Most AI visibility reports end one step too early. They tell you where your brand appeared. They do not tell you who called.
For a national software company, that gap is an attribution annoyance. For a dental clinic, a law firm or a roofing contractor, it is the whole question, because the lead that matters arrives as a phone call, and a phone call carries no referrer.
I spend a lot of time on this problem at GEOflux. The pattern is consistent: businesses are being discovered through AI assistants far more than their analytics show, and much of the measurement gap is self-inflicted.
The call that seems to come from nowhere
Consider how a local lead moves through an AI answer. Someone asks ChatGPT, Gemini or Perplexity for a recommendation, gets three names, and taps a phone number or searches one of the names directly. Your analytics record nothing, a direct visit, or a branded search.
ChatGPT alone now serves more than a billion users, according to OpenAI. Even a small share of local discovery moving through that interface is a meaningful number of calls, and almost none of them will be labelled correctly.
There is also a technical trap most teams have not noticed. Dynamic number insertion, the standard call tracking method, usually swaps phone numbers in the browser with JavaScript.
AI crawlers typically read the raw page without running that script. So the number they learn and repeat is your untracked default number.
The result is a channel that routes calls to the one line your tracking treats as unattributed. If your direct and unknown call volume has crept up this year while organic clicks flattened, that is not a coincidence worth ignoring.
Measure it in three layers
You will not get perfect attribution here. You can get directionally honest attribution, which is more than most of your competitors have.
The first layer is clicks you can see. Some assistants pass referral data when a user taps a cited link. Segment those sources in GA4 rather than letting them dissolve into referral noise, and watch the trend rather than the absolute number.
The second layer is calls you can hear. Record and transcribe calls, with consent, and ask one question at intake: how did you find us? Then search the transcripts for phrases like "ChatGPT said" or "the AI recommended," which customers volunteer far more often than marketers expect.
The third layer is the infrastructure fix. Serve tracking numbers server-side, or dedicate a number to the version of your business data that AI systems read, so AI-sourced calls stop hiding inside your default line.
I made a related argument about how your Search Console is leaking conversations. The data is usually already there. Nobody set it up to be read.
Use customer language as the input
The transcripts do a second job, and it is the more valuable one. They show you exactly how customers describe their problem before they learn your vocabulary.
A clinic writes "periodontal treatment." A patient says "my gums bleed when I brush." AI assistants answer the patient's question, and they tend to pull from sources that use the patient's words.
Build your prompt testing library from those real phrases, not from the keywords your team brainstormed in a meeting. Then run the same prompts repeatedly across assistants, because answers drift, and a single check tells you very little about how often you actually appear.
Structure matters as much as wording. Write your core claims as specific, checkable statements: which services you offer, in which areas, with which qualifications, at what price range. A model can reuse a precise claim, while a vague one gets replaced by a competitor's precise one.
Finally, make those facts consistent everywhere a model might read them: your website, Google Business Profile, directories and review platforms. Contradictory hours, service lists or locations give a model a reason to recommend someone whose information is cleaner.
If you want to measure where your brand actually stands in AI-generated answers, GEOflux.ai (geoflux.ai) is built specifically for this. It maps not just whether your brand is mentioned, but why, and what to do about it.
Visibility is not the outcome
The AI visibility conversation has spent a year on mentions, share of voice and citation counts. Those metrics are useful as leading indicators, but they are not commercial outcomes.
A few months ago I wrote that AI mentions now move real traffic. For local businesses, the next step is admitting that the traffic often is not traffic at all. It is a phone ringing.
There is a budget consequence too. A channel you cannot attribute is a channel that loses every internal argument about spend, no matter how much revenue it actually drives.
Connect the mention to the call, and AI visibility becomes a budget line you can defend. Leave the gap open, and it stays a slide in a quarterly deck.