Your GEO tracker is not measuring what you think it is

Published 17 July 2026 · Based on June 2026 field data · Last reviewed 17 July 2026

For two years the GEO industry has sold a number it cannot actually measure. Every tool in the category, ours included, reports something that looks like a citation count. Buyers read it as volume. It is not volume. It never was.

In late May 2026, Google started reporting AI surface data inside Search Console. For the first time, a site owner can see first party impression counts for AI Overviews, AI Mode and generative surfaces in Discover. Not an estimate. Not a sampled panel. Google's own log of when it put your page inside an AI answer.

We had that data land on an account we had already been tracking with a synthetic prompt panel for eight months. Same site, same month, two instruments. What came back is worth publishing, because it changes how the number on your dashboard should be read.

Prompt panel vs Google Search Console: what each one actually measures

This is the part the category glosses over, so it is worth being slow about.

An experiment and a census answer different questions. The experiment answers "when this specific question is asked, who gets named". The census answers "how often did we actually show up". Both are useful. Neither substitutes for the other, and dividing one by the other produces a meaningless ratio.

The account: one B2B software vendor, two instruments

A B2B software vendor, anonymised. Long sales cycle, technical buyers, a category where four or five vendors get compared constantly. We have run an AI visibility program on it since late 2025 and tracked it in PromptWatch throughout.

Read those three numbers again. They do not combine. There is no arithmetic that turns 32.3% into 15,088, because one is a share of a controlled prompt set and the other is a count of real impressions. That is the entire point.

What Google's first party AI data shows that a prompt panel cannot

Google's data starts at 18 May 2026. The series so far:

Before this data existed, the best available proxy for the same surface was a third party estimate: roughly 45 AI Overview appearances for this domain. That estimate was not wrong either, and this is the trap. An appearance is a position in an AI Overview. An impression is one instance of that overview being shown to someone. Forty-five appearances can produce tens of thousands of impressions, because a single appearance on a query with volume gets served over and over.

What the census does reveal is concentration. Of 198 pages surfacing in Google AI, the top two account for over half of all impressions:

That table is a strategy document. Six of the seven pages Google puts inside AI answers are explainers and comparisons. The first commercial product page ranks seventh. If you had been optimising product pages for AI visibility on this account, the census just told you that you were working on the wrong pages, and no prompt panel would have told you that, because the panel only sees the prompts you thought to write.

What a prompt panel shows that Google Search Console cannot

Now flip it. Here is the same account, same month, in the tracked panel:

Google's census cannot produce this chart. It has no idea who else was named in the answer, because Search Console only reports on your property. It also cannot see ChatGPT, Claude or Perplexity at all, which on this account carry most of the assistant referral traffic.

The panel can. And when it moved 6.1 points in a month against a stable 90 day baseline, that movement was real and attributable, because the prompt set did not change. That is what a controlled experiment buys you.

How to measure AI visibility: three instruments, three jobs

Three instruments, three jobs, no substitutions:

The uncomfortable consequence: there is no single number for AI visibility, and any tool that gives you one is giving you an artefact of its own configuration. We sell GEO. We would rather sell one clean number. It does not exist.

What to do with this: four actions

• Turn on the GSC AI surface report today. It is free, first party, and it is the only census you have. If your reporting still runs on a third party AI Overview estimate, replace it.
• Read your panel as shares and trends only. If your agency reports a citation count as a headline KPI, ask them how many prompts are tracked and how often they run. The count moves when that config moves.
• Let the census pick your content targets. Six of seven top surfacing pages on this account were explainers. That was not the plan going in. It is the plan now.
• Never divide one instrument by the other. Different units, meaningless ratio.

We told this client, in writing, in their June report, that our own tracked panel captures only a slice of real citation activity and that its absolute counts sit far below citation reality. It is an awkward thing to put in a deliverable you are being paid for. It is also the only defensible position once the census exists, and we would rather say it before someone else says it for us. See our monthly Run program or book a GEO Audit call.