How we measure AI visibility

Five signals. Ranked by what they can prove.

Most AI-search reporting is one dashboard number with no idea where it came from. We rank every signal by what it can prove and lead with the layer that holds up.

Hawaiian hibiscus botanical specimen, 18th century natural history plate style

The evidence ledger

Every signal, ranked by what it can prove.

First-party and verifiable at the top; everything else labeled for what it is.

Load-bearing

Server & CDN access logs

The proof, because it can't be faked

Source

First-party

What it proves

A platform read the page

How we use it

We lead every report on this.

Supporting

GA4 AI Assistant channel

Real, with a real blind spot

Source

First-party

What it proves

AI-referred sessions that arrive tagged

How we use it

Supporting, blind spot stated out loud.

Supporting

Self-reported attribution

Signal, with a known bias

Source

Third-party

What it proves

How customers say they found you

How we use it

Corroborates other layers; never the sole basis.

Directional

Share-of-voice tools

Direction, not proof

Source

Third-party

What it proves

Rough trend across sampled prompts

How we use it

Color and context; never a headline number.

Omitted

Incrementality

We don't claim it, nobody can

Source

None

What it proves

Nothing yet, no validated method

How we use it

Left out on purpose.

The through-line

One proof spine, several supporting angles.

Every report leads with what the server logs show. GA4 comes next, blind spot stated plainly. Share-of-voice and self-report round it out as directional and supporting. Nothing is presented with more confidence than its method can support.

FAQ

Questions about measuring AI search visibility

We start with server and CDN access logs, which record when OAI-SearchBot, PerplexityBot, or Googlebot crawls a page. That’s first-party and verifiable. From there we layer in GA4’s AI Assistant channel, share-of-voice tracking, and self-reported attribution, each weighted by how much it can prove.
Because a single score hides which parts of it are provable and which parts are estimates. Share-of-voice tools sample a fraction of possible prompts, and a citation has well under a 1 in 100 chance of reappearing identically across runs. Collapsing that into one number makes a volatile, partial signal look like a settled fact. We show the layers instead, so you can see what’s solid and what’s directional.
No, and any agency who tells you they can is overclaiming. Isolating incrementality, whether a customer would have converted anyway, is the hardest open problem in marketing measurement, and there’s no validated method for doing it in AI search yet. We report what the layers can show and leave out the number nobody can currently prove.
Being crawled means a platform’s bot read the page, which shows up in server logs. Being cited means the platform used that page in an answer it gave someone. A page can be crawled constantly and cited rarely, since most retrieved content is never surfaced in a final response. Crawl logs prove the first. Share-of-voice tools give a directional read on the second.
GA4’s AI Assistant channel depends on the referring platform sending attributable signals with the visit. It’s documented to be blind to Google AI Overviews entirely, and it likely undercounts Perplexity as well, since not every AI-originated visit arrives with a trackable referral. It’s real data, we just don’t treat it as the full picture.

See these layers applied to your own presence.

The Context Map is our fixed-scope diagnostic, same proof spine, run against your site.

Get your Context MapOr apply to work with us