TechmadeazyA founder showed me her analytics last month and asked why we were talking about AI at all. Her ChatGPT referrals for the whole of August came to 41 sessions. Direct traffic was up 60% year on year and nobody could explain it.
Those two facts were the same fact.
Google Analytics 4 was built before anyone was getting shopping advice from a chatbot, and it has never really caught up. There is no default channel for AI traffic. Depending on how a visitor arrives, an AI referral can land in Referral, in Direct, or in Unassigned, and a large share of it lands in Direct. Estimates vary, but somewhere between a third and two thirds of AI referral sessions arrive with no referrer header at all, which GA4 reads as somebody typing your URL into a browser.
So before you decide AI is not sending you customers, it is worth checking whether you can see them.
Why the referrer disappears
When you click a link in a normal web page, your browser passes along where you came from. That is the referrer header, and it is what analytics runs on.
Click a link inside a mobile app, or inside a desktop assistant, or inside an AI answer embedded in another product, and that header often gets stripped. Nothing malicious is happening. It is a mix of app behaviour, privacy defaults and integrations that were never built with attribution in mind.
The result is that the traffic arrives, converts, and gets filed under Direct alongside people who typed your name in.
The fix, in the order I would do it
1. Build a custom channel group
In GA4, Admin, Data display, Channel groups. Copy the default group, add a new channel at the top called AI, and give it a rule matching your session source against a regex. Start with the obvious ones: chatgpt, openai, perplexity, gemini, google.*ai, claude, anthropic, copilot. Add to it as you find others in your referral report.
Put the AI channel above Organic Search in the ordering, or Google's own AI surfaces will get swallowed by it.
2. Check the referral report by hand first
Before you write the regex, spend ten minutes in Reports, Acquisition, Traffic acquisition, and look at session source and medium with no filters applied. You will find hostnames you did not expect. Every market throws up a slightly different list.
3. Use the one assistant that plays nicely
ChatGPT appends utm_source=chatgpt.com to its outbound links. That is why ChatGPT traffic shows up more cleanly in most accounts than anything else, and why people assume it is the only assistant sending them anyone. Gemini and Perplexity mostly do not tag their links, so their share of your traffic is almost certainly understated relative to ChatGPT's.
4. Separate AI from everything else in your conversion reporting
Once the channel exists, build one exploration: AI as the dimension, sessions, engaged sessions, conversions and revenue as the metrics, compared against Organic Search and Paid. That single table is the one you will actually look at each month.
What you still will not be able to see
Google AI Overviews and AI Mode clicks pass through Google's normal organic channel. There is no clean way to split them out in GA4, and Search Console does not give you a filter for it either.
This matters more than it sounds. For a lot of ecommerce categories, AI Overviews is the surface with the widest reach, and it is the one your analytics is least able to describe. You can infer movement from page-level patterns and from your own crawler logs, but inference is not measurement.
It is also the reason answer tracking exists as a separate discipline from analytics. Analytics tells you what happened after someone clicked. It cannot tell you how often you were named in an answer that produced no click at all, and for a considered purchase the answer with no click is often the one that decided it.
Read the numbers honestly
You will find plenty of published multiples for how much better AI traffic converts. Adobe reported AI-referred visitors to US retail sites converting 54% higher than other sources. Other studies put it at four times organic. Some put individual clients far higher than that.
Two cautions.
The first is that the multiple varies enormously by category and by price point. A £25 T-shirt and a £900 mattress do not behave the same way, and anyone quoting you a single number is quoting one company's quarter.
The second is more awkward. When AI referral traffic to your site triples, part of that is your work and part of it is the assistants themselves growing. One study that ran for several months on a live site tracked a set of untouched comparison pages alongside the pages being optimised. Traffic from ChatGPT to the optimised pages grew 5.7 times. The untouched pages grew 3.5 times on their own. Strip out the platform growth and the real effect of the work was closer to 1.8 times.
That is still a good result. It is just not the result the headline implies, and you should know which one you are looking at before you present it to a board.
What does not work
Judging AI performance from GA4 alone, in either direction. It undercounts your traffic, it cannot see the answers you appeared in, and it flatters you on growth that the platforms would have delivered anyway.
The honest setup is analytics for what converts, and a tracked prompt set for what gets said about you. Neither substitutes for the other.
Where to start
- Open your traffic acquisition report and sort by source. Note anything that looks like an assistant.
- Build the custom channel group. It takes about fifteen minutes.
- Set a baseline this week, before you change anything on your site, so that whatever you do next has something to be measured against.
- Accept that a portion of your AI exposure will never appear in analytics, and track answers separately.
