Why GA4 shows almost no ChatGPT traffic when buyers name it on calls

The complaint arrives in a consistent shape. Sales are hearing the name on calls, sometimes in the first two minutes, and the buyer describes a shortlist an assistant handed them. Marketing opens GA4, filters for anything resembling ChatGPT, and finds a number small enough to round to zero. Somebody then concludes the tracking is broken, and spends a quarter trying to repair a measurement that is working correctly.

Both numbers are right. They count different events. GA4 counts clicks that reached your server carrying an identifiable source, and ChatGPT referrals were 0.32 percent of traffic across 101,574 websites in May 2026, the highest share in seventeen months of tracking. 1 Sales are hearing about a recommendation, and 51 percent of B2B software buyers now start research in an AI chatbot. 2

The influence is real and it is large. It mostly does not arrive as a click. Once that is accepted the work stops being an attribution repair job and becomes a measurement design problem, which is a different and far more tractable thing.

Setting the two sides beside each other is worth doing once, because the ratio between them is what makes people assume a bug. The first rows are what buyers report doing. The last row is what analytics records. Neither instrument is faulty.

A useful way to hold this is that the assistant is doing the job a category listicle and three peer conversations used to do. None of those ever appeared in GA4 either. What changed is that the whole of that stage now happens inside one product your analytics can see the edge of, which makes the absence feel like a failure rather than a normal condition.

This is the most repeated explanation and, for the web app, it is wrong. We checked rather than assumed. On 4 September 2026 we put 50 B2B SaaS buying questions through ChatGPT in a desktop browser, one software category each, read the link markup out of every rendered answer, then clicked through and read what the destination received.

Two things follow. The referrer is sent, trimmed to the origin so the conversation URL never leaks, which is enough for any analytics package to attribute the session. And OpenAI tags every link with a campaign parameter, so even a visitor arriving with no referrer at all stays identifiable from the query string, as long as your site does not discard it on redirect. 4

One more thing is visible while the data is open, and it needs stating in the right unit. 215 of the 271 links, or 79 percent, pointed at a domain belonging to a product the answer was recommending, and 28 of the 50 answers linked to nothing else. That counts anchors rendered inside an answer, which is not the same measurement as counting citations across a corpus, so it does not compare with the published studies of what engines cite. Read narrowly it says one useful thing. A recommended vendor is not short of links. The gap opens after the link already exists. 4

Referrer loss is a genuine problem in the places we did not test. Links opened inside the mobile apps and inside Atlas arrive with the referrer removed, which drops the session into Direct or Unassigned rather than an AI channel. 5 So the honest version of the claim is narrower than the one in circulation. Some ChatGPT traffic is invisible, the web app's is not, and neither fact explains a number as small as 0.32 percent.

Into two places a referral report was never going to see. The first is that the click does not happen. A confidential OpenAI report shown to a publisher put the best-performing URL at 185,000 conversation impressions against 3,800 click events, a 2 percent rate, with 0.1 percent described as typical across the wider set. 6

Ahrefs measured the same effect from the receiving end and put it plainly: visitors from AI search click links 75 percent less than they do in traditional organic search. 7 An agency tracking its own citations found its most-retrieved page appeared in AI answers 210 times and drew 10 referral sessions. 8

The second place is your own branded search line. When an assistant names a vendor, clicking the citation is not the main thing buyers do next. 3

• 71 percent go to the vendor's website, by whatever route. 3
• 63 percent search Google for the company by name, which arrives in GA4 as organic search. 3
• 46 percent compare it against alternatives. 3
• 38 percent check the G2 profile. 3
• 14 percent ask a colleague, which leaves no trace anywhere. 3

That second line is the whole gap in one number. A buyer who reads your name in an answer and then types it into Google is recorded as branded organic, and the assistant that created the demand is credited with nothing. The traffic is not missing. It is filed under the wrong cause.

Partly, and it is worth turning on first because it costs nothing. Google added an AI Assistant channel to the default channel group in May 2026. When the referrer matches a recognised assistant, GA4 sets the medium to ai-assistant and the campaign to (ai-assistant) with no configuration on your side. The documentation names five: ChatGPT, Gemini, Deepseek, Copilot and Grok. 9

It also states what it excludes, and the exclusion matters more than the feature. Google's own AI surfaces are deliberately kept out, so AI Overviews and AI Mode continue to count as organic search. 9 The channel applies going forward only, so nothing from before it launched is reclassified.

Four of those five rows are the same buyer having the same experience, and one of them is counted as an assistant. A custom channel group built on source rather than medium collapses the first four back into a single line, which is worth an afternoon. It does nothing for the fifth, and the fifth is the largest. 5

Asking the buyer. It is unglamorous and it is the only instrument positioned to see what your logs cannot. One agency ran the comparison on its own pipeline between 6 July and 5 August 2026. Of 213 leads who answered a free-text question about how they heard of the company, 189 named a generative engine. First-touch attribution credited 28 of those to AI. 8

The other 161 were filed as organic search, direct, or referral. Every one of those labels is defensible on the evidence the tool had. None of them is what happened. That ratio answers the question this page is about: the channel report was not undercounting by a little, it was seeing roughly one lead in seven.

1. 01Put a required free-text question on every demo request and contact form asking how the buyer first heard of you, and read the answers rather than counting them. This is the highest-yield single change, and in the one published comparison it recovered close to seven times what first-touch caught. 8
2. 02Ask the same question on discovery calls and write the answer into a CRM field rather than into call notes, so it can be counted later. Sales already hold this information and are already saying it out loud.
3. 03Build a custom GA4 channel group keyed on source rather than medium, so ChatGPT sessions fragmented across AI Assistant, Referral, Unassigned and Direct collapse into one line. 5
4. 04Confirm your site keeps the utm_source parameter through every redirect. A canonical redirect that drops the query string throws away the one identifier that survives when the referrer does not. 4
5. 05Log the four OpenAI agents server-side and verify them against OpenAI's published address files rather than trusting the user-agent string. OAI-SearchBot builds the search index, ChatGPT-User fetches a page live when somebody asks, GPTBot is the model crawler, and OAI-AdsBot checks pages submitted as ads. 10
6. 06Watch branded search volume in Search Console as the second-order signal. If assistants are recommending you, demand surfaces there before it surfaces anywhere you can attribute. 3
7. 07Run a fixed prompt set against a baseline taken before the work starts, and read the sources in the answers rather than the sentiment. A screenshot of one good answer proves nothing, because the same prompt returns different answers on different days.

Self-reported attribution is a survey and carries the biases of surveys. Buyers misremember, they name the tool they used most recently rather than the one that changed their mind, and people who fill in a free-text box are not a random sample of the people who visited. Treat the number as a corrected floor rather than a true value.

Server logs have the opposite problem. A crawler hit proves an engine read the page and proves nothing about whether the page was then used in an answer, and retrieval agents fetch far more than they cite. Neither instrument is sufficient alone. Read together, one bounds what is possible and the other measures what was noticed.