The AI traffic GA4 cannot see

For about two years the only way to see AI traffic in Google Analytics was to build the channel yourself: a custom channel group, a regex full of hostnames, and a note to update it every time somebody shipped a new assistant. Half the industry had a slightly different regex, which meant no two sites were counting the same thing.

That ended in May 2026. GA4 now carries an AI Assistant channel in the Default Channel Group, with no configuration required. The rule is public and it is mechanical: when the referrer matches a list of AI assistants, medium is set to ai-assistant, campaign is set to (ai-assistant), and the session lands in a channel of its own instead of being buried in Referral.

This is a real improvement and it deserves the credit. It is also the moment to be precise about what that row counts, because the number in it is going to be small, and the reason it is small has almost nothing to do with how often assistants name you. Three separate things have to go right before a session reaches that channel, and the surface that changed search the most is excluded from it by definition.

Every figure below was checked on September 4, 2026. All of them are somebody else’s published data rather than ours, with the sample size and the window attached, so you can go and disagree with the methodology directly.

Start with the honest size of the thing. SE Ranking looked at 101,574 websites across 250 countries and territories, using aggregated Google Analytics data from January 2025 through May 2026. In May 2026, ChatGPT accounted for 0.32% of all referral traffic worldwide.

Not 0.32% of AI traffic. 0.32% of referral traffic, which is itself a slice of the total. If you were expecting the channel row to explain a quarter of anything, it will not.

The growth is the part worth taking seriously. That 0.32% was 0.23% in April, a 36.7% jump in a single month, and it beat the previous peak of 0.29% set back in October 2025. The EU grew 42.7% month over month, the UK 38.7%, the US 23.2%. For the four months before that, worldwide growth was about 1.5% in total, with the UK and the EU actually declining. So this is not a smooth ramp you can extrapolate. It is flat for a while and then it moves.

Both readings are true at once, and picking one of them is how people end up wrong. A rounding error growing 36.7% in a month is worth a channel row and worth watching. It is not worth a quarter of your roadmap on the strength of the sessions alone, and anyone selling you that quarter is quoting the growth rate without the base.

Yes, when three things are all true at once. Somebody clicked. The referrer header survived the trip. And the source is on Google’s list. Miss any one of the three and the visit still arrives, it just arrives wearing someone else’s label. Here are the three holes, in the order they cost you the most.

AI Overviews and AI Mode are not in it

Google’s own definition of the channel says so out loud: the AI Assistant group excludes AI Overviews and AI Mode. A click out of an AI Overview carries a google.com referrer, so it lands in Organic Search, sitting in the same bucket as a blue link at position three and indistinguishable from one.

That is the exclusion that matters, because AI Overviews are the surface with the volume. They now appear on more than 20% of queries. SparkToro, working from Similarweb clickstream data, put zero-click US Google searches at 68.01% for January through April 2026, up from 60.45% in 2024. AI Mode was only 0.34% of searches in that window, so almost none of that shift is AI Mode yet.

The behavioural detail is sharper still. Pew Research tracked 68,879 searches from 900 consenting US adults in March 2025. When an AI summary appeared, users clicked a search result on 8% of visits. Without one, 15%, nearly twice as often. And the links inside the summary itself, the citations everyone is optimising for, were clicked on 1% of visits.

Put those together and the shape of the problem is clear. The surface that changed the most produces the fewest clicks, and every click it does produce is filed under Organic Search. Your AI Assistant row can be flat for a year while AI Overviews quietly rewrite your organic numbers underneath it.

Claude and Perplexity are not on the list

The live definition of the channel names ChatGPT, Gemini, Deepseek, Copilot and Grok. Claude appeared in Google’s launch note and is not in the list as it currently reads, and Perplexity has never been in it. Traffic from either one keeps landing in Referral, exactly where it was before the channel existed.

This is not a complaint, it is a maintenance job. The list is Google’s to change and it will change, silently, without a migration note. Read the current definition rather than trusting a blog post about it, including this one, and keep a custom channel group alongside the native one for the assistants the default has not picked up yet.

Worth noting what the mismatch does to your history. GA4 processes channel assignment daily and the change was not retroactive, so everything before May 2026 stays classified the way it was classified then. A year-over-year comparison on this channel is comparing a definition to its own absence.

The app sent no referrer at all

The channel is built on the referrer header. A large share of assistant usage happens in native mobile and desktop apps, and a fair amount of the rest happens by someone copying a URL out of an answer and pasting it into a browser. Neither produces a referrer. Both land in Direct.

ChatGPT patches part of this by appending a utm_source=chatgpt.com tag to links it cites while browsing the live web, which is why some of your ChatGPT clicks have arrived correctly labelled for a while now. It is applied inconsistently in the native apps, and a page the model names from memory rather than from a live fetch usually is not a link at all. No other assistant tags reliably.

So Direct has been carrying a share of your AI traffic since the beginning, it still is, and there is no clean way to pull it back out. A rising Direct number on pages that assistants cite is a symptom worth watching. It is not evidence, and treating it as evidence is how people talk themselves into a result.

Here is the part that reframes the whole report. Cloudflare publishes a metric called the crawl-to-refer ratio: how many pages a platform’s crawler fetches for every one visitor that platform sends back. When they introduced it in July 2025, Anthropic’s ratio for the week of June 19 to 26 was 70,900 to 1.

Those numbers have improved enormously. Cloudflare Radar keeps the metric on a rolling 28-day window, and the window ending July 21, 2026 read roughly: Mistral 3,389 to 1, Anthropic 2,237 to 1, Perplexity 225 to 1, OpenAI 217 to 1, Microsoft 35 to 1, Google 4.6 to 1, DuckDuckGo 2.5 to 1.

Read why they fell, because it is not the obvious reason. The crawling did not slow down. Referrals appeared in the denominator for the first time as ChatGPT and Claude shipped consumer surfaces that link out. Anthropic going from five figures to four in six months is a product change, not a politeness change.

What the ratio tells you about your analytics is this. For every session in that AI Assistant row, hundreds or thousands of fetches happened that produced no session at all. Those fetches are the common event. The click is the rare one. And the fetch never touches GA4, because nothing there executes your JavaScript, which means the most frequent interaction between an assistant and your site is invisible to the tool you are using to measure assistants.

Your server logs do see it. That is the same place the answer to a different question lives, and Nobody is reading your llms.txt goes through what a month of them actually records, including how to tell a retrieval bot from an SEO audit tool wearing a similar name.

This is the counter-argument, and it is a decent one. Semrush published a study in June 2025 finding the average AI search visitor 4.4 times as valuable as the average organic search visitor, measured by conversion rate. The mechanism is believable without any hand-waving: somebody arriving from an answer has already had the basic question resolved and has probably already seen the comparison. They are further down the decision than the person who typed a keyword.

Now the caveats, in the order that matters. Semrush sells AI visibility tracking, which does not make the number wrong and does mean the incentive points one way; their plans are priced out against ours in Beseen vs Semrush. We sell the same category, so read this paragraph with the same suspicion.

The bigger problem is the denominator. At 0.32% of referral traffic, most sites are computing a conversion rate on a few dozen sessions a month. A conversion rate on forty sessions is a story, not a result, and it will swing wildly month to month for reasons that have nothing to do with anything you did.

There is a selection effect underneath it too, pointing the same direction. A click out of an assistant happens either because the answer did not finish the job or because the person is ready to act. Both of those select for intent. The 4.4x is real and it is partly measuring who bothers to click, not what the channel is worth in total.

One more finding from the SE Ranking data makes the sessions less useful than they look: 60% of AI-referred traffic lands on homepages, against 17% from organic search. The visit usually cannot tell you which page won the answer, because the answer often did not send anyone to a page. It sent them to your front door with your name already in their head.

None of the above is a reason to skip the setup. A channel that counts a small thing correctly is worth far more than one that counts an unknown mixture. An hour of work, in order.

• Keep the native AI Assistant channel and add a custom channel group beside it, with a regex covering the sources the default misses: claude.ai, perplexity.ai, and whatever else shows up in your Referral report with an assistant's hostname. Order matters inside a channel group: the first matching rule wins, so it has to sit above Referral.
• Read Google's current definition of the default group rather than a summary of it, and write down the date you read it. The list of recognised assistants changes without an announcement, and a channel definition you cannot date is a number you cannot compare.
• Segment the AI Assistant channel by landing page. If most of it is the homepage, you have confirmed the pattern above, and your session data cannot tell you which page earned the mention. That is a limit to work around, not a reporting bug.
• Do not try to force AI Overview clicks into this channel. They are organic and they should stay organic. Watch them in Search Console instead: impressions holding while clicks fall, on the queries where an overview now appears, is the shape of that loss.
• Grep 30 days of server logs for AI user agents and count the fetches. Compare that count to the sessions in the channel. The gap between the two is your own crawl-to-refer ratio, and it is the only version of that number that is about your site.
• Watch Direct on the pages assistants cite, weekly, as a symptom rather than a metric. Never put it in a report as AI traffic. You cannot separate it from a bookmark and you should not pretend otherwise.

Finish that and you will have a clean, small, correctly labelled number that is genuinely comparable week to week. Which is exactly what it is for, and it still will not answer the question you actually have.

Every hole above comes from the same source: analytics measure visits, and the thing you need to know is whether you were named. Those two come apart because the assistant answers the question itself. Being named without a click is the normal outcome, not the failure case, and it is the outcome no session-based tool can see by construction.

The measurement that does not depend on any of this is the prompt. Twenty to thirty of them, worded the way a buyer asks rather than the way a keyword tool phrases things, run against ChatGPT, Claude, Perplexity and Gemini, recording which brands get named and which URLs get cited. No referrer header involved, no channel definition to age, nothing for a native app to strip.

The rules for doing that well are boring and load-bearing. Freeze the wording, because a reworded prompt is a different question and you will read the difference as a result. Repeat weekly, because one run is an anecdote. Change one thing at a time on the pages you touch. Ranking is not being named covers how to pick the prompts, and why a page can rank at position one and still never appear in the sentence.

Keep the GA4 work anyway. Sessions and citations answer different questions, and the useful report has both: what the assistants say about you, and what the small number of people who click go on to do.

The prompt half is the part with no free version. Running one prompt against one assistant and recording who gets named is cheap in data fees and expensive in discipline: the same wording, the same schedule, week after week, with the results written down somewhere you can diff them. Spreadsheets lose that fight around week three.

That loop is what we built. How it works covers what gets pulled, how often, and what a rewrite looks like when it lands. If you would rather see your own prompt coverage than another study about somebody else’s, start the free trial: 14 days free on Starter, then $99 a month. The first check is the whole diagnosis: every question you lose, who is named on it instead of you, and the wording of the answer that did it. That is the reading your analytics were never going to give you.