead.

The discipline that makes this useful is consistency rather than volume. Same prompts, same cadence, recorded the same way, so the comparison over time is meaningful even though any single result is noisy. Answers vary between runs, and a single check tells you almost nothing.

Be explicit that this is sampling. It is a panel of questions you chose, not a measurement of the whole surface, and the honest word for the output is indicative. I would rather hand a client a small number I can defend than a confident percentage nobody can reproduce.

Do the vendor tools help?

They can, as long as you read what they are actually measuring. Webflow, for example, says its AEO product pulls data from ChatGPT, Claude, Gemini and Perplexity so you can see where you are mentioned, where you are cited, and who is getting cited instead of you. That is the right set of questions to be asking.

Webflow also published a figure about the underlying problem at its 2026 conference, saying it analysed the websites of 2,000 companies and found the median company appeared in 16 percent of the AI answers they would want to be part of, and were cited in those answers only 6 percent of the time. Those are Webflow's own numbers about a market it sells into, so weigh them accordingly.

The thing to check with any tool in this category is how it gets its data, because the underlying method is almost always prompting engines at scale rather than receiving data from them. That is a legitimate approach. It is also sampling, and it should not be presented to a board as if it were server logs.

Which metrics are actively misleading?

Any single percentage describing your share of AI answers. The denominator is unknowable, the results vary between runs and between users, and the number will move for reasons that have nothing to do with your site. It looks like a KPI and behaves like a mood.

Mention counts without citation counts are the second trap. Being named in an answer and being linked from it are different outcomes with different commercial value, and collapsing them into one figure hides the distinction that matters most to anyone hoping for traffic.

The third is comparing engines directly. Different systems have different retrieval behaviour and different citation habits, so a lower number on one engine is not evidence of a weakness. Track each engine against its own history instead.

What do you actually change once you have the data?

Work on the pages the report says are already appearing, before you write anything new. A page that surfaces in AI features but fails to earn the click usually has a fixable problem in its opening lines, and improving it is cheaper than manufacturing a new candidate.

If a page appears but the quoted passage is not the one you would have chosen, that is a structural issue in how each section opens rather than a content quality issue. I went through the mechanics in why AI search quotes one paragraph and ignores the rest and, on the writing side, in writing answer blocks that get cited.

If you are being contradicted by other sources on the same question, that is a third and different problem, and it is worth understanding how engines handle disagreement before you rewrite anything. I covered that in how AI answer engines handle conflicting information.

What should you do next?

Open Search Console and find the generative AI performance report. Google says these insights reached all websites worldwide on August 31, 2026, so it should be there, and the pages dimension alone is usually worth the ten minutes it takes to read.

Then write down twenty questions your buyers actually ask and run them once this week, recording mentions and citations separately. That gives you a baseline you can repeat monthly, which is the only thing that turns a noisy measurement into a useful one. If you want help building a prompt panel that reflects how your buyers really search, reach out and tell me what you sell.