How do I measure AI search visibility when there is no ranking and barely any referral data?
You measure it the way a pollster measures opinion: by sampling. Define a fixed set of prompts a real buyer would ask, run them on a schedule across the assistants you care about, and record whether you were cited, mentioned without a link, or absent. The output is a citation rate per prompt, tracked over time. It is an estimate rather than a count, and it is the only honest number available.
Three things that are actually measurable
None of these is a ranking. All three are rates, which means they need repeated sampling to be stable and should always be reported with the number of runs behind them.
- Citation rate. The share of runs where your domain appears as a linked source. The cleanest signal, and the hardest to earn.
- Mention rate. The share of runs where your company is named without a link. Often higher than citation rate, and it still produces branded searches and phone calls.
- Source-set share. Of all sources cited across a prompt set, how many are yours or are pages describing you. This tells you whether the conversation about your category includes you at all.
Build the prompt set from the buyer's sequence, not your keyword list
Keyword lists describe how people search when they already know the vocabulary. Assistants get asked in the language of the problem: the noise it makes, what it smells like, whether it can wait until Monday.
The best source of that language is your own phone. Real callers describe problems in exactly the words a person types into an assistant, and pulling those phrasings out of recorded calls is a natural byproduct of call analysis. A prompt set built this way is worth far more than one written in a conference room.
Freeze the set, then defend the freeze
Same prompts, same location, same signed-out conditions, same cadence. When you think of a better prompt, add it as a new one with its own start date. Editing a prompt that already has history destroys the trend without making it obvious that it did.
Record the date and the assistant version alongside every run. Vendors change models without announcement, and a sudden step in your numbers is more often theirs than yours.
Never report it alone
Citation rate measures presence in answers. It does not measure demand. Pair it with branded search volume, direct sessions, and the answer to the question you ask every caller about how they found you.
One of those numbers moving without the others is usually noise. All of them moving together is a real trend. Reporting them side by side in one place is what a marketing intelligence view is for, and it keeps an unauditable number from being mistaken for a result — the same standard we apply to everything in revenue reporting.
Topics: measurement · reporting · visibility · methodology
Have a version of this question about your own business?
The useful answer usually depends on which systems you run and how they're connected. That's a conversation, not a blog post.