Pull up your GA4 acquisition report. Somewhere near the bottom, under "referral," there is a row for chatgpt.com. It is probably under two percent of sessions. Now look at the conversion rate on that row. In most published studies, visitors who arrive from an AI assistant convert at several times the rate of organic search, because they show up after the assistant has already answered their first five questions and named you as the fit.
That row is AI referral traffic. It is small, it is growing every month, and it is almost certainly undercounted. If you run marketing for a company that sells anything researched online, this is the number your board will ask about within a year. Better to own it now.
What counts as AI referral traffic
AI referral traffic is any session that starts when a person clicks a link inside an AI assistant or AI search surface. That means ChatGPT (including its search mode), Perplexity, Gemini, Microsoft Copilot, Claude, Grok, and the citation links inside Google AI Overviews and AI Mode.
It is not the same as being mentioned. A mention is visibility. A referral is a click. Visibility is what you optimise for; referrals are the first hard proof that the optimisation turned into pipeline. You need both numbers, and you need them in the same report.
Why your analytics undercounts it
Three leaks hide most of this traffic.
First, Google. Clicks from AI Overviews and AI Mode carry a google.com referrer, so they land in your organic bucket and look like ordinary search. Search Console now folds AI Mode impressions and clicks into the Web performance report, but it does not break them out. You cannot see them in GA4 at all.
Second, apps. The ChatGPT mobile and desktop apps, Copilot inside Windows and Edge, and in-app browsers often strip the referrer. Those sessions arrive as "direct." Your direct traffic has been quietly rising for eighteen months. Part of that is AI.
Third, naming. Some engines pass a referrer, some pass UTM parameters, and some do both inconsistently. ChatGPT appends utm_source=chatgpt.com to many outbound links; Perplexity passes a clean perplexity.ai referrer; Copilot can show up as copilot.microsoft.com, bing.com, or nothing. If you are not grouping these by hand, GA4 scatters them across four channels.
Build a dedicated AI channel in GA4
This takes about twenty minutes and pays for itself the first time someone asks "is AI sending us anything."
Open Admin, Data settings, Channel groups, and create a custom channel group. Add a channel named "AI Assistants" with a rule that matches source against a regex such as
chatgpt|openai|perplexity|gemini|bard|copilot|claude|anthropic|grok|x\.ai|mistral|you\.com|phind. Place it above Referral and Organic Search so it wins the match.Add a second rule for session campaign or source containing chatgpt.com to catch the UTM version.
Build an exploration with sessions, engaged sessions, key events, and conversion rate, segmented by that channel. Pin it to your weekly dashboard.
Annotate the date you created the channel so nobody mistakes the jump for growth.
Once it exists, compare the AI channel against organic on two metrics only: conversion rate and pages per session. Volume will look embarrassing for a while. Quality will not.
Attribute the clicks you cannot see
For the leaks above, there is no clean fix, so use proxies that hold up in a board deck.
Track branded search volume alongside AI mention rate. When an assistant names you in a buying question, a share of those people search your brand on Google ten minutes later. A rising branded-search curve that tracks your AI mention curve is the attribution story for the invisible clicks.
Add "How did you hear about us?" as a free-text field on demo and signup forms. Within a quarter you will have a list of answers that say "ChatGPT recommended you." Count them. That is pipeline you can name.
Watch direct traffic to deep pages. Nobody types a 90-character URL to a comparison page. Direct sessions landing on long URLs are, in most cases, copied or clicked from an assistant.
Which prompts send the clicks
Referrals come from citations, and citations come from specific prompts. The question is not "does ChatGPT send us traffic" but "which buyer questions make ChatGPT cite our page instead of a competitor's."
Take the ten pages that already receive AI referrals in your new channel. For each, ask the engines the questions a buyer would type that would plausibly surface that page. Note where you are cited, where a competitor is cited, and where nobody is. Pages that get mentioned but never cited are usually missing one thing: a clear, extractable answer near the top with a stat or a definition an engine can quote.
Re-run that set monthly. Models change their retrieval sources without warning. A page that earned citations in spring can go dark by autumn while your SEO rankings sit exactly where they were.
Report it like revenue, not like traffic
Three numbers belong in your monthly report: AI referral sessions, AI referral conversion rate against the site average, and the count of high-intent prompts where your brand is cited. The first one will be small. The second one is your argument. The third one is your leading indicator, because citations this month become referrals next month.
Tracking that third number by hand works for a few weeks. LLM Search Console runs your prompt set across ChatGPT, Perplexity, Gemini and the rest on a schedule, records which pages get cited, and lines that up against the referral data so you can see which prompts are actually producing clicks.
Do this before Friday
Create the AI Assistants channel group in GA4 today, backdate nothing, and screenshot the conversion-rate comparison after fourteen days. Then subscribe to the newsletter and we will send you one attribution teardown a week, built on real data from real sites.