Zero-Click Search: What to Measure Instead of Traffic

"How much traffic is ChatGPT sending us?" It is the question most marketing teams ask first about AI assistants, and it is the wrong one. Traffic counts the buyers who left an answer to visit you. In zero-click search the buyer does not leave: the answer is the destination. The useful question is another one. In how many answers about your category are you named, and in what terms?

Traffic still matters, and it is worth tracking properly. But a referral report can only see the visits that happened. By construction, it is blind to every answer that named you and sent nobody, and to every answer that named a competitor instead. This piece explains why that gap exists, where the evidence of presence actually ends up, and what can be measured in place of clicks.

What zero-click search means once an assistant writes the answer

Zero-click search is a search that ends without a visit to any website. The term is older than generative AI: a results page showing the weather, a phone number or a definition already settled the query on the spot. What has changed is the size of the answer. An assistant does not show a single fact. It composes a shortlist, compares suppliers and explains how they differ. For a buyer at the early stage of a purchase, that is often everything they wanted from the session.

The sequence behind one of those answers has several steps. The assistant may search the web, read some pages, decide which ones to rely on, write the reply, name some companies and, sometimes, attach links to its sources. The visit to your site is the last link in that chain, and the only one your analytics can see. Everything before it happens on somebody else's servers.

That is why a fall in referral traffic and a fall in visibility are not the same event. Traffic can drop while you are named more often, because more buyers are satisfied without clicking. It can also hold steady while you disappear from the shortlists of a market you have never looked at.

Where the evidence of presence ends up

Each step leaves a trace in a different place, and most of those places are not yours.

The fetch. When an assistant reads a page while answering, the request can appear in your server logs under the assistant's user agent. That tells you a page was read. It does not tell you whether it was used, or how.
The citation. Two search engines now report part of this. Google Search Console has a generative AI performance report, announced on 3 June 2026 and rolled out to all websites on 31 August 2026. It shows impressions from AI Overviews and AI Mode by page, country, device and date. It does not report clicks. Bing Webmaster Tools has an AI Performance report listing the grounding queries in which your site appears and the pages cited for each; on 13 October 2026 it gained a Topics Portfolio view. Both cover their own engines only. Neither says anything about ChatGPT, Claude or Perplexity.
The mention. Whether your company is named in the text of the answer, where it sits in the list and what is said about it exists only in the answer itself. Nothing on your side records it.
The click. The one step your analytics sees, and the one that zero-click search removes.

There is a further reason the click is a weak proxy: many answers have nothing to click. In the 17,719 answers PSentry collected between December 2025 and October 2026 through the assistants' APIs, Claude reported at least one source in 97 answers out of 100 and Perplexity in 94. Gemini did so in about 56 out of 100, and ChatGPT in 44. When an answer carries no source, a company can be named, described and recommended without a single link to follow. The sample has limits that should be stated plainly: the prompts are those of PSentry's customers, mostly B2B services and industrial firms, and an API is not the consumer app. The figures describe those answers, not every search.

What to measure in place of traffic

If the click is not the unit, the answer has to be. Visibility in AI search is the share of relevant answers in which your company is named, read across a fixed set of questions and repeated over time. A handful of measures follow from that definition, and each can be checked by hand before any tool is involved.

Mention rate. Out of a fixed set of buyer questions, how many answers name you. The questions must be the ones a buyer would type, not ones that already contain your brand.
Who is named instead. The companies that appear where you do not. In zero-click search, a competitor's mention is a sales conversation you never hear about.
How you are described. Recommended, listed in passing, or credited with a product range you dropped years ago. A mention with the wrong description can do more harm than no mention.
Which sources the answer leans on. Your own pages, a trade directory, a distributor, a forum thread. This is the part you can work on, by keeping those sources accurate.
Language and assistant, kept apart. The same question in Italian and in English can produce different shortlists, and four assistants rarely agree. One blended figure hides exactly the differences an exporter needs to see.

None of this means anything as a one-off. Answers vary between runs, so a single check is an anecdote; a fixed question set asked on a regular schedule turns it into a trend. How to count that share without flattering yourself is covered in share of voice in AI answers. The traffic side, still worth having, is in tracking ChatGPT and Perplexity traffic in GA4.

What these measures cannot tell you

A mention rate is not demand. It says how often you appear when a question is asked, not how many buyers asked it. No assistant publishes how many times a question was put to it, so any figure that claims to know is an estimate.

It is not revenue either. A buyer who reads your name in an answer may ring you a month later, through a distributor, without ever opening the website. Nothing in that path links the call back to the answer. The honest position is to treat presence in answers as an early signal, and to ask new contacts where they first heard of you.

Nor is it the model your buyer used. A measured answer comes from a specific model, at a specific moment, through a specific interface. The buyer may have been on another version, logged in, with a history that colours the reply. Measurement gives a consistent sample, not a replay of anyone's conversation.

Questions that follow the traffic report

Does a fall in referral traffic from AI assistants mean we are less visible?

Not necessarily. Traffic can drop while you are named more often, because more buyers are satisfied by the answer without clicking. It can also hold steady while you disappear from the shortlists of a market you have never checked.

Can Google Search Console show whether ChatGPT mentions our company?

No. Its generative AI report counts impressions in AI Overviews and AI Mode, without clicks, and Bing's AI Performance report covers Bing only. Neither says anything about ChatGPT, Claude or Perplexity.

Can a mention in an AI answer be linked to a sale?

Not directly. A buyer may read your name in an answer and ring you weeks later through a distributor, without ever opening the website. Asking new contacts where they first heard of you is the simplest way to see that path.

Why check the same questions more than once?

Answers vary between runs, so a single check is an anecdote. A fixed set of buyer questions asked on a regular schedule, kept apart by language and assistant, is what turns it into a trend.

PSentry asks the same set of questions to ChatGPT, Claude, Gemini and Perplexity twice a month, in each language you sell in, and records who is named and which sources are cited. It measures presence. It does not promise more of it, and it does not change what the assistants say.