Since 13 May 2026, GA4 has had its own AI Assistant channel in the Default Channel Group. No configuration needed, no regex tinkering: sessions from ChatGPT, Gemini and Claude come in with medium ai-assistant and land in their own row in your reports. That is progress. Before that, this traffic disappeared into the Referral catch-all or, worse, into Direct.

But if you think measuring AI traffic is now solved, you are using a yardstick that is structurally too short. The channel only registers visits that send along a recognisable referrer. And a sizeable share of AI traffic does not.

What GA4 does show since May 2026

The new channel does exactly what you would expect. You see how many sessions come in from AI assistants, which landing pages they hit and what they do next. You can put the channel alongside Organic Search and Direct and compare month by month. Semrush described the launch with three named sources: ChatGPT, Gemini and Claude.

And that is about it. Perplexity falls outside the standard definition. So do smaller assistants and AI browsers. And every session without a referrer stays invisible, however clear the pattern may otherwise be.

What you miss: a third to two thirds

The measured estimates vary, but they all point in the same direction. Veza Digital puts the share of AI traffic arriving without a referrer at 35 to 71 percent. Clickport analysed 371,847 sessions across client sites and arrived at 35.7 percent without a referrer in April 2026. In that same dataset, ChatGPT accounted for 76 percent of all AI sessions.

Across 371,847 measured sessions, 35.7 percent of AI traffic arrived without a referrer. In GA4, that traffic sits under 'Direct'.

Four mechanisms strip that referrer. The assistant opens the link in an in-app browser or desktop client that does not send a referrer. The user copies the URL and pastes it into a new tab. An intermediate redirect throws the referrer away. And, the most painful one: your own site's referrer policy. A strict Referrer-Policy or a CDN setting can wipe out the very origin data you are trying to measure. That is the first thing to check before you draw any conclusions about your AI channel.

Small volume, high value

The absolute numbers are usually disappointing. Conductor benchmarks put AI referrals at roughly 1.08 percent of all web traffic. That is little, certainly next to organic search.

Quality is another story. Semrush calls AI visitors roughly 4.4 times more valuable than average organic traffic. SimilarWeb measured an e-commerce conversion rate of 11.4 percent on AI referrals against 5.3 percent on organic search. That difference makes sense: someone who reaches your site through an assistant has already worked through most of the orientation phase. The assistant has already drawn up the shortlist.

Which makes it painful that a third of those visitors disappear under Direct. You underestimate your best converting channel, and you attribute the conversions to "people who already knew us anyway".

Measuring AI traffic is not the same as measuring visibility

This is the fundamental point. Even a perfect channel would give you only a fraction of the picture, because most AI interactions never lead to a click at all.

Cloudflare Radar shows how lopsided the ratio between crawling and referring is. OpenAI fetches around 217 pages per referral. Anthropic 2,237. Mistral 3,389. Google sits at 4.6. In other words: for every visitor an AI assistant sends to your site, hundreds to thousands of pages are read without anyone clicking through.

Those pages read are not wasted. They determine whether your brand gets mentioned when someone asks ChatGPT which agency, which supplier or which product to consider. The answer appears in the chat, the user takes it away, and your analytics sees nothing. If you want to know whether you show up there, you have to measure AI visibility in the answers themselves, not in your session logs. A first check along the lines of is my company in ChatGPT often says more than a month of channel data.

The playing field shifts faster than your reporting cycle

Measuring once is pointless. In one GA4 panel, ChatGPT dropped to 63 percent of AI traffic, while Claude went from 1.4 to 18.5 percent in the same period and Gemini quadrupled. Anyone basing their GEO strategy on a measurement from six months ago is optimising for a distribution that no longer exists.

The large panels are honest about this too. Trakkr reports movements across thousands of GA4 properties, but with explicit coverage limitations: only visible referrers count. Nobody has the full picture. That is no reason not to measure, but it is a reason to phrase your conclusions carefully.

What to do in practice

A workable approach in five steps:

  1. Check your own referrer policy. A Referrer-Policy of strict-origin-when-cross-origin usually works well. no-referrer wipes your data.
  2. Build a custom channel group alongside the default one, where you manually add Perplexity, Copilot, AI browsers and new assistants. Google is by definition behind the market.
  3. Segment Direct by landing page. Direct sessions landing on a deep, specific URL with no brand history are almost certainly AI traffic. Put that segment next to your ai-assistant channel and see whether it moves in step.
  4. Read your server logs. There you do see which crawlers come by and how often. That is the only place where the crawl-to-refer ratio becomes visible.
  5. Measure visibility separately from clicks. What share of the relevant prompts mentions your brand, and in what context. See our measurement method for how we record that.

The summary

The AI Assistant channel in GA4 is a useful lower bound, not a complete measurement. Count on missing at least a third of your AI traffic, on the traffic you do see converting better than your organic traffic, and on clicks being only a fraction of the times an assistant reads your content anyway.

Use GA4 for the trend and to underpin the business case. Use a direct visibility measurement to know where you stand and what there is to gain. Combine both and you can improve AI visibility based on demonstrated positions instead of assumed effect.