If you opened a GA4 property recently and found a channel called AI Assistant that you never created, that's Google's doing. On May 13, 2026 it became part of the default channel group, and it needs zero configuration.
Here's the short version: GA4's AI Assistant channel automatically splits ChatGPT, Gemini, Deepseek, Copilot and Grok traffic out of Referral by stamping the session with medium ai-assistant — but it deliberately excludes Google's own AI Overviews and AI Mode, it isn't retroactive, and it doesn't name Perplexity or Claude. So it's a real improvement and an incomplete answer at the same time, and if you report it as "our AI traffic" you'll be wrong in both directions.
We've been through this on client properties for the last few months. Below is what the channel does, the three gaps it leaves, and the setup we now run instead.
What the AI Assistant channel actually does
Mechanically it's simple, which is why it works without setup. When a session arrives with a referrer GA4 recognises as an AI assistant, three things happen automatically:
- The medium is set to
ai-assistant. Notreferral, not(none). - The campaign is set to
(ai-assistant). A reserved value, so it won't collide with your UTMs. - The session is bucketed into the AI Assistant channel in the default channel group.
Google's default channel group documentation defines it as the channel "by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok," with the rule being that the medium exactly matches ai-assistant.
That's genuinely useful. Before this, chatbot traffic sat in Referral next to newsletters, partner sites and forum links, and every agency that wanted to report it had to build a regex. Now the baseline is free.
The number itself is usually smaller than clients expect. Across the properties we've looked at this year, AI Assistant sessions land in the low single-digit percentage of total sessions for most service businesses — but they convert noticeably better than the average referral, because someone who asked a chatbot a buying question and then clicked through is already deep in intent. Small volume, good quality. That's the pattern worth reporting, not the raw session count.
Gap one: it excludes the AI surface that matters most
Read the definition again and you'll notice what isn't in it. Google AI Overviews and AI Mode traffic stays inside Organic Search. They are not part of the AI Assistant channel and never were.
This is not an oversight. Google's own AI surfaces don't pass a distinguishing referrer, so there's nothing for the rule to match on. But it produces an absurd situation for anyone reporting on "AI traffic": the channel named after AI excludes the AI product with by far the widest reach, and the one most likely to be cutting your clicks.
For most of our clients, third-party chatbots are a rounding error next to what AI Overviews does to a page that used to rank third. We wrote about the visibility side of this in the AI search citation gap — the measurement side is the same problem wearing a different hat.
So if a client asks "how much traffic are we getting from AI?", the AI Assistant channel answers a much narrower question than the one they asked. Say so before you hand them the number.
Gap two: it isn't retroactive, and that breaks your trend line
The channel started collecting when the rollout reached your property. It did not go back and reclassify old sessions. Traffic from ChatGPT in March 2026 is still sitting in Referral, exactly where it landed.
Two consequences, and the second one causes arguments:
- Any date range before your switchover shows zero AI Assistant sessions. Not "low" — zero. That's a reporting artefact, not a finding.
- Your Referral channel drops on the switchover date. If nobody annotated it, a year-over-year Referral comparison shows a decline that looks like lost partner traffic, and someone will spend a week investigating it.
We've now seen that second one play out twice on inherited accounts. Both times the "referral collapse" was just the AI split doing its job. Annotate the date in the dashboard and the conversation ends in thirty seconds instead of a week. It's the same failure mode we hit with consent mode and the conversion drop that wasn't real — a definitional change that reads as a performance change.
Gap three: the recognised source list is shorter than the market
Google names five: ChatGPT, Gemini, Deepseek, Copilot, Grok. Perplexity isn't on that list. Neither is Claude.
Perplexity in particular is a search product that sends real referral traffic and is very much in the consideration set for the kind of research query that turns into a lead. If you report the native channel as your AI number, Perplexity sessions stay in Referral and quietly don't count.
There's also a category the rule simply can't catch: assistants and AI browsers that strip the referrer, or that fetch and summarise your page without a click at all. Those sessions land in Direct, and no channel rule will ever pull them back out. We covered that failure mode in more depth in what AI browsers are doing to ad attribution.
The setup we run instead
Two channel definitions, not one, plus a proxy for the thing GA4 can't see.
Keep the native channel. It costs nothing, it's clean going forward, and it's the definition Google will keep updating as it adds sources. Use it as your forward-looking number.
Add a custom channel group for AI traffic. This is the important one. Custom channel groups in GA4 are applied at report time, which means they rebucket your existing data — the moment you create it, you can see AI traffic going back as far as your data goes. Match source against a regex covering chatgpt|openai|gemini|bard|perplexity|claude|anthropic|copilot|deepseek|grok, and keep it in one channel rather than splitting per-tool until the volume justifies it.
That single group gives you two things the native channel can't: history, so you have a trend instead of a start date, and coverage, so Perplexity and Claude stop hiding in Referral.
Add a Search Console proxy for AI Mode. Since GA4 won't separate it, approximate it where the data does exist. Pull the Pages report in Search Console, compare a rolling window to the same window a year earlier, and watch impressions against clicks on your informational pages. A page holding impressions while shedding clicks is being answered above the results. That impression-to-click gap is the honest number, and it's what we put in the report — clearly labelled as an estimate, because it is one.
If you already run server-side tagging, capture the raw referrer. Anyone on a server-side GTM setup can log the full referrer header before GA4's rules touch it, which gives you ground truth on assistants Google hasn't classified yet. If you're weighing that infrastructure anyway, we broke down the options in self-hosted versus managed server-side GTM.
What to change in client reporting this week
Four things, in order, and none of them take long:
- Annotate the AI Assistant switchover date on every dashboard that shows a Referral trend. Do this first — it prevents the fake-crisis conversation.
- Build the custom channel group so the AI line has history behind it.
- Rename the report section. "AI Assistant traffic" is accurate. "AI traffic" is not, because AI Mode isn't in it. That one word saves you a bad meeting.
- Add the Search Console impression-to-click gap as a separate row, labelled as an estimate.
The bigger shift underneath all of this: the number of people who read about you inside an AI answer and never click is now larger than the number who click. Measurement can only chase part of that. The rest is upstream — being the source the model quotes in the first place, which is a content problem, not an analytics one. That's the case we made in publishing daily content that actually gets cited.
Get this set up properly
If your GA4 is still reporting AI traffic as one uninterrupted line, or your Referral channel took a mystery dip in May and nobody has explained it, that's a ninety-minute fix, not a project.
Get a Free Automation Audit — we'll look at your GA4 channel setup, your Search Console trend, and where your attribution is quietly losing sessions, and tell you exactly what to change. No pitch deck.