You opened your acquisition report sometime in June and there was a channel you did not create. AI Assistant. It already had sessions in it. Nobody on your team set it up.
For about ten minutes that felt like a win. Two years of ChatGPT clicks buried inside Referral, finally named.
Then you looked at the number next to it, and you were exactly where you started.
Your traffic was never unlabelled. It was unexplained.
Google shipped a nameplate on May 13, and the industry called it a fix
Here is what actually happened. On May 13, 2026, Google added an AI Assistant channel to GA4's Default Channel Group. When a visit arrives with a referrer Google recognises, the medium is set to ai-assistant, the session lands in the new channel, and the campaign gets stamped (ai-assistant). No configuration. It reached most properties by early June.
Every SEO blog on the internet wrote the same post about it. No more regex. No more burning a custom channel slot. AI traffic has grown up.
They are describing a naming fix: a change that gives your problem a label and stops there. The bucket got a better name. The bucket still does not tell you anything.
That distinction matters more for you than for the agencies writing those posts. They have an analyst who will go read the row. You have eleven minutes between a support email and a deploy.
A channel that starts counting on launch day has no history, and history is the entire point
The AI Assistant channel is not retroactive. GA4 processes data daily, so everything before May 13 stays filed wherever it landed at the time. Your launch week is in there. Your first Show HN spike is in there. The three months where you were actually trying things and needed to know which one worked are in there, sorted under Referral next to a directory listing and a forum backlink.
It is also incomplete in a way nobody flags on the way in. Perplexity is absent from Google's official channel definition and still lands in Referral. AI Overviews count as Organic Search. And any session where the referrer header never arrives, which is most of them when a link opens in an in-app browser, is still Direct and always was.
The Digital Bloom's early-2026 report put that last number at 70.6% of AI referrals showing up in GA4 as direct. Google just improved the visible minority of a channel and left the majority where it was.
So the row is real, it is new, and it is a floor rather than a measurement.
Knowing 312 people came from ChatGPT is not knowing what question you were the answer to
Take a solo dev running an AI receipt parser for freelancers. Nineteen hundred dollars MRR, eight months old, shipping alone.
In June the new channel tells her 312 sessions came from AI assistants. Good. She now knows a number she did not know in May.
What she needs to know is different. Which prompt produced those 312 sessions. Whether ChatGPT is describing her product correctly or confusing it with the expense tool she competes with. Why those visitors landed on a comparison page she wrote as an afterthought instead of the homepage she rewrote four times. Whether the eleven that signed up came disproportionately from one assistant, which would mean writing for that assistant is worth an afternoon and the other three are not.
None of that is in the row. The row is a count. Attribution, the way the industry uses the word, means correctly sorting visitors into named piles. That was the job when the piles were Google, Facebook and email, and you could infer intent from the pile name. It is a much thinner job now. A visitor from ChatGPT arrived because a model read something, decided you were the answer, and said so. The interesting object is the recommendation. The click is its shadow.
Conductor's January 2026 benchmark found AI-referred sessions run about 2.3 times longer than sessions from traditional search. These are not people browsing. They arrived already convinced by something you did not write and cannot see.
You are building an AI product and you cannot tell whether AI can find you
The volume is small and moving fast, which is exactly the condition under which a baseline is worth having. SE Ranking tracked 101,574 sites and found ChatGPT's share of referral traffic went from 0.23% in April 2026 to 0.32% in May. The jump traces to around May 7, when ChatGPT started surfacing links more prominently in answers. Similarweb measured referral visits climbing roughly 150% comparing the weeks either side.
That is one product change, at one company, in one week, moving a channel by half again. It will move like that repeatedly, and you will find out afterwards.
Meanwhile there is a founder on Indie Hackers this year describing the moment his conversions fell while his sessions held steady, asking which reality was lying: "My dashboard or my eyes?" Underneath, somebody else pointed out that once you add a consent banner you are no longer observing your users. You are observing the ones who agreed to be observed.
Twelve months from this Tuesday, if nothing changes: you will have shipped four more launches and still be reading a row that counts clicks. A competitor will have spent the same year learning which questions the models hand them and writing directly into those answers. The gap will not announce itself. It will just be there in the signup graph, and you will call it luck.
Five things worth doing this week, none of which need a new tool
Ask the assistants the five questions your product is supposed to be the answer to. ChatGPT, Claude, Perplexity, Gemini. Write down whether you appear, and what it says you do. That is your AI visibility baseline and it costs ten minutes.
Export ninety days of Referral and Direct out of GA4 and sort by landing page. Not by source. The pages assistants cite are almost never your homepage, and the list of what they cite is a content brief you did not have to think up.
Build the custom channel group anyway, with a Perplexity source rule, ordered above Referral. The native channel does not cover it and does not cover your history. Ten minutes, once.
Separate crawlers from humans before you believe any launch spike. Pull your server logs for launch day and count what identified itself as a bot. Most of a Show HN spike is machines, and the celebratory screenshot you posted was mostly machines too.
Add one free-text field to signup: how did you hear about us. It is the only instrument that catches the referrer-less majority, and after forty responses it will disagree with your analytics in a way worth a Saturday morning.
The arrivals board
GA4 added a flight to the arrivals board. It does not tell you who got off the plane, or why they flew, or what someone told them about your city before they booked.
That is the whole gap. The industry spent May celebrating a better-labelled board, and the question you actually walked in with was never a labelling question. You wanted to know what to do on Wednesday.
Clerion breaks out ChatGPT, Perplexity, Claude, Gemini and Copilot as their own channel, and then writes you a plain-English briefing about what happened in it, using real figures rather than generated ones. No cookies, no banner, one script tag, and the whole product on the free plan. If the five things above turn up something you want explained rather than counted, it is about a minute to install.