Why ChatGPT Traffic Shows Up as Direct in GA4: The GPT-6 UTM Drop and Recovery

ChatGPT traffic lands in GA4's Direct channel for one reason: GA4 received no signal telling it where the visit came from. In the second week of October 2026, many sites watched their ChatGPT sessions fall by close to half while Direct rose on the same pages. The cause was traced to the GPT-6 rollout, after which ChatGPT links lost the utm_source=chatgpt.com tag and the referrer at the same time^[1]. The short version: the visits didn't disappear, the labels did, and on October 15 tagging was reported to be coming back^[2]. This page covers what happened, what it left behind in your reports, and what to fix now.

30-second definitions: UTM, referrer, Direct

• UTM parameters are tracking tags appended to the end of a link URL that tell GA4 the source and medium directly. ChatGPT has been adding utm_source=chatgpt.com.
• The referrer (Referer header) is an HTTP signal in which the browser passes the address of the previous page to the next site. A link opened with the noreferrer attribute does not send it.
• The Direct channel is the GA4 default channel group bucket for sessions whose source is (direct) and whose medium is (not set) or (none)^[5].
• The AI Assistant channel is the GA4 default channel group bucket for sessions whose medium exactly matches ai-assistant, excluding Google AI Overviews and AI Mode^[5].

GA4 can recognize a ChatGPT visit if either the UTM tag or the referrer survives. Lose both and the visit goes to Direct. That is exactly the combination October produced.

What happened in October 2026: timeline

GPT-6 rolled out to paid plans on October 7 and to Free and Go plans on October 8^[2]. Adobe's Cedric Huesler found that GPT-6 opens links through window.open with the noopener,noreferrer options and does not append utm_source=chatgpt.com^[1]. In his words, with no tag and no referrer, "analytics classifies that visit as Direct." The impact was uneven. Inline links in the answer body lost the tag, while links in the Sources panel and shopping links kept it^[2].

The recovery rests on observation, not an official announcement. Huesler reported that as of 9 a.m. Pacific on October 15, tagged ChatGPT clicks across the thousands of domains he tracks were back to roughly 95 to 100% of the same hour in the prior week^[2]. OpenAI's release notes for October 1 to 15 say nothing about link tracking^[2].

Evidence: how big was the impact?

Several measurement teams, working from different samples, reported numbers pointing the same way. Their samples and windows differ, so the figures shouldn't be compared head to head. The overall picture still lines up: ChatGPT sessions fell by roughly 40 to 50%, and Direct rose by about the same amount.

In Huesler's 500-page analysis, Direct rose on two of every three pages that lost ChatGPT traffic, compared with roughly an even split the week before^[1]. The share of mobile Direct visitors who saw a consent banner (a sign of a first visit) climbed from 36% to 44%, moving closer to the 55% seen among ChatGPT visitors^[1]. He described the results as an early analysis based on sample data from October 9 to 13^[1].

LLM Pulse found that total traffic on the same sites fell only about 2%, and that Perplexity and Claude referrals showed no break on October 9^[3]. If every other engine held steady and only ChatGPT dropped, a measurement problem explains the data better than falling demand. At Delante, ChatGPT traffic declined for 11 of its 12 clients, with a median drop of 40%^[4].

Why only part of it vanished: new answer formats and link types

If the tag had disappeared everywhere, the ChatGPT line would have hit zero. Only about half of it went. Profound pointed to Intelligent UI, a new answer format. According to Profound's early analysis as cited by Delante, the share of answers containing tagged source links fell from 98% on October 8 to 75% on October 9 and 66% on October 14, while the share of Intelligent UI answers rose from 2% to about 35% over the same period^[4]. Profound's later, updated post says the rate dropped further to about 10% from 19:00 UTC on October 14 and recovered to about 82% by 03:00 UTC on October 16^[6].

The links shown inside answers changed shape as well. Tomek Rudzki of Peec AI ran 1 million prompts each on October 2 to 7 and October 9 to 13. He found that the share of responses containing a direct webpage URL rose from about 8 to 10% to 42%, and that 54% of those URLs pointed to a homepage^[2]. Those links carried neither a tag nor a referrer. Keep in mind that these figures count displayed links, not actual clicks^[2]. Read alongside the Profound data, the most natural interpretation is that the same GPT-6 answer kept or lost its tags depending on which format it rendered in.

Fewer visits, or lost labels?

The diagnosis converges on lost labels. In an updated post, LLM Pulse concluded that people are still clicking, but many of those clicks arrive as Direct^[3]. The same post left open the possibility that the new interface answers more questions inside the chat and so reduces clicks outright, noting that analytics alone can't measure this^[3]. Search Engine Journal pointed out that Graphite, LLM Pulse, Profound, and Peec AI, the companies behind these measurements, all sell SEO or AI search services, and that LLM Pulse and Profound did not disclose sample sizes^[2]. When you cite these numbers, the honest move is to note those interests alongside them.

The real lesson here is structural, not about one bug. ChatGPT traffic measurement already rested on a single UTM tag. On the sites Graphite tracks, 95% of mobile ChatGPT visits relied on the tag alone, and the referrer is still mostly missing after the recovery^[2]. One wobble in tagging policy, and the AI Assistant channel can drop by half overnight. Under that setup, judging GEO performance by a single traffic line doesn't hold up. You need to track metrics that don't depend on tags, such as citations and mentions inside AI answers, which is the same principle laid out in GEO KPIs and prompt monitoring.

What to do: correcting your GA4 reports now

Search Engine Journal and LLM Pulse both recommended steps 1 and 2^[2][3]. Huesler first used October 8 as his reference point and now treats October 8 to 15 as a window; he advises counting UTM tags, not referrers^[2]. Delante added checking server logs and adding a "How did you hear about us?" survey question^[4]. GA4 does not reclassify processed sessions, so visits that went to Direct during this period stay in Direct^[3]. The correction happens when you interpret the report.

If your channel groups aren't set up yet, start with the GA4 AI assistant channel measurement guide, which walks through a custom channel group that also covers Perplexity and Claude. For baseline traffic shares by engine, see AI referral traffic share 2026. For why different tools report different numbers, see AI visibility tool measurement discrepancies.

How this differs from the Gemini traffic drop in the same period

A Gemini traffic drop was reported around the same time, but its cause was different. LLM Pulse found that gemini-tagged visits nearly vanished between September 28 and October 2 while visits from gemini.google.com held steady or grew, and explained it as a change in how Gemini tags its links^[3]. Search Engine Journal reported that Gemini had added UTM parameters to its links about two weeks earlier^[2]. On the Gemini side, the classification label changed. On the ChatGPT side, the label went missing. Don't lump the two into one chart as if they shared a cause.

Summary

ChatGPT traffic showed up as Direct in GA4 because GPT-6 inline links dropped both the UTM tag and the referrer between October 8 and 15, 2026^[1]. Depending on who measured it, ChatGPT sessions fell by roughly 40 to 50%, with Direct rising on the same pages^[2]. Tagging was reported to start recovering on October 15, but OpenAI has not confirmed this publicly^[2]. The job now is to fence off that window with an annotation, correct your reading through page-level totals and UTM-based counts, and track tag-independent metrics such as citations inside AI answers. For measurement that covers several engines at once, see multi-engine measurement and what GEO is.