When the AI recommendation calls instead of clicking: what is still measurable

The call that replaces the click

A tax firm asks ChatGPT which AI phone assistant suits it. The answer names three providers and draws on six sources. The partner reads, remembers a name, looks up the phone number and calls.

Up to that point everything is traceable: the question asked, the providers named, the sources cited, the order of the names. From the call onwards, one trace remains, the voice at the other end. Referrer, campaign parameters, form and session all stay behind on the web. The most valuable contact in the whole chain is also the quietest.

This guide covers both halves: what can be measured before the contact, and what is left of the origin when the contact is a conversation.

What is measurable before the contact

On 26 August 2026 we put twelve buying questions in German to ChatGPT, Claude and Perplexity through their official interfaces. Subject: AI phone assistants, a young market with dozens of providers competing for it. The result: 36 answers, 366 cited sources and 168 distinct domains.

The first finding is about the sources, ahead of the brands. Of the fifteen most cited domains, thirteen belong to companies that sell on this market themselves. Those thirteen alone account for 122 citations. The most cited page in the whole survey is a ranking of seven providers, published by a company that sells one, and it appears in 10 of the 36 answers.

For a company that wants to appear on this market, the consequence is immediate: an AI answer is built to a large extent from comparisons written by competitors. Its own website contributes to it, and those comparisons carry the larger share. How to open their doors is covered in our guide to mentions on third-party sites.

Three engines, three webs

The three engines searched the web for 35 of the 36 answers. They each read their own pages.

Of the 168 domains in the survey, 125 are cited by exactly one of the three engines, that is 74.4%. Nine domains are cited by all three. Question by question, two engines share 10.9% of their sources on average.

This continues what we measured in July and August on a far larger corpus: six AI engines, asked the same question on the same day, share 6.4% of their sources. In practice, visibility measured on one engine holds for that engine. A provider can lead on Perplexity and hold a very different place on ChatGPT, because each engine answers from its own web. Why an answer comes out the way it does is the subject of the guide Why ChatGPT recommends your competitors.

Tracking archives these answers with their sources, question by question and engine by engine. The Atlas orders the sources by domain and by engine and shows which ones keep coming back. Competitors starts from a basket of competitors and returns the sources of the market ranked by frequency.

What the call carries over

A form walks a prospect through fields, and every field can be read back afterwards. A call comes in on a number that is the same for every question, and what triggered the contact is known at that moment to the caller.

The prospect stays. What travels with the call is the origin, and with it the answer to whether the visibility work produced anything. A company can lead the answers of several engines and still leave open which share of its business comes from there.

The one field that survives the call

There is an exception, and it lies in the conversation itself. Callers very often say of their own accord where they come from:

“ChatGPT recommended you for our firm. We are looking for something to take our calls, but we have three offices, and I wanted to know first whether it fits our calendar.”

That single sentence holds the origin of the attention, the sector, the need, a complexity marker, a technical question and the buying stage. A form would have captured two fields of it. The sentence was spoken, so it is there. The only question is whether anyone records it.

That is exactly the job of a layer that takes calls and turns them into structured data. One German example is the AI phone assistant from Teloro, built by Teloro GmbH in Mannheim. The system takes incoming calls in natural language, captures the need, asks targeted follow-up questions and prepares a defined next step: an appointment, a call back, or a handover to a person who already has the context. Putting the question “how did you hear about us” into that intake gives the phone channel exactly what a form gives elsewhere. The provider describes the flow for customer service in detail.

Epovest measures which questions a market asks, who is named in the answers and which sources the engines read to produce them, and derives the actions from it: where to appear, what to change, and the evidence of what moved. What happens next on the phone belongs to another category of tools, and Teloro is one of them.

A self-declared origin is an indication, and reads as one: someone who read an AI answer, then opened two comparison pages and called the next day will usually name the last stop. For the phone channel, that sentence remains the input an analysis can rest on.

What is worth doing about it

1. Read the market through its sources before reading it through its brands. On these twelve questions, knowing the fifteen most cited domains says more about the answers to come than the list of known providers.
2. Appear in the comparisons that keep coming back. Three domains alone carry 52 of the 366 citations. An accurate, current entry on those pages weighs more than one more page on your own website.
3. Measure engine by engine. With 74.4% of domains cited by exactly one engine, the per-engine figure is the telling one.
4. Put the origin question into the first contact, on the phone too. It costs one sentence and it is the origin signal a call can carry by itself.
5. Repeat the same measurement. A dated point says where you stood on one day. It is the second wave that says whether anything moved.

Method

Protocol. On 26 August 2026, twelve buying questions were put in German to ChatGPT, Claude and Perplexity through their official interfaces, once per engine: 36 answers. On each engine the provider's default model was queried. 35 of the 36 answers triggered a web search while answering.

How domains were classified. A domain counts as selling on this market when a product page on the same domain offers an AI phone assistant or its setup. The classification was done by hand for the first fifteen domains, and for those only. The remaining 153 were left open.

Counting unit. A citation is one source cited in one answer. A domain cited twice in the same answer counts as two citations. The unit of the domain table is the registrable domain, and a subdomain is grouped into it.

Limits. One repetition per question and engine, on ChatGPT, Claude and Perplexity. The survey is a dated point: it says what these three engines answered on 26 August 2026.