Measuring AI Visibility

Track 2 · Lesson 6 of 6 · about 15 minutes

Measuring AI visibility

The final lesson is the humbling one: there is no Search Console for generated answers. Whatever you measure here, you build yourself — which is why measurement discipline separates the honest practitioners from the dashboard sellers.

Why there is no ranked list to read

Classic rank tracking works because a query has an ordered result list you can observe. A generated answer has no such list — the same prompt can cite different sources across runs, sessions, model versions and days, because retrieval is probabilistic and generation adds its own variance. Any single observation is an anecdote. That is not a tooling gap waiting to be fixed; it is the nature of the thing being measured.

Sampling is the only honest method

What works is boring: a fixed prompt set that represents the questions you actually care about; run on a schedule against the engines that matter to you; repeated — multiple runs per prompt, because one run tells you nothing; and logged with dates, engines and raw outputs. Out of that you get a defensible trend: cited in 12 of 40 prompts this month against 8 of 40 last month means something. “We appeared in ChatGPT once” does not.

Keep the prompt set stable even when it tempts you to fiddle — the moment you change prompts, you have restarted the series, not continued it.

Reading vendor scores without being sold to

Every “AI visibility score” on the market is prompt sampling with branding. Some are competent. The test is disclosure: does the vendor state the prompt set size, the engines, the repeat count, the dates? A score with a published method is a useful outsourced sample. A percentage with no method is a marketing number — and you already know from the research quiz what to do with unverifiable numbers.

Signals adjacent to the answer

Two supporting reads cost nothing. Referral traffic from assistant domains in your analytics shows citations converting into visits — scarce but real. And log hits from retrieval agents (OAI-SearchBot, PerplexityBot and friends, from lesson two) confirm you are being fetched at all, which is the precondition for everything upstream. Neither replaces sampling; both anchor it in observed behaviour.

Track complete. Test it with the GEO & AEO quiz · go deeper: AI visibility vs rank tracking · AI citation tracker.