Method
How to Measure a Site's AI Visibility
AI visibility isn't one number. It's a set of separate observations, each tied to a specific engine and to the part of the page that engine actually reads.
If you're asking how to measure a site's AI visibility, start with a concrete point: you don't measure the page as you see it, but the portions of the page that each engine reads. Say you've published a guide and want to know whether ChatGPT, Perplexity or Gemini cite it. First you have to establish which parts of the document those systems actually read. If the useful content sits inside a script or behind a block they can't access, the measurement doesn't cover that page. That's where RankAI starts: from the public pages engines read, separating what is observable from what stays outside.
What observing AI visibility means
AI visibility is not a single number. It's a set of separate observations, each tied to an engine and to a type of reading. RankAI's public pages name ChatGPT, Perplexity, Gemini, Grok and Google AI. That list is the declared scope: it doesn't include other systems, and it shouldn't be extended by analogy.
Two distinctions matter, and both are verifiable:
- Claude and Grok are measured only after you connect them. Before that step, no measurement is available.
- Gemini and Google AI Overviews are not the same observation. Treat them as two distinct readings, even when they concern the same site.
So a single aggregate figure can't answer the reader's question. You need to know which engine read which part of the page.
RankAI's proposed method
What follows is an operational proposal from RankAI, not a universal standard. Its purpose is to make the measurement verifiable step by step.
- Define the observable page. Verifiable criterion: the portion of content you want to measure must be readable without depending on elements the engine doesn't process. If the useful text isn't on the public page, the measurement doesn't cover it.
- Assign every observation to an engine. Verifiable criterion: every result must name the engine that produced it. A figure with no engine attached is not an AI visibility measurement.
- Treat unconnected engines as unmeasured. Verifiable criterion: for Claude and Grok, a measurement exists only after connection. Before that, the correct state is no measurement, not a low value.
- Keep Gemini separate from Google AI Overviews. Verifiable criterion: record the two observations in separate fields, even when they concern the same page and the same period.
- Record absence as N/A. Verifiable criterion: when a visibility figure is unavailable, the product shows it as N/A, never as zero. Zero is an observed value; N/A is an absence of observation.
The trickiest part of AI visibility is the temptation to fill the gaps. If an engine isn't connected, or a page isn't readable in the part that matters, the instinct is to write zero: it looks like a clean, comparable figure, useful for an average. But zero means the engine read and found nothing; N/A means we didn't observe. Those are two different claims, and they lead to different decisions: the first suggests rewriting the content, the second suggests enabling a reading that's still missing.
That's why, in how we set up the measurement, the basic unit isn't the site but the engine-page pair. A page can be observable for one engine and not for another, and that isn't an error to fix: it's information. Keeping the fields separate, declaring the scope of the named engines and preserving N/A values makes the picture comparable over time without inventing continuity where there is none. The check we propose is simple: for every figure shown, it must be possible to trace it back to the engine, the page and the type of reading that produced it. If any of the three is missing, it isn't an AI visibility measurement but an approximation.
A verifiable limit
The limit lies in the scope itself: RankAI's public pages name ChatGPT, Perplexity, Gemini, Grok and Google AI, and Claude and Grok are measured only after you connect them. Anything outside that list isn't covered by the measurement and shouldn't be inferred by similarity. Also, an API response is not a person's screen: what we observe doesn't match a user's visual experience in front of the engine.
For the full picture of how the measurement works and what stays outside it, see the in-depth page.