ChatGPT referral traffic fell by half this month, and it was a tracking bug, not fewer visitors

Software sites watched their ChatGPT-referred traffic in Google Analytics 4 appear to collapse by roughly half starting October 9, and the real cause was a tagging glitch inside ChatGPT, not a sudden loss of visitors. Four separate measurement efforts caught the same event within days of each other. The monitoring service LLM Pulse put the decline in ChatGPT referral sessions at 46%, comparing October 9-11 against the two weeks before, with overall site traffic barely moving. The SEO platform Graphite logged declines ranging from 22% up to 57% on sites that normally get meaningful AI-assistant traffic, averaging out to roughly a 45% drop from baseline, per Search Engine Journal's reporting. Adobe's Cedric Huesler pulled a sample of roughly 500 pages that each got at least 5% of their traffic from ChatGPT and compared the week before and after: those pages lost close to 31,000 ChatGPT sessions daily, almost matching a roughly 26,000-session daily gain in the unlabeled Direct bucket, according to Search Engine Roundtable.

What actually changed

The timing lines up with OpenAI's GPT-6 rollout: paid accounts got it October 7, free accounts October 8. LLM Pulse found no break in the data on October 7, only after free users were upgraded the next day. Huesler traced the mechanism: GPT-6 began opening inline links in a way that withholds the referrer and drops the utm_source=chatgpt.com tag ChatGPT had carried on its links since 2024, so the visit lands in analytics as Direct traffic instead of a ChatGPT referral.

Not every link broke the same way. Per Search Engine Journal's account of Graphite's numbers, citations inside the body of a ChatGPT answer lost their tracking tag, but links shown in the separate Sources panel, plus shopping links, kept theirs — and because 95% of Graphite's mobile ChatGPT visitors arrived only through a tagged link, phones took the brunt of the miscount. The AI-visibility firm Profound ran its own count of citations across ChatGPT answers and watched the portion carrying any working tracking link slide from near-universal on October 8 to just two-thirds by October 14, a slide it linked to a newer "Intelligent UI" answer layout that went from a sliver of its sample to more than a third over the same stretch. A fourth firm, Peec AI, re-ran a fixed set of one million prompts before and after the change and saw ChatGPT embed a bare homepage or article link directly in its answers roughly four times as often as before — about 42% of responses, versus 8% to 10% beforehand — with slightly more than half of those links pointing at a brand's homepage rather than a specific page.

Worth flagging: three of the four measurement firms here — Graphite, Profound and Peec AI, plus LLM Pulse — sell SEO or AI-visibility tooling, and Huesler's own employer, Adobe, sells analytics software. Each has a commercial stake in marketers caring more about AI referral numbers. That bias doesn't erase the point that four unrelated methodologies converged on the same week and a similar magnitude of change.

Who this affects, and for how long

Huesler told Search Engine Roundtable that OpenAI acknowledged the issue to him as a bug on October 14, and he saw tagged links reappear the next morning. By the start of the business day on the West Coast, his tracking across thousands of domains showed tagged ChatGPT clicks back in line with where they stood seven days prior.

"Relieved. It was a bug." — Cedric Huesler, Adobe, via Search Engine Roundtable

The window was brief only in retrospect. LLM Pulse's own data shows the drop held for roughly a full week, which is long enough to distort any week-over-week or month-over-month report that reads GA4's channel grouping without a note attached. LLM Pulse also flagged a related, earlier change: Gemini referrals broke in a similar way between September 28 and October 2, with sessions tagged "gemini" nearly disappearing while visits from gemini.google.com held steady, which points to a tagging change rather than a traffic change there too.

What a software marketer should do

• Annotate October 8 through 15 in GA4 and in any dashboard that reports a ChatGPT or "AI Assistant" channel, so nobody reads that week as a drop in AI-driven demand.
• Check whether Direct sessions rose on the same landing pages that lost ChatGPT-attributed sessions during that window. A rise there points to mislabeling, not fewer visitors.
• Hold off on rebuilding attribution models or cutting AI-referral reporting based on one week of data — the tagging issue already appears fixed.
• Treat single-week swings in any AI-assistant channel as a tagging question first, before a demand question. This is the second such change in two months, after a similar shift in how Gemini tagged its links in late September.

None of this was disclosed by OpenAI. Its ChatGPT release notes for October 1 through 15 do not mention a change to link tracking, and the fix was confirmed only through Huesler's own testing, not an official statement. For marketers who built reporting around a ChatGPT referral line, the lesson isn't really about ChatGPT: a tracking tag you don't control, and that no vendor is obligated to announce, can vanish from your dashboard overnight.