Key Takeaways
- Your AI referral traffic is already in GA4, but it is split between the Referral bucket and Direct rather than labeled as AI.
- GA4 added a native AI Assistant channel on 13 May 2026 that needs no setup, though it does not recognize Perplexity.
- A custom channel group closes the gap in about fifteen minutes, as long as you place the rule above the built-in Referral rule.
- Between roughly a third and two thirds of AI sessions arrive with no referrer and land in Direct, which no configuration can recover.
- Referral traffic is a lagging indicator; how much of your catalog an assistant can read is the one that moves first.
Your AI Referral Traffic Is Already There, Filed Under the Wrong Name
The question comes up in every conversation about AI shopping, usually with a slightly defensive edge: is any of this actually sending me traffic?
Fair question. The reason it is hard to answer is not that the traffic does not exist. It is that your analytics were built for a world where every visit arrived with a return address, and a good share of AI visits do not.
Some of it is sitting in GA4 right now under Referral, mixed in with everything else. Some is in Direct, indistinguishable from someone typing your URL. Almost none of it is labeled in a way that lets you say “this came from an assistant” without doing some work first.
The work is worth doing and takes about fifteen minutes. What matters more is knowing what the number means once you have it, because it will still be an undercount, and it will still be describing last month.
What GA4 Fixed in May 2026, and What It Left Out
Google closed part of this gap. On 13 May 2026 it added a native AI Assistant channel to GA4’s Default Channel Group, and unlike most analytics improvements it needs no configuration at all. When GA4 recognizes a referrer as an AI assistant it stamps the medium as ai-assistant and files the session under its own channel, so chatbot traffic stops hiding inside Referral.
That is a real improvement. Open your Default Channel Group report and you may already have a row you have never noticed.
The catch is what the channel recognizes. It covers the big assistants, ChatGPT, Gemini and Claude among them. It does not cover Perplexity, which keeps landing in plain Referral. For a merchant that is an awkward omission, because Perplexity’s shopping users are among the most purchase-minded traffic any AI surface sends. The default channel will show you a number while quietly leaving out one of the sources you would most want to see.
So treat the native channel as a floor. If you want the full picture, you build it yourself.
Building a Custom Channel Group That Catches Everything
In GA4, go to Admin, then Data display, then Channel groups, and create a new group. Add a channel called AI Assistants and give it one condition: Source matches regex.
For the pattern, start with the hosts that actually turn up in shopping journeys, and extend it as you spot new ones in your Referral rows:
chatgpt\.com|chat\.openai\.com|perplexity\.ai|gemini\.google\.com|claude\.ai|copilot\.microsoft\.com
Then the step everybody misses. Drag your new AI rule above the built-in Referral rule. GA4 evaluates channel rules in order and stops at the first match, so if Referral sits above yours it claims the traffic first, and your new channel sits there reading zero while you wonder what you got wrong.
Two things to know before you lean on it. Custom channel groups do not backfill, so the data starts the day you create the group and your history stays as it was. And this is a reporting layer rather than new collection, so it reorganizes what GA4 already captured. Which brings us to the part no configuration fixes.
The Bigger Gap: Visits That Arrive With No Return Address
A large share of AI referrals never identify themselves at all.
When someone taps a citation inside the ChatGPT mobile app, the in-app browser on iOS or Android often passes no referrer header. Several desktop apps do the same. With no referrer, GA4 has nothing to attribute and drops the session into Direct, where it is indistinguishable from a bookmark or a typed URL. Independent analyses put this somewhere between roughly a third and two thirds of AI sessions, depending on the platform mix. Even at the low end, that is a lot of buyers you cannot see.
No regex fixes this. The information was never sent.
There is a second and larger blind spot underneath it. Analytics only ever counts people who clicked. It cannot count the shopper who asked for a recommendation, got your product named in the answer with its price and a description, and decided from that. Some of them buy later by searching your brand directly. Some were never going to click at all. Either way, being recommended is the thing that happened, and the click is only the part your analytics catches.
Being named in the answer is the outcome. The click is just the fraction of it your analytics can see.
Why This Number Lags, and What to Watch Instead
Set all this up and you will have a chart. It will trend, and it is genuinely useful for arguing internally that AI discovery deserves attention. Be honest with yourself about what it can steer, though.
It tells you what already happened, to the products that were already readable, over a window that has already closed. Change your product data today and this chart will not move for weeks, and when it does you will not be able to attribute the change cleanly, because the size of the undercount shifts with your traffic mix.
For steering you want something that moves before the traffic does, and there is an obvious candidate. An assistant can only recommend products whose data it can read, so the question that leads the traffic is not “how many sessions came from ChatGPT last month” but “how much of my catalog is currently in a state where an assistant could recommend it at all.” That you can answer today, across the whole catalog, without waiting for anything. It is the difference between checking the scoreboard and checking whether your team is on the pitch.
So use both, in that order. Set up the custom channel group, because a directional number is worth having and fifteen minutes is a cheap price. Then manage against readiness. UCP Radar scores every product in your catalog from 0 to 100 on how ready it is for AI surfaces and reports it as a before and after around each optimization run, so the improvement is a number rather than a hope. Most catalogs land between 25 and 40 before anyone touches them, which is why so many get skipped. That figure moves the week you fix your data, and it covers every product rather than the handful that happened to get clicked.
The two then work together. Readiness tells you whether you are eligible to be recommended. The referral chart, undercounted as it is, tells you the market is responding. Nobody can promise you a ranking, but the merchants turning up in those answers are the ones whose data could be read.
Conclusion
If AI referral traffic looks small in your reports, the likeliest explanations are that it is misfiled, that it arrived without a referrer, or that the recommendation never needed a click to do its job. Probably all three at once.
Fix the reporting so you can see the shape of it, then stop using it as your steering wheel. The number worth managing is how much of your catalog an assistant can read right now, because that one moves first and you control it directly. It is the same shift generative engine optimization asks for everywhere else: optimize the data, not the report. If you want to see where yours stands, UCP Radar will score your feed and show you which products are currently unreadable.
Frequently Asked Questions
Since 13 May 2026, GA4 has a native AI Assistant channel in the Default Channel Group that captures ChatGPT, Gemini and Claude with no setup required. For fuller coverage, create a custom channel group under Admin, Data display, Channel groups, with a channel whose Source matches a regex of AI hosts. Place that rule above the built-in Referral rule, otherwise Referral matches first and your channel stays empty.
Google's native AI Assistant channel does not recognize Perplexity, so those sessions still land in ordinary Referral. That matters for merchants because Perplexity's shopping users tend to be high intent. A custom channel group with perplexity.ai in the regex is the way to pull it into the same reporting line as the other assistants.
Because a large share of AI referrals arrive with no referrer header. In-app browsers inside the ChatGPT mobile app, and several desktop apps, pass no referrer at all, so GA4 has nothing to attribute and files the session under Direct. Independent analyses put this somewhere between roughly a third and two thirds of AI sessions. No configuration recovers it, because the information was never sent.
No. Custom channel groups do not backfill. Reporting starts from the day you create the group, and your historical data stays classified the way it was. That is a good reason to set one up now rather than when you next need the number, since the history you build is only as long as the group has existed.
It is useful but lagging and incomplete. It counts only sessions that clicked through and were correctly attributed, missing both the unattributed visits and every shopper who acted on a recommendation without clicking. A leading indicator is how much of your catalog is currently readable enough to be recommended, which you can measure today rather than waiting weeks for a traffic chart to respond.
