How to Track Traffic From ChatGPT, Perplexity and Other AI Assistants in GA4

Everyone asks whether AI assistants send traffic. Few sites can answer, because the numbers hide in the referral report under names nobody grouped. This post shows how to group them in GA4, gives you a script to classify a referrer export, and spells out what no report will tell you.

The classifier script was run on invented sample data to demonstrate the logic. The counts below are not real traffic and say nothing about how much AI traffic any site gets.

How does GA4 classify traffic by default?

Google's help page on channel groups says GA4 sorts traffic into channels with rule-based definitions, and that default channel definitions cannot be edited. Referral is described as non-ad links on other sites or apps, with a manual-traffic rule of medium equal to referral, app or link. Custom channel groups, per the same page, let you create your own rule-based categories of web-traffic sources.

I could not confirm from that page exactly which match types a custom rule supports, so open the custom channel group editor in your property and check that you can match on Source with "contains" or "matches regex" before you plan around it.

How to build an AI assistants channel group

1. In GA4, open the channel group settings for your property and create a custom channel group.
2. Add a channel named AI assistants with a rule on the source dimension that matches your list of hostnames.
3. Put it above broader rules. Order matters: if "Search" matches google.com first, gemini.google.com visits land in Search. My script below hit this exact trap.
4. Save, then look at it in Acquisition reports as a secondary dimension. Expect it to apply from the point you set it up in the way GA4 documents, so check the help page about whether past data is reprocessed.

A referrer classifier you can run on an export

If you prefer to analyse a CSV export of referrers and sessions, this script groups hostnames with ordered rules, strips www., and prints hosts it could not classify so you can extend the list.

RULES = {
    "ai-assistant": ["chatgpt.com", "perplexity.ai", "gemini.google.com", "copilot.microsoft.com", "claude.ai"],
    "search": ["google.com", "bing.com", "duckduckgo.com"],
    "social": ["linkedin.com", "t.co", "reddit.com", "youtube.com"],
}

def classify(h):
    for label, hosts in RULES.items():
        if any(h == d or h.endswith("." + d) for d in hosts):
            return label
    return "direct" if h.startswith("(") else "other"

The dict is ordered, so ai-assistant is checked before search. Output on the invented sample (9 referrer rows, 273 sessions):

search         142  52.2%
direct          80  29.4%
ai-assistant    28  10.3%
social          17  6.2%
other            5  1.8%
unclassified hosts to review: {'news.ycombinator.com': 5}
gemini.google.com -> ai-assistant

The last line is a regression check: with the rules in the other order, gemini.google.com would count as search. The hostname list is my starting guess. Replace it with hosts from your own report.

What will no report show you?

Four blind spots, so you do not over-read the numbers:

• AI Overviews clicks look like Google search. Google says AI feature traffic is included in overall Search Console data, in the Performance report under the Web search type, and describes no separate breakdown. Do not expect a clean split.
• Zero-click answers. A user reads your fact in an assistant and never visits. No referrer, no session.
• Missing referrers. Some visits will arrive as direct. I did not find a primary source for how often, so I give no percentage.
• Brand lift after the answer. The later branded search is not attributed to the assistant.

So report AI referrals as a floor, never the whole effect.

What else should you measure alongside it?

• Citation tests. Ask a fixed set of questions in each assistant monthly and log whether you are cited. The method is in the AI citation test protocol.
• Landing pages. Which pages receive AI referrals? They show what the assistants find quotable, and they feed the rewrite list from the AEO checklist.
• Branded search impressions in Search Console, month on month.
• Server logs for crawlers such as OAI-SearchBot, if you can access them.

If you do not want GA4 at all, cookieless analytics like Plausible or Umami show referrers on a single page, and the same hostname grouping applies.

A monthly routine that takes twenty minutes

1. Export referrers and sessions for the last 30 days.
2. Run the classifier. Review the unclassified list and add any AI hosts.
3. Record the AI assistant share and the top five landing pages.
4. Run your ten citation prompts.
5. Compare with last month. Do not act on a change smaller than the month-to-month noise you have already seen.

What I could not verify

• Which hostnames each assistant sends today. My list is a starting guess.
• GA4's handling of historical data when you add a channel group.
• Any figure for how much AI traffic a typical site receives. I give none.