There is a group of visitors on your site right now that your reporting cannot see as a group. They are scattered across Direct and Referral, mixed in with newsletter clicks and people who typed your URL.
They arrived from ChatGPT, Perplexity, Gemini, Copilot, or Claude. There are probably not many of them yet. They are also, in most accounts we look at, converting at a different rate than everyone else, and nobody knows in which direction because nobody has separated them out.
This is a measurement problem with a one-hour fix. Here is the fix, then what to do with what it shows you.
Why they do not behave like search visitors
A search visitor and an AI-referred visitor are at different points in the same journey.
The search visitor is still gathering. They opened four tabs, yours is one of them, and they are comparing as they read. Your page has to do the work of establishing who you are, what you do, and why you are credible.
The AI-referred visitor already had that conversation. They asked an assistant a question, got an answer that mentioned you, possibly alongside two competitors, and clicked through with a pre-formed impression they did not get from you. Some of what the assistant said about you was accurate. Some of it was assembled from sources you do not control.
They arrive later in the decision, in smaller numbers, and with different questions. Treating them as generic referral traffic wastes the most interesting signal available about how AI systems are describing your business.
Build the channel group
GA4 supports custom channel groups as rule-based categories built on source, medium, and campaign fields, usable as a dimension anywhere the default channel group works. That is all you need.
In Admin, under Data display, open Channel groups and create a new one. Copy the default group so you keep everything else intact, then add a new channel at the top of the order.
Channel name: AI Assistants
Condition: Source matches regex, using something like:
chatgpt\.com|chat\.openai\.com|openai\.com|perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|bing\.com/chat|claude\.ai|you\.com|poe\.com
Put the channel above Referral and Organic Search in the ordering, since channels evaluate top down and these hostnames would otherwise be swallowed by broader rules.
Two things to know before you trust the output.
Google AI Overviews traffic does not appear here. Clicks from an AI Overview are still Google organic search, indistinguishable in GA4 from a normal blue link click. What you can see there is the pattern described in the Ahrefs study of 300,000 keywords, where the presence of an AI Overview correlates with a sharply lower clickthrough rate for the top-ranking page. Impressions holding while clicks fall is your signal, not a channel label.
Assistants also do not all pass a referrer. Some traffic from AI tools lands in Direct no matter what you configure, so treat this channel as a floor rather than a complete count.
What to look at once it exists
Give it a month, then run four comparisons.
Conversion rate against organic. The volume will be small, so do not read too much into a single month. What you are watching for over a quarter is whether this group converts meaningfully better or worse than search traffic.
Landing pages. Which pages are assistants sending people to? This is the most actionable output of the whole exercise, because it tells you which of your pages the models consider authoritative on which topics. It is frequently not the page you would have chosen.
Pages per session and time on page. AI-referred visitors often either convert quickly or bounce quickly, because they arrive with a specific question already framed. A middling engagement number usually means the page did not confirm what the assistant told them.
Lead quality by source. Once you have enough volume, run these leads through the same disqualification reason codes as everything else. If AI-referred leads disqualify for a consistent reason, an assistant is describing your business inaccurately, and that is fixable.
What it changes about what you publish
The landing page data is the part that should change your content plan.
If assistants are sending people to a blog post rather than a service page, that post is doing your positioning for you and deserves a clear path to conversion. If they are sending people to a page that describes a service you no longer emphasize, the models are working from stale information about you.
Both are content problems with known fixes, which we cover in what makes content citable by AI engines. What the channel group adds is the feedback loop. Most businesses optimizing for AI visibility are flying without one, publishing and hoping. Measuring the referrals tells you which pages actually got picked up, and pairing that with tracking whether models mention your brand at all gives you both halves: whether you are mentioned, and what happens when someone acts on it.
What not to conclude
Three cautions, because this is new enough that overreaction is the default.
Small numbers are still small numbers. Forty sessions converting at 8% is not a finding, it is three conversions. Wait for a quarter before you redirect a budget.
Assistants change their behavior without telling anyone. A model update can shift which sources get cited across an entire category in a week. A drop in this channel is not necessarily something you did.
And do not rebuild your site for this traffic yet. For nearly every business we work with, AI referrals remain a small fraction of total sessions. The right posture is to measure it properly now, so that when it does matter you have a year of history instead of starting from zero.
If you want the visibility side of this handled deliberately rather than hoped for, that is what our work on AI and generative search visibility covers. And since these visitors still have to make it through your forms, your response time, and your qualification process like everyone else, the rest of the path is mapped in the full diagnostic from search to closed deal, with the measurement layer underneath it in measuring past the form.