How to Track ChatGPT and AI Referral Traffic in GA4

GA4 will not tell you what ChatGPT sends you unless you teach it to. Out of the box, a visitor who clicks through from chatgpt.com or perplexity.ai is filed under generic Referral traffic, mixed in with links from forums and news sites. The fix is a custom channel group — a reporting-layer change that takes about 20 minutes and touches no code on your site.

We built this channel on our own production site first, then wrote it up. The steps below are the setup we run, and the observations at the end are from our own build, marked [our data].

Why doesn't GA4 show ChatGPT traffic as its own channel?

GA4's default channel group has no definition for AI assistants, so their sessions match the generic Referral rule and disappear into it. The referrer arrives like any other website referrer — chatgpt.com is just a hostname to GA4 — and nothing in the default logic treats it differently.

This matters because AI-assistant sessions behave differently from forum or news referrals. The visitor already read an answer about their question inside the assistant; the click-through is a deliberate step past that answer. Blending those sessions into Referral hides the one channel most teams are currently asked to report on.

The fix is a custom channel group. It relabels sessions at the reporting layer, leaves your raw source and medium data untouched, and works in the standard Traffic acquisition report.

Which referrer hostnames identify AI assistants?

In production we match 6 hostnames, which cover the assistants that appear in our referral logs [our data]:

The regex we deploy, as one line:

chatgpt\.com|chat\.openai\.com|perplexity\.ai|claude\.ai|gemini\.google\.com|copilot\.microsoft\.com

Semrush's published GA4 method takes the same approach with a wider net — their pattern adds you.com, search.brave.com, and bard.google.com (Semrush, 2025). The wider list costs nothing in accuracy; we run the shorter one because the extra hostnames have not appeared in our own logs at meaningful volume [our data]. Start wide or start narrow — what matters is matching on source hostname, which survives changes to landing pages and campaign tags.

One distinction worth keeping straight: these hostnames identify humans clicking through from an assistant. Bot fetches — GPTBot crawling for training, or ChatGPT-User fetching a page live during a conversation — do not run JavaScript and never appear in GA4 at all (OpenAI bot documentation). Those show up in server logs, not analytics. Our complete AI crawler list covers that side.

How do you set up the custom channel group in GA4?

The whole setup happens in the GA4 admin, in 6 steps:

1. List your hostnames. Start from the 6 above, or Semrush's wider set.
2. Copy the default channel group. Admin → Data display → Channel groups → create a copy. GA4 only allows custom channels on a copy, never on the default group.
3. Add the channel. Create a channel named "AI Assistants" with one condition: source matches regex and the pattern above.
4. Order it above Referral. GA4 assigns each session to the first channel whose condition matches, top to bottom. If AI Assistants sits below Referral, it will never match anything.
5. Verify. In Reports → Traffic acquisition, switch the channel dimension to your new group. Cross-check by filtering session source for the same hostnames — the two views should agree.
6. Annotate the go-live date. Future-you will want to know exactly when the channel started separating cleanly.

How do you verify the channel is capturing AI traffic?

Filter the Traffic acquisition report by session source for any one hostname and compare it to the new channel's totals — the counts should reconcile. If the channel reads zero while the source filter shows sessions, the channel is sitting below Referral in the evaluation order, which is the single most common setup mistake we see.

A second check: open an assistant yourself, ask a question your site answers, and click your own citation if one appears. The session should land in the AI Assistants channel within the normal GA4 processing delay. It is a crude test, but it exercises the full path.

What did our production data show?

On our insurance build, we shipped the channel group alongside a first-party tag gateway in August 2026, and the first days of separated data are on the record [our data]. One observation, and one deliberate refusal.

The observation: volume is small. AI-assistant referrals are a thin slice next to organic search in the channel's early data [our data]. That is consistent with the best external measurement available: Pew found that when a Google search produced an AI summary, users clicked a source cited inside it in just 1% of visits (March 2025 browsing data, 900 U.S. adults, Pew Research Center, published July 2025). The assistant answers; the click is the exception.

The refusal: we are not publishing conversion quality yet, because the channel is weeks old and days of separated data cannot carry a finding. Answering that question honestly is the reason the channel exists, and the answer will be published here when it has months behind it — whichever direction it points. Conversion claims for AI traffic also carry heavy survivorship bias, because the sites receiving AI clicks are precisely the ones the assistant already vetted. Our page on whether AI citations are worth anything works through that reconciliation.

Is AI referral tracking worth setting up yet?

Yes, because it costs 20 minutes and nothing else — but be honest about what it will show. For most sites, the channel will read near zero for weeks. That is a true reading, not a broken setup, and it is worth having on the record before anyone proposes spending real money on AI visibility.

The channel also measures the wrong end of the funnel if citations are what you care about. GA4 sees clicks; it cannot see the far larger number of times an assistant mentioned or cited you without anyone clicking. For that you need direct observation — the free share-of-voice method — and for Google's AI surfaces specifically, the Search Console approximation. Referral tracking is one instrument on the panel, and the operator's guide to generative engine optimization shows where it fits among the others.