utm_source=chatgpt.com tag and opened without a referrer. Google Analytics files a visit with neither as Direct, so the ChatGPT line in GA4 fell by roughly 30% to 46% in the two datasets that gave a percentage. OpenAI confirmed a known bug on 14 October and a fix began rolling out on 15 October. None of the sources reports any change in brand visibility that week. What changed was the label on a click, and for six days nobody could tell a relabelled visit from a lost one.
What happened, with dates
OpenAI rolled GPT-6 into ChatGPT for paid plans on 7 October and for Free and Go plans on 8 October, according to its own announcement. From about 9 October, many links in answers opened with a browser call that set noopener,noreferrer and dropped the ?utm_source=chatgpt.com parameter ChatGPT had appended since June. Cedric Huesler at Adobe described the mechanism, reported by Search Engine Roundtable on 14 October: "There's no tag and no referrer, so your analytics files the visit under direct."
Huesler contacted OpenAI on 14 October and the company confirmed the missing parameters were a known bug, with a fix expected that week. The rollout began early on 15 October, and by 9 a.m. Pacific time tagged ChatGPT clicks were back to roughly 95% to 100% of the level recorded at the same hour the previous week, based on telemetry covering thousands of domains, according to Search Engine Land's report of 20 October.
What the week looked like in the data
- Adobe (Huesler). Across about 500 pages where ChatGPT drove 5% or more of visits, 9 to 13 October against the week before: ChatGPT down about 31,000 visits a day, Direct up about 26,000 a day on the same pages, Direct rising on two-thirds of the pages that lost ChatGPT traffic.
- LLM Pulse (Esteve Castells, 13 October). Across the sites it monitors, ChatGPT sessions in GA4 fell about 46% on 9 to 11 October against the same days in the two previous weeks; more than eight in ten sites lost over 30%; total traffic stayed flat, around 2% down. Its first version could not see Direct rise at site level; its 16 October update, citing Huesler's page-level data, concluded the evidence "points clearly to lost labels". No sample size given.
- Profound (Josh Blyskal, LinkedIn). Between 8 and 14 October the share of collected answers whose cited sources carried a UTM tag fell from 98% to 66%, alongside the rollout of ChatGPT's new interface. He reported teams seeing falls of roughly 30% to 40%.
- Peec AI (Tomek Rudzki, LinkedIn). On the same one million prompts, 2 to 7 October against 9 to 13 October, ChatGPT placed a direct link to a website inside 42% of answers, against 8% to 10% a week earlier, with no tag and no referrer; 54% of those links pointed at a homepage.
Four sources, four readings of the same week. Adobe saw the visits reappear as Direct on the pages that had depended on ChatGPT. LLM Pulse could not see that at site level and only agreed once it looked page by page. None of the underlying data is published, and while the bug lasted none of them could prove whether people had stopped clicking or had kept clicking without a label. LLM Pulse still allows that the new interface may resolve more questions inside the chat and send fewer clicks.
What we saw ourselves. On the Google Analytics properties we could read on 15 October, where there was enough ChatGPT volume to read, sessions from the ChatGPT source between 9 and 13 October ran roughly two-thirds below the daily average of the three weeks before, while total sessions did not fall in step. ChatGPT is small relative to Direct on every property we read, so we could not see the missing visits land anywhere. A single-property read confirms the fall and cannot settle the cause.
What the week showed
A referral line measures a label, not a behaviour. It is the same trap as the visibility dashboard that reports a number without a decision. The GA4 ChatGPT row only ever counted visits that arrived with a tag or a referrer. The tag is OpenAI's to add or drop, and this time it dropped by accident. The next change may be deliberate: ChatGPT's new interface can answer with interactive elements inside the chat, and Peec found that over half of the new in-answer links point at a homepage rather than the page that answered. Either could move the referral line without moving your visibility.
Aggregate analytics cannot separate relabelled from lost. A few dozen visits a day moving from one bucket into a Direct channel of thousands is invisible at the top line. Page-level comparison, as Huesler did, was the one GA4 view that showed something, and even that could not prove the cause.
Visibility and traffic are different measurements. Whether an answer names your brand can only be read by asking the questions and reading the answers. Whether somebody clicked, and whether the click was labelled, is a separate question with a separate owner. The week sharply degraded the second measurement and left the first untouched.
What to keep doing
- Annotate 9 to 15 October in GA4 and in any dashboard that charts AI referrals, so the dip is never read later as a loss of visibility.
- Stop using the ChatGPT referral line as a visibility measure. Measure the mention itself: a fixed panel of buyer questions, put to each engine on a schedule, with the answers kept. Our travel edition shows what that reading looks like across six engines and two captures.
- Watch Direct to the pages ChatGPT sends people to, and to your homepage, page by page, whenever the referral line moves sharply.
- Keep your custom channel rules for the engines that still pass a referrer, and treat every engine's tagging as temporary, as Dalin de Graff advises. Google, Perplexity and the others can change theirs just as quietly.
- Add a self-reported source to your main form ("How did you hear about us?") and reconcile it against GA4 each month. It is crude, and it survives tag changes.
Frequently asked questions
Is ChatGPT traffic back to normal in GA4?
According to Search Engine Land's 20 October report, tagged ChatGPT clicks were back to roughly the previous week's level by the morning of 15 October. Check your own ChatGPT line from 15 October onwards; if it has not recovered, the cause is something other than this bug.
Why did GA4 show ChatGPT visits as Direct?
GA4 attributes a session from the referrer header or from UTM parameters in the URL. For about six days many ChatGPT links arrived with neither, so those sessions fell into Direct, the same bucket as a typed address or a bookmark.
Does referral traffic measure AI visibility?
Not on its own. Visibility is whether the answer names you. Referral traffic is whether somebody clicked and the click was labelled. The second can vanish for a week while the first does not move. If you want an outside read of the first, our free AI visibility check puts your brand to five engines, and the AI Search Diagnostic builds the question panel properly.
Sources
- OpenAI, GPT-6 and Intelligent UI for everyone, 7 October 2026.
- Barry Schwartz, OpenAI Drops UTM Parameters From ChatGPT Results (GPT-6), Search Engine Roundtable, 14 October 2026, quoting Cedric Huesler (Adobe).
- Anu Adegbola, ChatGPT drop in UTM tracking was a bug, Search Engine Land, 20 October 2026.
- Esteve Castells, ChatGPT Referral Traffic Fell by Almost Half on October 9, 2026. Here's What We Know, LLM Pulse, 13 October 2026.
- Josh Blyskal (Profound), LinkedIn post, 14 October 2026.
- Tomek Rudzki (Peec AI), LinkedIn post, 14 October 2026.
- Dalin de Graff, ChatGPT's UTM Tags Disappeared With GPT-6, 14 October 2026.