How to track ChatGPT and AI traffic in GA4

You cannot improve what you cannot see, and by default GA4 barely shows you your AI traffic. Worse, a big chunk of it is quietly mislabeled as 'Direct.' Here is how to build a channel that surfaces AI referrals from ChatGPT, Perplexity, Claude, Gemini and Copilot - and why even then the number is a floor.

Marketers keep asking "how much traffic are we getting from AI?" and then trusting whatever GA4 shows. The honest answer is: more than GA4 shows, and you have to set GA4 up correctly to see even that much. Both problems are fixable.

Step 1: build a custom AI channel group

GA4 does not, by default, group AI assistants into a clean channel you can watch. So make one:

1. Go to Admin, then Channel groups, and create a new custom channel group.
2. Add a channel called AI (or "AI Assistants").
3. Set the rule to match Source matches regex, and use a pattern covering the major AI hosts:

Host-based: chatgpt\.com|perplexity\.ai|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com|meta\.ai

Many practitioners use a broader token pattern to catch variants and subdomains: .*chatgpt.*|.*openai.*|.*perplexity.*|.*claude.*|.*gemini.*|.*copilot.*.

Then move the AI rule to the top of the ordering so those sessions get claimed by your AI channel before another rule grabs them.

"By default GA4 barely shows your AI traffic - and a big chunk of it is mislabeled as Direct."

Curious how AI engines describe your brand right now? Get a free visibility audit and see where you stand across ChatGPT, Gemini and Perplexity.

Step 2: understand the "Direct" undercount

Here is the catch that trips everyone up. Many AI clicks arrive with no referrer. Free-ChatGPT links and native-app clicks frequently pass no referring URL, so GA4 has nothing to attribute them to and files them under Direct. One 2026 analysis found about 36% of AI traffic arrived with no referrer.

The consequence: even a perfectly-built AI channel group shows you a floor, not the true total. A meaningful slice of your real AI traffic is sitting in Direct, indistinguishable from bookmarks and typed-in visits. Read your AI channel number as "at least this much," not "exactly this much."

Step 3: know about the native channel (and its limits)

Around 13 May 2026, Google added a native "AI Assistant" channel to GA4's default channel group. It helps, but it is not a complete fix: it rolled out gradually, it is not retroactive, and it does not cover every platform (Perplexity referrals, for instance). So keep your custom group - it applies to historical data and you control the coverage - and treat the native channel as a supplement, not a replacement.

What to do with the data

• Trend it, do not trust the absolute. Because of the Direct undercount, the direction and relative movement of your AI channel matter more than the exact number.
• Segment by engine. Break the AI channel into ChatGPT, Gemini, Perplexity and the rest, so you know which assistants actually send you people.
• Pair it with citation tracking. Referral traffic tells you who clicked through. It cannot tell you whether you were named in answers that never sent a click - which, as zero-click grows, is most of them. Measure citation directly too.

The takeaway

Two setup steps get you most of the way: build a custom AI channel group with the right regex, and remember that a big share of AI traffic hides in Direct with no referrer. Do both and you will see far more of your AI traffic than the default report shows - while remembering that clicks are only half the story, since the biggest AI exposure is being named in answers that never send a click at all.

Common mistakes that quietly break the setup

Most broken AI channels are not broken by a bad regex. They are broken by small ordering and scoping mistakes that GA4 never warns you about. If your AI number looks suspiciously flat or large, start here before touching the pattern.

• Rule order left at the bottom. Channel rules are evaluated top to bottom, and the first match wins. If a broad "Referral" or "Organic" rule sits above your AI rule, it can claim ChatGPT sessions before your AI channel ever sees them. Your regex can be perfect and still capture almost nothing.
• Anchoring the pattern too tightly. A rule like ^chatgpt\.com$ matches only that exact string and misses subdomains, tracking parameters and the mobile hosts. If you want reliability, match on a token that appears in the host rather than demanding an exact equality.
• Confusing source with medium. AI referrals usually arrive with a medium of "referral" and a source that is the AI host. Building the rule on medium instead of source lumps them in with every other referrer and defeats the point.
• Expecting it to be retroactive without checking. A custom channel group reprocesses historical data, but the native "AI Assistant" channel does not. Comparing the two over the same date range and finding a gap is normal, not a bug.

None of these throw an error. They just hand you a plausible number that is wrong - the most expensive kind of mistake in a report your team acts on.

A worked example: reading the report honestly

Say your custom AI channel shows 1,200 sessions last month, and Direct shows 40,000. The naive read is "AI is a rounding error, ignore it." That read is almost certainly wrong, and here is how to pressure-test it without pretending you have numbers you do not.

Start with the shape of Direct. A healthy site with steady branded search should have a fairly stable Direct channel. If Direct has climbed month over month while your obvious campaigns have not, some of that climb is very likely referrer-less AI traffic that GA4 could not label. You will not get an exact figure, but you can see the drift, and the drift is the signal.

Next, look at what those Direct sessions do. Filter Direct to landing pages that are deep and unlikely to be bookmarked - a comparison page, a pricing breakdown, a long guide. Bookmark and typed-in traffic clusters on the homepage and a few shallow pages. A stranger landing cold on your "X versus Y" page with no referrer looks a lot more like someone who just asked an assistant which tool to pick.

"The AI number GA4 hands you is a floor. The interesting question is how far above the floor you actually are."

The point is not a precise second number. It is to stop treating the AI channel figure as the ceiling. Once you read it as a floor and check Direct for corroborating shape, you make far better calls about where AI sits in your mix.

Why clicks are the smaller half of the story

Even a flawless GA4 setup only ever measures the people who clicked. That was a reasonable proxy in the search era, when ranking well and getting the click were nearly the same event. In AI answers they are not. The assistant reads your page, synthesises an answer, and often resolves the user's question inside the chat. The user never arrives, so GA4 never sees them, no matter how good your channel group is.

This blind spot grows as more queries end in an answer instead of a link. If your brand is named, summarised and recommended inside the response, that is a real marketing outcome with zero sessions attached to it. Conversely, you can win the click war on the small slice that clicks through and still be invisible in the answers that shape the buying decision.

So treat GA4 as one instrument, not the whole dashboard. It tells you who the AI channel sent to your door. It cannot tell you whether the assistant named you, cited you, or recommended a competitor while your page sat unread in its context window. Measuring that means watching the answers themselves - visibility a referral report was never built to give you.