The measurement problem in one sentence
Somewhere between 35% and 70% of AI referral sessions arrive without a referrer header and get filed as Direct traffic, which means the AI channel in your analytics is plausibly showing about half of what is really happening — and 89% of brands cannot attribute this traffic properly at all.
That is an unusual situation. Marketers are used to channels being imperfectly measured; they are not used to a channel where the majority of the evidence is missing and the visible remainder is systematically biased.
What the conversion numbers actually say
The figures circulating in 2026 are striking: ChatGPT referral traffic converting at 15.9%, Perplexity at 10.5%, Claude at 5.0%, against 1.76% for Google organic. A broader claim puts AI search traffic at roughly 4.4 times the conversion rate of traditional organic.
These deserve careful handling, and the sources are unusually honest about why. The 15.9% and 10.5% figures come from a single B2B company's data over one quarter. Anyone quoting 15.9% as a benchmark is quoting one firm's ideal customer profile, not an industry norm, and the multiple varies enormously with who your buyers are.
What is well supported is the direction and the mechanism. Assistants do a large amount of qualifying before the click: the user has already described their situation, had options compared, and received a recommendation. The click is a late-funnel action. Compare that with a Google result, where the click often begins the research. Higher conversion from AI referrals is what you would predict from the behaviour, which is why the finding replicates even where the magnitude does not.
The practical consequence is that this channel is easy to dismiss on a traffic report and hard to dismiss on a revenue report. If you judge it by sessions, it looks like a rounding error. If you judge it by outcomes, it frequently is not.
Which platforms send the traffic
Through 2026 the distribution concentrated on one player and then began to spread. ChatGPT's share of AI referral traffic fell to around 76.85% by April 2026, with Gemini second at roughly 9.0% and Perplexity third at about 7.73%.
Two implications. First, instrument ChatGPT carefully and group the rest — the effort does not divide evenly. Second, the trend is toward dispersion, particularly as assistants get embedded into browsers and operating systems, which is a shift we cover in how AI browsers are changing the way customers find you. Do not hard-code a platform split that will be wrong in six months.
Making it visible in GA4
GA4 does not classify AI assistants as their own channel by default, so they scatter across Referral and Direct. The fix has three parts, and none of them is complete on its own.
1. A custom channel group. In Admin, create a channel group with a rule matching AI referral hostnames — chatgpt.com and chat.openai.com, perplexity.ai, gemini.google.com, claude.ai, copilot.microsoft.com, and the others you see in your own referral list. This captures every session that does carry a referrer, and it is the foundation for everything else. Review the list quarterly; new surfaces appear constantly.
2. An exploration that segments by it. Channel groups apply going forward in standard reports, so build an exploration with landing page, sessions, engagement rate, and your key conversion, segmented to the AI group. The landing page dimension is the valuable one — it tells you which pages assistants are actually sending people to, which is rarely the pages you expected.
3. UTM tags wherever you control the link. Anywhere you place a link that an assistant might surface — documentation, partner listings, profiles — tag it. This is partial by definition but it is the only fully reliable signal in the stack.
The Direct traffic problem, and the honest workaround
None of the above recovers the sessions that arrive with no referrer, and those are the majority. There is no complete technical fix available today, so what follows is inference rather than measurement, and it should be labelled as such in any report you circulate.
The most useful proxy is unusual Direct traffic to deep pages. Direct traffic to your homepage is ordinary — people type the domain or use a bookmark. Direct traffic to a specific article three levels down, from a new user, on a first session, is not something people type from memory. Trend that segment over time. A rise in deep-page Direct that tracks with your citation visibility is the best available evidence that assistants are sending people.
Two supporting signals are worth watching alongside it. Server logs show assistant crawlers fetching your pages, which tells you what is being read even when nothing is attributed. And self-reported attribution — a "how did you hear about us" field on your enquiry form — has become genuinely useful again, because it captures what analytics structurally cannot. A meaningful share of businesses instrumenting AI properly first noticed the channel from that free-text field rather than from a dashboard.
Citations and sessions are different questions
Worth separating clearly, because they get conflated and they need different work. Citation visibility is whether assistants mention and recommend you at all, most of which produces no click. Referral traffic is the smaller subset who click through.
Optimising for the second while ignoring the first badly understates the channel, because the recommendation that never produces a click still influences the purchase. Our method for measuring the first, without buying a tool, is in finding out whether ChatGPT recommends your brand, and the structural work that makes a page citable in the first place is in schema markup for AI search.
What to do with the data once you have it
The landing page report is where the value is. Assistants cite specific pages that answer specific questions, and the pattern is consistent across the sites we have instrumented: comparison pages, pricing explanations, and process or methodology pages get cited far more than homepages or service overviews.
That gives a clear instruction. Find the pages already earning AI referrals, and make more of them — not more pages generally, but more of that specific kind. The pages that get cited share a shape: a direct answer near the top, specific numbers, and a scope stated plainly enough that a model can tell when the page is relevant.
And set a floor on how you evaluate the channel. Judging it on session volume against Google organic will always make it look negligible. Judge it on conversions and on assisted revenue, and instrument it well enough that the comparison is fair. Send us your GA4 property and we will tell you how much of your Direct traffic is not really direct.