We were doing routine tag maintenance in a client’s Google Tag Manager container and found something that hadn’t been there a few months ago: an OpenAI conversion pixel, quietly installed and firing.
That pixel is real, and it’s new. OpenAI shipped an ads measurement pixel this year along with a Conversions API, and a small ecosystem of GTM templates has already grown up around it. If you’re buying ads inside ChatGPT, you can now measure them the same way you measure Meta or Google.
But that’s the small half of the story, and it’s the half that gets attention because it comes with a setup guide.
The larger half is this: most of the traffic AI assistants send to your site isn’t paid, and it doesn’t come with a pixel. Someone asks an assistant for a recommendation, gets your name, clicks through, and buys something. That’s real revenue arriving through a real channel. And most GA4 properties are counting a fraction of it.
In May, GA4 added AI Assistant as a default channel group. Traffic gets classified there when the medium is exactly ai-assistant, or when the referrer matches Google’s list of recognized assistants. If you haven’t looked at your channel report since spring, it’s already there.
It’s a genuine improvement. It’s also counting less than it appears to, for three reasons.
It requires a referrer. Traffic that arrives without one — mobile app clicks, in-app browsers, copied links — still lands in Direct, which is where measurement goes to die. That isn’t a small residual. App usage is a large share of how people talk to assistants, and it’s the share that’s hardest to see.
It isn’t retroactive. Sessions from before the change stay in whatever bucket they landed in at the time. Pull the last twelve months and you’re reading two different measurement regimes stitched together, with a seam in mid-May that nothing in the interface labels.
And the list is Google’s, not yours. Google hasn’t published the full set of referrers it recognizes, or said how it will be updated as new assistants launch. The documentation names a handful of examples. Whether the specific assistants sending you traffic are on that list isn’t something you can confirm from the outside — you can only check whether they show up in your data.
So the channel exists now, which is better than it was. What it reports is the portion of the traffic that arrived with its paperwork in order.
It’s tempting to treat the OpenAI pixel as the fix. It isn’t, for two reasons.
First, it only measures what you paid for. It’s a conversion pixel for an ad product. If you’re not running ads inside ChatGPT, installing it measures nothing.
Second, and more to the point: not every assistant sells ads, and the ones that don’t still send traffic. Anthropic published a formal commitment in February that Claude will stay ad-free — no sponsored links beside conversations, no advertiser influence on answers, no product placements the user didn’t ask for. So there is no Claude pixel, and unless that position changes, there won’t be one.
That doesn’t mean Claude sends you nothing. It means the traffic it sends can only be measured the ordinary way: by configuring your analytics to recognize it.
Which is the actual work here, and it applies whether or not you ever spend a dollar on AI advertising.
You can find out where you stand in an afternoon.
Start with the AI Assistant channel, if your property is showing one. That number is what GA4 could identify. Treat it as the low end. Note where it begins — if it starts abruptly in May, that’s the feature switching on, not your traffic appearing.
Then go looking for what it missed. Check your referral traffic for assistant domains that aren’t being rolled into the AI Assistant channel. Anything you find there is a gap in Google’s list as it applies to you, and it will keep accruing quietly until someone defines it.
Then look at Direct over the same window. A steady rise in Direct that doesn’t correspond to any campaign, any brand spend, or any email volume is worth taking seriously — and the new channel will not fix it, because the traffic causing it is arriving without the referrer the channel depends on.
Then decide whether you need a channel definition of your own. In most cases you do — not to replace the default, but to catch what it drops and to hold the definition steady over time rather than inheriting changes to a list you can’t see. Once that number exists, you can ask the questions that matter: whether this traffic converts better or worse than search, whether it’s landing on the pages you’d expect, whether it’s growing.
If you’re running server-side tagging, there’s a further step worth taking: preserving the referrer through the request, so traffic that loses it on the client side can still be attributed correctly. That’s the one change that addresses the Direct problem rather than working around it.
None of this is exotic. It’s the same work we do on every tracking audit — finding the places where real activity is happening and the measurement isn’t following it.
What makes this one worth flagging is that the channel grew quickly, the tooling arrived late, and a lot of properties are now running a configuration that is technically current and still undercounting. The organizations that sort this out will be able to say what AI-sourced traffic is worth to them. The ones that don’t will keep watching Direct go up.
We do the data engineering and measurement work behind this — tagging, channel definitions, server-side configuration, and the reporting layer that makes the answer usable. If you want to know what your setup is currently missing, that’s a short conversation.
