Docs

GA4's AI Assistant Channel: Catches and Gaps

GA4 now auto-tags ChatGPT, Gemini and Copilot traffic as an AI Assistant channel — but it misses Direct sessions and skips Perplexity entirely.

Acquisition

Google Analytics 4 quietly solved half a problem this year. Since a rollout that reached full availability in early June, GA4 now automatically reclassifies sessions from ChatGPT, Gemini, Copilot, Grok, and DeepSeek out of the catch-all Referral bucket and into their own Default Channel Group, tagged AI Assistant with medium ai-assistant. No configuration, no custom channel group rules — it just appears in reports that used to lump this traffic in with backlinks and forum referrals.

That’s a genuine improvement, but treating it as the finished picture is a mistake. The classification only covers sessions that arrive with a matching referrer header, and it’s forward-only — traffic before the mid-May cutover was never reclassified, so month-over-month comparisons of the new channel are comparing a full month against a partial one. More importantly, Perplexity isn’t in Google’s official list, so it still lands in ordinary Referral, and any AI assistant session that arrives without a referrer at all — increasingly common as assistants strip referrer data or open links in ways that don’t pass one through — still falls into Direct, indistinguishable from someone who typed your URL from memory. Teams that stop instrumenting because “GA4 handles this now” will keep undercounting exactly the traffic this channel was meant to surface.

Data Points to Track

  • Referrer domain, captured raw before GA4’s channel grouping logic runs, so you can audit which assistants are and aren’t being classified correctly as the list of recognised referrers changes
  • Landing page path for AI Assistant sessions specifically, since this traffic tends to land deep on product or pricing pages rather than the homepage
  • Session engagement rate and pages-per-session, segmented by ai-assistant medium versus ordinary Referral and Direct, to see whether AI-referred visitors behave differently
  • Query or prompt fragment, where an assistant passes one, logged as a custom parameter rather than discarded — this is the closest thing to search-term data this channel offers
  • First-touch vs. this-session channel, so a user who first arrived via AI Assistant but converts on a later Direct visit doesn’t get silently attributed away from where they actually started

Setup Steps

  1. Audit your current Direct traffic for AI-shaped sessions before assuming the new channel has closed the gap — look for spikes in Direct sessions landing on deep, non-homepage URLs, which is the signature of an AI referral with a stripped referrer.
  2. Add a custom dimension for raw referrer so you retain the pre-classification signal even after GA4 buckets a session into AI Assistant, Referral, or Direct — this is what lets you catch reclassification changes early.
  3. Build a saved segment for medium = ai-assistant and compare its conversion rate and average order value against your other acquisition channels on a monthly cadence, not just page views.
  4. Cross-reference server logs or a lightweight edge function for known AI crawler and assistant user-agents, since GA4’s channel only covers referred human sessions, not the separate question of AI agents fetching your pages directly.
  5. Re-run the audit whenever Google updates the AI Assistant referrer list — it has already changed once since launch, and a channel definition that changes underneath you will quietly shift historical comparisons.

Actionable Insights

Once raw referrer data sits alongside GA4’s own classification, a sudden drop in ai-assistant sessions combined with a rise in Direct is a strong signal that an assistant changed how it passes referrer data, not that AI-driven traffic actually declined. A high engagement rate but low conversion rate on AI Assistant sessions landing on product pages usually points to a mismatch between what the assistant told the user and what the page actually delivers, worth checking against the answer text if you can retrieve it. And because the channel is forward-only, treating everything before mid-May as a rough undercount rather than a true zero baseline avoids drawing false “AI traffic just started” conclusions from what’s really a measurement cutover.

Expert help

Need help tracking this in your app?

Our team sets up analytics pipelines for mobile and web teams every day. Talk to us and get your first events flowing in under an hour.

Talk to an expert