Obility Editorial · · 4 min read
GA4 labels an AI visit by matching its referrer
On 13 May 2026 Google Analytics added an AI Assistant channel to its default channel group. Google’s release note says that when the referrer of a visit matches a recognized AI assistant, GA4 sets the medium to “ai-assistant”, sets the campaign to “(ai-assistant)”, and files the session under the new channel. Before that date, most of these visits sat in Referral alongside every other site that linked to you.
Google does not publish the list it matches against. The release note names ChatGPT, Gemini and Claude. The channel definition page in Analytics Help names ChatGPT, Gemini, Deepseek, Copilot and Grok, and says the channel excludes Google’s own AI Overviews and AI Mode. Neither page mentions Perplexity, and the downloadable spreadsheet of source categories that drives the other default channels has no AI category at all. So the honest starting point is to look at your own data: open Traffic acquisition, set the dimension to Session source / medium, and see which assistants actually appear with “ai-assistant” as their medium.
A visit with no referrer cannot be matched
The rule depends on the referrer, and plenty of AI visits arrive without one. A link copied out of an answer and pasted into a new tab carries no referrer. Neither does a link opened by an app that does not pass one along. Google’s documentation on traffic sources says that when no information about the referral source is available, the session is processed as direct traffic. Those visits land in Direct, with nothing to show where they came from.
ChatGPT is the one assistant with a documented second trace. OpenAI’s publisher FAQ states that ChatGPT adds utm_source=chatgpt.com to the links it sends from search results. A URL parameter survives when a referrer does not. But Google’s documentation also says campaign parameters in the page URL map directly to the source and medium dimensions, ChatGPT supplies only a source, and Google’s channel pages do not say what medium such a visit receives when its referrer is missing.
You do not need to guess. In an exploration, filter Landing page + query string for utm_source=chatgpt.com and break the sessions down by Session default channel group. If some of them sit in Unassigned, Referral or Direct instead of AI Assistant, the channel is undercounting ChatGPT on your property, and the gap between the two numbers is the size of the problem.
Google’s sample regex counts the wrong things
Before the default channel existed, the usual fix was a custom channel group, and Analytics Help still offers one for AI assistants as a worked example. Its regular expression begins with an alternative that, under GA4’s default full match, accepts any source ending in “ai”. That includes perplexity.ai and claude.ai, which is the intent, but also every partner blog, directory or review site on a .ai domain, which is not. The same expression would miss chat.deepseek.com and grok.com, two sources Google’s own channel definition names.
A safer pattern lists exact hostnames you have seen in your own source report, with the dots escaped, for example chatgpt\.com, (www\.)?perplexity\.ai, claude\.ai, gemini\.google\.com and copilot\.microsoft\.com, joined with a vertical bar. GA4 places traffic in the first channel whose definition it matches, so the AI channel has to sit above Referral in the group order. Review the list each month, because new assistants appear and hostnames change.
The trend line breaks on 13 May unless you rebuild it
Google’s release note does not say that sessions recorded before 13 May were reclassified. If your report compares channels across that date, Referral will appear to drop and AI Assistant will appear from nothing. That is a labelling change, not a shift in buyer behaviour, and it is easy to misread in a quarterly review.
Analytics Help states that custom channel groups can be applied to reports retroactively. A custom group built on the exact hostname list above gives you one consistent series for AI referrals on both sides of the date, which the default channel cannot. Google’s 11 June release note adds a second option: a Source Group dimension that groups sources such as ChatGPT and Perplexity and is populated retroactively. Where it appears in your reports, it is a useful check on your own hostname list. Use one consistent series for trends, and treat the default channel as a cross check.
A session is the last and smallest trace of an AI answer
Even a perfectly configured channel only counts people who clicked. Google’s definition puts AI Overviews and AI Mode clicks into Organic Search, where GA4 cannot separate them from ordinary results. An answer that names your company and the buyer acts on later, by searching your brand or typing your address, produces no AI session at all.
That is why Obility’s Agent Analytics keeps referral activity, the landing pages and actions of visitors arriving from answer engines, separate from crawler requests and from observations of what the answers actually said. Read the GA4 channel as a floor on AI referral traffic, check which landing pages it sends people to, and compare those pages with the ones assistants cite when they answer your buyers’ questions. Do not read a flat or falling AI channel as proof that AI answers ignore you.