How to see AI referral traffic in GA4 across every assistant

The hosts worth watching

AI assistants send traffic from a handful of hostnames, and the list is short enough to hold in your head. ChatGPT sends from chatgpt.com, and historically from chat.openai.com, which still appears in older exports. Perplexity sends from perplexity.ai and its regional subdomains. Google's Gemini sends from gemini.google.com, and older sessions carry bard.google.com. Claude sends from claude.ai. Microsoft Copilot sends from copilot.microsoft.com, and Copilot answers served inside Bing refer as bing.com.

Newer entrants worth adding: grok.com, chat.deepseek.com and meta.ai. None of these will be large for most sites yet, but adding them costs nothing and means you notice the day they start.

One of these deserves a warning. Traffic referred as bing.com may be a Copilot answer or an ordinary Bing search result, and nothing in the referrer distinguishes them. If you fold bing.com into an AI channel you will overstate; if you exclude it you will understate. Pick one, write down which, and be consistent.

One expression that catches them all

In Admin, then Data display, then Channel groups, create a new group and add a channel with a Source condition matching a regular expression. This one covers the set above:

chatgpt\.com|chat\.openai\.com|perplexity\.ai|gemini\.google\.com|bard\.google\.com|claude\.ai|copilot\.microsoft\.com|grok\.com|chat\.deepseek\.com|meta\.ai

Escape the dots. An unescaped dot matches any character, so chatgptxcom would match too — harmless in practice, but the habit matters when a pattern is doing something less forgiving.

Order the AI channel above Referral in the group, because GA4 evaluates top down and stops at the first match. This is the single most common reason a correctly written channel appears to do nothing.

Splitting by assistant rather than lumping them

A single AI assistant channel answers whether the category sends anyone. It does not tell you which assistant, and the assistants behave differently enough that the distinction is worth having.

For a per-assistant view, use an exploration rather than more channels. Reports, then Explore, then a free-form exploration with Session source as the row dimension and Sessions, Engagement rate and Key events as values, filtered to sources matching the same expression. That gives you one row per host without cluttering your channel group with ten entries.

If you would rather have it in standard reports, create separate channels per assistant inside the custom group. It is more configuration to maintain and it survives staff turnover better, which is a real consideration.

The assistants you will not see

Claude and Meta AI send outbound links with the referring page removed. Visits from them arrive with no source and land in Direct. No expression catches them, because there is nothing to match against.

This is not a small caveat. If Claude is a meaningful source of traffic for your category, your AI channel will show it as approximately zero while the traffic is real and arriving. The measurement gap is largest for exactly the assistants that most respect user privacy, which is an uncomfortable but honest thing to know.

The only signals that survive are indirect. A campaign tag on a link you control survives referrer stripping, because it lives in the URL rather than in a header. And crawler activity in your server logs shows which assistants are reading your pages, whether or not their readers arrive with a referrer.

A sensible reporting cadence

Weekly is too often for most sites. AI referral volumes are small enough that week-to-week movement is mostly noise, and reading noise as signal produces confident wrong decisions.

Monthly, with landing pages broken out, is usually the right rhythm. Look at whether the set of pages being cited has changed, not just whether the total moved. A shift in which pages arrive is a real signal about how assistants see your site; a ten per cent move in a small total is not.

Whatever cadence you pick, write down the definition you used — which hosts, whether bing.com is included, which channel group. In six months nobody will remember, and an undocumented definition is how two people end up reporting different numbers for the same month.

Keeping the definition alive as the list changes

New assistants appear and existing ones change hostnames. A regular expression written once and never revisited slowly stops matching reality, and the failure is silent — the channel keeps reporting, just less of the truth each quarter.

Set a reminder to re-check the host list quarterly. The check is short: look at your unattributed referral sources for anything that looks like an assistant, and look at the crawler user agents in your logs, which usually show a new entrant reading you before its readers start arriving.

When you add a host, note the date beside the expression. A step change in the series then has an explanation attached rather than becoming a mystery that somebody attributes to their own work six months later.

And keep the expression in one place. If the same list is duplicated across a channel group, three explorations and a spreadsheet, they will drift apart, and the version somebody quotes in a meeting will be whichever one they opened last.

Reading the result honestly

Once the channel works, the most likely outcome is that the numbers are smaller than expected. AI referral volumes on most sites are modest, and the gap between the discourse and the data is wide.

That is worth absorbing before making plans. A channel sending a few hundred sessions a month deserves proportionate effort, and the correct response to a small number is often to stop investing rather than to buy a tool that measures it more precisely.

The exception is conversion quality. A small channel that converts several times better than search is worth more attention than its volume suggests, and that is a comparison you can make today from data you already have.