You can stand up a working AI-referral view in GA4 in an afternoon, with no tagging changes. The build has four parts: a channel that captures visible AI referrals, an exploration that ties them to outcomes, an estimate for the dark portion, and a maintenance habit so it doesn't rot. This is the hands-on companion to why AI traffic shows up as "Direct". Read that first if you want the mechanics.
Step 1: Create the "AI / LLM referral" channel
In GA4 (as of mid-2026): Admin → Data display → Channel groups. Don't edit the default group. Create a copy, then add a new channel at the top of the rule order so it claims AI hits before Organic or Referral can.
Define it as Source matches regex:
chatgpt\.com|chat\.openai\.com|openai\.com|perplexity\.ai|gemini\.google\.com|
copilot\.microsoft\.com|bing\.com/chat|claude\.ai|you\.com|poe\.com|phind\.com|
duckduckgo\.com/aichat|meta\.ai
Two notes. First, rule order is the whole trick: if Organic Search sits above your AI channel, Google-property AI sources can get claimed there. Second, channel groups apply from creation onward (they don't rewrite history), so the sooner this exists, the sooner you have a trend.
Step 2: Tie the channel to outcomes
Reports answer "how much"; you want "is it worth anything." Build a Free-form exploration: dimension = your new channel (from the custom channel group), metrics = sessions, engaged sessions, key events (conversions), and revenue. Add landing page as a second dimension and you'll see which pages AI answers send people to. Usually deep content and PDPs, not the homepage.
This is the table that makes the case internally: AI-referred visitors arrive pre-recommended, so their conversion rate typically reads meaningfully above generic organic. Report it as revenue, not sessions. Sessions undersell it.
Step 3: Estimate the dark portion
The visits that lost their referrer are sitting in "Direct." You can't reclaim them, but you can bound them. Build a second exploration on Direct sessions with three filters:
- Landing page is not the homepage (bookmark/typed traffic overwhelmingly lands on
/). - New users (AI answers skew toward first-time visitors).
- Landing pages that match your citable content: the blog posts, guides, and PDPs that appear in AI answers.
Trend that segment next to your Step-1 channel. When they rise and fall together, you're looking at the same demand split across two buckets; the ratio between them is your house dark-traffic multiplier. It's an estimate. Defensible, directional, honest, and far better than reporting only the visible floor. (The reasoning behind each signal is in the companion piece.)
The channel you built in Step 1 is the floor. The Step-3 segment is the shadow. Report both, labelled as such. A floor plus an estimate beats a precise-looking number that's silently wrong.
Step 4: Keep it honest
- Update the regex when a new engine ships or a domain changes. Put a quarterly reminder on it. New AI sources appearing as plain "Referral" are your tell.
- Annotate launches (site changes, AI-visibility pushes) so movements in the trend have context.
- Don't over-read small numbers. Pre-scale, week-to-week noise is large; judge the monthly trend.
- Remember what GA4 can't see: answers that mention you but send no click. A zero-click recommendation still shapes the shortlist, which is why this dashboard pairs with answer-level metrics, not replaces them.
What you'll have by evening
One channel showing visible AI referrals and their revenue, one view estimating the hidden remainder, and a trend you can defend in a planning meeting. What it still can't tell you: which prompts, which engines, and which of your pages earn the answers upstream of the click. That's the answer-level layer. Share of voice, citation rate, sentiment, and tracking it daily across every engine is what Buffy Intel does.