Create a custom channel group with an "AI Assistants" channel at the top, defined by a regular expression on the source.
The native "AI Assistant" channel, launched in May 2026, misses some assistants and has no history. In your own channel, read revenue rather than sessions.
That revenue is a floor, and the channel will probably stay much smaller, in sessions, than the "Unassigned" row in the same report.
The benchmark
The order of magnitude comes from a study published in Marketing Science, covering 973 e-commerce sites between August 2024 and July 2025. Assistants send them less than 0.2% of visits, and ChatGPT accounts for more than 90% of that.
Set that figure next to twelve months of GA4 data from a store the authors work with.
The "Unassigned" row made up nearly 6% of sessions there, and bots, by our estimate, 5 to 7%, for less than 0.5% of revenue.
The comparison is ours and rests on just one study and one case: the bucket of misfiled sessions is about thirty times bigger than the channel you're trying to measure.
That's why we read revenue, which this noise barely touches.
The rule
Read this channel as a trend, and wait for 30 transactions before comparing its conversion rate with that of other channels.
In a made-up example, a store with 50,000 monthly sessions gets about a hundred visits from assistants, or one or two orders a month at a 1.5% conversion rate.
Three more orders triple that rate, and reaching 30 transactions takes twenty months. So no budget should be decided on a first figure. We set ourselves the same rule for our own media, with a revenue check once a month.
The steps
- The native channel. In the Traffic acquisition report, the "AI Assistant" row follows a list of referrers for which the documentation gives only five examples: ChatGPT, Gemini, Deepseek, Copilot and Grok. Search Engine Journal noted in July 2026 that Perplexity was still in Referral. Note the date this row appears in your property, because it's zero before then, which skews any year-over-year comparison.
- The group. Open Admin > Data display > Channel groups, then click Create new channel group. The group applies retroactively, which gives you the history the native channel lacks, and you can then select it as the primary dimension of the same report.
- The channel. Add an "AI Assistants" channel with the condition "Source matches regex," then move it to the first position. GA4 keeps the first channel whose definition matches. If yours sits below Referral, it stays almost empty, and you can spend a long time hunting for the error in the formula.
^.*(chatgpt|openai|perplexity|gemini\.google|copilot\.microsoft|m365\.cloud\.microsoft|edgeservices\.bing|claude\.ai|mistral|deepseek|grok|x\.ai|you\.com|meta\.ai|poe\.com|phind|iask\.ai|andisearch|huggingface|writesonic).*$- The test. Before saving, run the formula against your real sources, exported from this report. A GA4 regular expression must match the entire value, hence the
.*at both ends. Because it works on the source alone, it also catches visitschatgpt.com / (not set), which are filed under "Unassigned" by default. - The check. Compare GA4 purchases with Shopify orders, all channels combined, using the calculation in the consent guide. This spring, an e-commerce manager at a brand told us: "we have a conversion rate of 0.1, when our conversion rate, normally, is 0.6, 0.7." According to her, half of her sales went through an express payment that didn't send the purchase back to GA4. My rule, which is purely personal, tolerates a 15% gap for reading a trend. Beyond that, the revenue in your AI channel is understated too.
- The blind spots. Assistants' mobile apps rarely pass along their origin, so watch for Direct traffic landing on deep product pages. On Google's side, clicks from AI Overviews and AI Mode stay in Organic Search.
The trap
The false positive comes from the broad version, which matches any source containing one of its words. In two constructed examples, "grok" catches ngrok, your developers' tunneling tool, and a newsletter tagged utm_source=mistral for a collection of the same name inflates the channel from day one. Nobody notices, since the number looks plausible.
If the test turns up intruders like these, use the strict version, limited to full hostnames, and complete it with your real sources:
^(chatgpt\.com|chat\.openai\.com|openai\.com|perplexity\.ai|www\.perplexity\.ai|gemini\.google\.com|copilot\.microsoft\.com|m365\.cloud\.microsoft|edgeservices\.bing\.com|claude\.ai|chat\.mistral\.ai|chat\.deepseek\.com|grok\.com|x\.ai|you\.com|meta\.ai|poe\.com|phind\.com|iask\.ai|andisearch\.com)$The sentence to present
Copy this sentence, then rewrite it with your own numbers before presenting it to your leadership.
"AI assistants brought us at least €[X] this month, from [N] orders. That's a floor: GA4 only sees [Y]% of our Shopify orders, and Google's AI answers are counted as organic search. I propose we track this number every month and decide nothing until we have 30 orders."
