When someone visits your site through ChatGPT, you often will not see it in Google Analytics. The visitor arrives as "direct", or worse, as "organic" while Google had nothing to do with it. So you have no idea that AI assistants send you traffic, let alone how much and from which source.
Most of that is fixable. GA4 does part of the work itself since this spring, and you cover the rest with UTM parameters and a custom channel group. Here is how, and where the measurement still breaks.
Your prospect asks ChatGPT, Perplexity or Gemini a question. The answer cites sources, often with a link to your site. The visitor clicks through. Fine, right?
Except you do not always see that click as a referral. If no referrer is passed, your analytics assumes someone typed your domain directly or came from a bookmark. You structurally underestimate what AI does for you.
Sound familiar? Every week you see dozens of visits land as "direct" with no idea where they came from.
We see at Dutch B2B companies that part of that "direct" traffic actually comes from AI answers. Generative assistants increasingly show citations with sources. That creates a kind of referral traffic that did not exist a few years ago.
A normal website passes a referrer on click. The browser tells GA4: "this visitor comes from LinkedIn" or "from Google". GA4 then sorts the traffic into the right bucket.
On 13 May 2026 Google added a channel for this. When GA4 recognises the referrer as an AI assistant, the session automatically gets the medium ai-assistant, the campaign (ai-assistant) and lands in the AI Assistant channel of the default channel group. You do not have to configure anything. Google documents the channel in its default channel group reference and names assistants such as ChatGPT, Gemini and Claude.
Good news, but it only solves half your problem. The channel works only when a referrer is passed. If someone clicks through from an assistant that sends no referrer, that session still sits under Direct. Google also does not publish a full list of recognised assistants, so you cannot be sure which tools are covered.
It gets more interesting when someone copies an answer and shares the link in Slack or WhatsApp. That is dark social: traffic through shared links without a referrer. It looks like direct navigation, but it is not.
Before you measure, you need to know what you are measuring. This is the order we use:
chatgpt.com, perplexity.ai and gemini.google.com. Place that rule before the standard rules for Organic Search and Referral.This is not a one-off job. New AI tools appear regularly, and the list Google recognises changes with them. Set a quarterly reminder.
UTM parameters get named as the fix for AI traffic. That is only half true, and the difference matters.
A UTM only works on a link you control yourself. You cannot force an assistant to cite a tagged URL; it takes the URL as it appears on your site. So for citations in AI answers, UTMs are not a measurement instrument.
Where they do work: every link you distribute yourself that might get forwarded later. Think of links in your newsletter, in a proposal, in a LinkedIn message or in documentation. That is exactly the traffic that otherwise disappears into dark social. Consistent tagging looks like this:
A complete URL then looks like this:
https://yoursite.com/blog/geo-for-it-companies?utm_source=newsletter&utm_medium=email&utm_campaign=geo
For actual AI citation traffic you lean on the AI Assistant channel, your own channel group and landing page analysis of your direct traffic. Not on UTMs.
The default channel group now catches part of your AI traffic. The rest still disappears into Direct, or sometimes into Organic Search. Then you report ChatGPT traffic as if it were Google. You do not want to explain that to your board.
The fix is a custom channel group alongside the default one. It catches the assistants Google does not recognise, and it keeps your own definition stable if Google changes theirs. Then build a report with the dimensions that matter:
What stands out with clients who set this up properly: AI traffic often shows higher engagement than average direct traffic. Visitors from an assistant have a concrete question, click through with intent and read longer. They are not there for a free whitepaper, they are solving a specific problem.
Four mistakes that make your measurement worthless.
Using the unwanted referrals list. This is the most common one. That GA4 setting does the exact opposite of what you want: it stops a domain from being counted as a referral, which pushes the traffic into Direct. Put AI domains in there and you make your own blind spot bigger. That list is meant for payment providers and login screens on your own subdomain.
Reading AI traffic as organic search. You see a rise in GA4 and assume your SEO is working. In reality an assistant is sending the traffic. The result is that you invest in the wrong channel and your GEO results never become visible.
Ignoring traffic quality. AI visitors convert differently from Google visitors. They are further along in their orientation, but they also compare more actively. Look only at conversions without splitting by source and you draw the wrong conclusions.
Skipping the baseline. You can only improve what you have measured. Most companies start measuring after three months of GEO work. By then you still cannot tell whether it worked.
Your board does not care about UTM parameters or channel groups. They want to know whether AI traffic delivers. So your reporting starts with the question you answer, not with the technique.
Build your reporting around three numbers: reach (how many citations and visitors), engagement (what they do on the site) and conversion (what it delivers). Put those three side by side and you have a story.
And be honest about the limits. Guaranteed positions in AI answers do not exist. A spot in an answer is not for sale. What you can report is whether your visibility grows, whether that traffic increases and whether it converts. Those are the numbers you can measure honestly every month.
Want to know where your site stands in AI answers? We measure in ChatGPT, Gemini, Claude and Perplexity and report with source URLs. Start with a free AI visibility scan through our GEO scanner and you will know within a day.
Start by checking whether the AI Assistant channel in your GA4 acquisition report is filling up. Since May 2026 Google files those sessions automatically when it recognises the referrer, with medium ai-assistant and source chatgpt.com. Clicks that arrive without a referrer stay under Direct; you only separate those with UTM parameters on links you distribute yourself, or by analysing your direct traffic per landing page.
AI traffic comes from visitors clicking through from an AI answer. Dark social is traffic from shared links without a referrer, for example through WhatsApp or email. Both look like direct traffic in analytics, but the source is different. Without UTM parameters you cannot tell them apart with certainty.
Partly. Since 13 May 2026 GA4 has an AI Assistant channel in the default channel group that separates traffic from recognised assistants without any configuration. That only works when the click sends a referrer. For everything else, and for assistants Google does not recognise yet, you need your own custom channel group or UTM tagging.
Measuring AI traffic is not rocket science. It takes an hour in GA4 and an appointment with yourself to keep it up. The payoff is that you see which assistants find you, on which questions, and whether those visitors do anything. That is not a vanity metric. That is marketing you can act on.