For about a week in October, the ChatGPT line in a lot of GA4 properties dropped by roughly half. Much of that traffic never left. It was relabeled and sat inside Direct.

I care less about that week than about what it exposed. If your analytics only recognizes the visit that arrives with a tag, then traffic from AI assistants was always partly hidden in Direct. October made it visible because the tag was there one day and gone the next.

What happened, with dates

Search Engine Journal reported on October 15 that GPT-6 began rolling out on October 7 for paid plans and October 8 for Free and Go users. The GA4 drop began October 9. Cedric Huesler of Adobe describes the mechanism: ChatGPT started opening links in a new tab with a setting that withholds the referrer, and stopped adding the utm_source=chatgpt.com tag its citations had carried until then. No referrer and no tag mean GA4 has nothing to sort the visit by, so it lands in Direct.

Huesler puts the affected window at October 8 to 15. He says OpenAI confirmed a bug on October 14 and that a fix started rolling out early on the 15th. By 9 a.m. Pacific that day, tagged clicks across the thousands of domains he tracks were back at 95% to 100% of the same hour a week earlier. That confirmation is secondhand, from his LinkedIn post. In the reports I read, there is no statement from OpenAI itself.

Melissa Rosenthal, co-founder of Outlever, put it on my radar with “around Oct 9” and “around the 15th.” The sources back her, with one correction: it was ChatGPT that restored its tagging, not GA4 that got fixed. And the mislabeled week stays mislabeled in your reports.

The numbers, and who is selling what

• Adobe (Huesler): about 500 pages where ChatGPT supplied at least 5% of visits, October 9 to 13 against the prior week. ChatGPT about 31,000 visits a day lower, Direct about 26,000 higher, search down 7%.
• Graphite (Smith): ChatGPT-attributed sessions down 22% to 57% on sites with meaningful AI traffic. He says about 95% of mobile ChatGPT visits to those sites depended on the tag alone.
• Profound: answers with at least one tagged link went from about 98% on October 8 to 66% on October 14.
• Delante: ChatGPT sessions fell in 11 of its 12 projects, median drop 40%.

Read those with two cautions. Graphite, Profound and Delante sell SEO or AI-visibility services, and Huesler works at Adobe. And the 31,000 and the 26,000 do not match, so Direct did not absorb everything ChatGPT lost. Huesler himself calls his results preliminary, from four days of sampled data.

Five checks for your own GA4

Start with the Direct channel, broken out by landing page. Compare October 1 to 7 with October 8 to 15, and look first at deep pages that used to receive ChatGPT sessions.

Then put ChatGPT and Direct in one table. A fall in the AI Assistant channel alone does not justify a strategy change. Total sessions and conversions tell you more.

Third, look at new users inside Direct. On mobile, Huesler saw the share of Direct visitors with a first-visit signal go from 36% to 44%. A Direct segment with more new people than usual points to untagged first visits.

Fourth, check the homepage. Peec AI found that ChatGPT linked straight to sites in 42% of answers, up from 8% to 10%, and 54% of those links went to homepages with no tag. That figure counts links displayed, not clicks.

Fifth, annotate October 8 to 15 in GA4 if your property supports it, and in any dashboard that feeds a board report. Months that straddle October 8 do not compare without an adjustment.

If you run ChatGPT ads with your own UTMs, those clicks were not affected. Do not extend what you learn about organic traffic to them.

And add one question to your contact or sign-up form: “How did you hear about us?” with “an AI assistant” as an option. No parameter can take that one away.

What not to conclude

Two tempting conclusions are both wrong. One: ChatGPT stopped sending you customers. You cannot know that from this data, and neither can anyone else. LLM Pulse saw ChatGPT sessions about 46% lower with total traffic flat, but its aggregates cannot tell lost visits from relabeled ones. Two: every extra Direct session in October is ChatGPT. That is a hypothesis your landing pages can support or kill.

Rosenthal’s advice is to keep tracking AI referrals without building strategy on them, and to lean on branded search, the how-did-you-hear question and net-new pipeline. I agree with half of that. A single AI-referral number is unreliable, which is why I would report it as a range.

Also, most ChatGPT clicks still arrive without a referrer. With the tag back, the signal is what it was before October: one parameter that another company can remove on a Thursday without telling you.

How we measure it at IQ Source

When we measure AEO, we use three layers, and none depends on a single parameter. Analytics comes first: the AI Assistant channel, a Direct segment on pages that used to receive assistant traffic, and annotations for every platform change we hear about. Visibility inside the answers comes second, measured by repeating each query with the method on the table (I wrote about why a visibility percentage without a method can be noise). What people tell you comes third: how they found you, logged by sales.

That gives you AI traffic as a band built from three sources instead of one GA4 line. It is the logic behind our AEO work, and it follows from what we wrote about being chosen, not just seen, in AI answers.

If your October report already went out showing the drop, open Direct by landing page before your next meeting. It may change what you tell your boss.

See how we measure traffic that arrives from AI assistants