Channel groups, referral exclusions, and a practical table of AI referrers worth watching — without over-reading small samples.
AI chat products send some referral traffic. Similarweb / TechCrunch reported in June 2025 that AI platforms sent about 1.13 billion referrals to the top 1,000 sites that month (up 357% year over year), with ChatGPT responsible for more than 80% of those AI referrals. Google Search still sent about 191 billion referrals in the same month. AI is growing; it is not replacing Search in the analytics account overnight.
This playbook sets up GA4 so you can see AI referrers without drowning in noise or over-reading a dozen sessions.
What you are trying to learn
You want a durable view of sessions that arrive from known AI hosts — ChatGPT, Perplexity, and peers — so you can tell whether access and content work correlates with visits. You are not trying to prove causation from a single spike, and you are not treating BrandKnown's readiness score as a traffic forecast.
Similarweb's Gen AI Landscape 2025 (US desktop, September 2025) reported ChatGPT referrals averaging about 15 minutes on site versus about 8 for Google, about 12 pages per session versus about 9, and about 7% conversion on transactional sites versus about 5% for Google. Those benchmarks explain why the channel is worth a custom report. Your site may look nothing like the top-site average.
Referrers worth watching
Hostnames change. Treat this table as a starting allowlist you maintain, not scripture.
| Referrer pattern (examples) | Typical product | Notes |
|---|---|---|
chat.openai.com | ChatGPT | Common web referrer |
chatgpt.com | ChatGPT | Newer hostname; include both |
perplexity.ai | Perplexity | Often visible citations |
gemini.google.com | Gemini | May appear for some shares/opens |
copilot.microsoft.com | Copilot | Watch volume before celebrating |
claude.ai | Claude | Lower volume for many sites |
Add hosts only when you see them in Acquisition → Traffic acquisition → Session source/medium or in Explorations. Do not invent a long list of domains you have never received.
Step-by-step GA4 setup
1. Confirm referrals are not stripped
In Admin → Data streams → your web stream → Configure tag settings → List unwanted referrals, make sure you have not excluded AI hosts by accident. Referral exclusions are for payment gateways and self-referrals, not for ChatGPT.
2. Find raw AI sessions
- Open Explore → Free form.
- Dimensions: Session source, Session medium, Landing page.
- Metrics: Sessions, Engaged sessions, Conversions (pick one that matters).
- Filter Session source contains
openaiORchatgptORperplexity(extend as needed). - Save the exploration as "AI referrals — raw."
3. Build a channel group or comparison
For ongoing reporting, create a custom channel group (or a comparison) that buckets AI referrers:
- Admin → Data settings → Channel groups (or use Comparisons on standard reports if you lack edit access).
- Add a rule: Source matches regex for your maintained host list, Medium equals referral.
- Place the rule above generic Referral so AI is not swallowed by the default Referral channel.
- Publish and wait for new data (channel groups apply going forward).
4. Register a useful audience (optional)
Create an audience of users whose first session source matched your AI regex in the last 30 days. Use it for path exploration, not for blasting ads on day one.
5. Annotate site changes
When you allowlist a bot, ship SSR, or rewrite a definition page, add a GA4 annotation or a spreadsheet note. Without annotations, you will invent stories about noise.
Decision table: how to read the numbers
| What you see | Sensible read | Bad read |
|---|---|---|
| 0–20 sessions / month | Tracking works; volume too small for CRO claims | "AI is dead for us" |
| Steady climb after access fixes | Directional win; keep watching | "We fixed rankings" |
| Spike with one landing page | That URL is citeable; improve its siblings | Rewrite the whole site tomorrow |
| High engagement, low convert | Content fit; offer/path issue | Delete the channel report |
| Referral from unexpected host | Add to allowlist if repeatable | Ignore forever |
Common GA4 pitfalls
- Self-referral pollution: Broken cross-domain tracking makes your own site look like the referrer. Fix measurement before blaming AI.
- Missing hostname variants: Tracking only
chat.openai.comwhile traffic shifts tochatgpt.com. - Default channel group only: AI sits inside Referral and never gets a slide in the client deck.
- Over-filtering bots in reporting identity: You want human sessions from AI products, not to hide all "bot-like" traffic so aggressively that real users vanish — know what your filters do.
- Calling branded direct traffic "AI": Direct is not proof of an assistant citation.
Tie tracking to AEO work
Tracking without a work loop is sightseeing. Pair GA4 with:
- Access audits for ChatGPT-User vs GPTBot and PerplexityBot.
- A prompt notebook for naming and citations.
- Branded query checks described in measuring branded search lift.
Agency teams can reuse one exploration template across clients, swapping property IDs — see the ten-client AEO workflow.
Honest ceilings
- Not every assistant click sends a referrer.
- In-app browsers and copy-paste visits look like Direct or Other.
- Top-1,000-site aggregates (Similarweb / TechCrunch, June 2025) will not match a mid-market SaaS.
- A BrandKnown scan does not create GA4 hits; it only flags website issues that block eligibility.
Set up the channel, maintain the host list, annotate changes, and refuse to narrate causation from a handful of sessions. That is enough analytics maturity for AI referrals in 2026.