Measure AI visibility engine by engine, not as a single AI score. Each platform reads different sources, cites different domains, and names different brands for the same question, so one blended number averages away the parts you can act on. Split the measurement in two. On-answer signals cover what shows up inside the reply: whether you are named, which sources back it, how you are described, and who else makes the shortlist. Off-answer signals cover what happens around the reply: traffic from assistants, AI crawlers reading your site, and what buyers say when you ask them directly.
Measure each engine separately
Start with the engines themselves. ChatGPT, Gemini, Claude, Perplexity, Grok, and Google's AI answers do not share a model, an index, or a set of source preferences. The same prompt can name you on one and skip you on another.
A single blended score collapses that spread. If you rate strongly on one engine and poorly on three, an average lands in the middle and tells you nothing useful. It hides both the win to defend and the gaps to fix.
So report per engine first. Pick the prompts that match how buyers actually ask about your category, run them on each platform, and read the results side by side. Roll up to a summary only after you can see the engine level detail underneath it.
Repeat the same prompts on a regular schedule. Answers shift as models update and as sources change, so a single snapshot ages quickly. Trend lines matter more than any one reading.
On-answer signals: presence, evidence, framing, gap
The most direct signal is presence: does the answer name you at all, and does it recommend you or merely mention you in passing. Track how often you appear across your prompt set, and measure it against rivals. This is a competitive read: of all the brands named for these questions, how often are you among them, and how prominently.
Next comes evidence. AI answers lean on a small set of sources, and each engine tends to trust different domains. Look at which pages and sites an engine cites when it answers your category questions, then check whether your own domain shows up among them. Being cited is stronger than being named without a source, because it means the engine is reading you directly.
Then framing. Being named is not enough if the description is wrong or unflattering. Read the sentiment of how you are characterized, and separately check whether the facts are correct: the products, the positioning, the claims. A confident answer that misstates what you do can cost more than silence.
Finally the competitor gap. For each question, note who else makes the shortlist. If the same two or three rivals appear beside you, or instead of you, that pattern tells you where the contest actually is. The gap between your presence and theirs, question by question, is often the clearest thing to work on.
Off-answer signals: traffic, crawlers, attribution
On-answer signals tell you what the engine says. Off-answer signals tell you what it does for your business, and they round out the picture.
The first is referral traffic from assistants. When someone follows a link out of an AI answer, that visit usually arrives with a source you can identify. Volume tends to be lower than classic search, because the answer often satisfies the question without a click. Intent, though, tends to be higher, because the person arrives already informed and closer to a decision. Judge this traffic on quality, not just count.
The second is AI crawler and bot activity. AI systems fetch pages to read and to build their sources. Your server logs show which AI crawlers hit your site and how often. Rising crawler activity is a leading indicator: engines cannot cite or recommend content they have not read, so being fetched often is an early sign you are in the running before it shows up in the answers themselves.
The third is self reported attribution. Ask new buyers where they first heard about you, in a signup field, an onboarding question, or a sales call. Some will say an AI assistant. This is softer than log data and depends on what people remember, but it captures influence that never leaves a clickable trail, which is common when the answer names you without a link.
Why one blended score hides more than it shows
A single AI visibility number is tempting because it is easy to report. It is also easy to misread. Averaging across engines, across signal types, and across questions folds together things that call for different responses.
Consider two brands with the same overall score. One is named often but described incorrectly. The other is described well but rarely named. The number is identical; the work is not. One needs to correct the record, the other needs to earn more mentions.
The same problem appears within a single brand over time. A score can hold steady while presence falls and framing improves, or the reverse. The headline stays flat and the real movement disappears.
Keep the components visible. Use a summary to orient, then always let people open the layer beneath it. The value is in the breakdown, not the roll up.
How the signals fit together
Read the signals as a chain, not a scatter of separate metrics. Crawler activity comes first: engines fetch your pages. Evidence follows: those pages start showing up as cited sources. Presence and framing follow from that: once an engine is reading you, it can name you and describe you. Traffic and attribution come last, as the downstream effect of appearing in answers people trust.
That ordering makes diagnosis easier. If crawlers are not reaching key pages, work on access and structure. If you are cited but not named, the framing or relevance of the mention may be the issue. If you are named well but see little traffic, the answers may be satisfying the question without sending a click, which is expected and not a failure.
No single number sits at the end of this chain, and it should not. The goal of measurement is to show which link is weak for which engine, so the next action is obvious. Report the parts, watch them over time, and let the pattern point to the work.