If your October dashboard shows ChatGPT traffic falling off a cliff around the 8th while Direct traffic crept up on the same pages, nobody stopped visiting you. ChatGPT stopped saying it had sent them.
For about a week, a bug in the way ChatGPT opened links stripped the little tag that tells your analytics a visitor came from ChatGPT. The visits kept arriving. They just arrived anonymous, and your reports did what reports do with anonymous visits: filed them under Direct, the junk drawer of web analytics.
The bug is fixed, mostly. The lesson isn’t. Every brand that now reports an “AI search” or “AI assistant” line to its leadership is reporting a number that rests on one query string a vendor appends as a courtesy. For a week the courtesy disappeared, and the only reason most marketers found out is that a product leader at Adobe went and asked.
What happened, in order
On October 7, OpenAI began rolling out GPT-6 with what it calls Intelligent UI to paid ChatGPT plans, with Free and Go users following on October 8. Intelligent UI lets ChatGPT answer with charts, comparisons, buttons and small tools instead of plain text. The Verge described responses filling up with pictures, charts and buttons. OpenAI’s release note said nothing about how links inside those new formats would behave.
Cedric Huesler, a senior director of product management at Adobe, puts the start at around 2 p.m. Pacific on October 8. In a LinkedIn recap he wrote that ChatGPT visits his team could attribute dropped “to about a third of normal within hours.” His earlier diagnosis, quoted by Search Engine Roundtable, was that links were opening in a new tab with settings that withhold the referrer, and without the utm_source=chatgpt.com parameter. No tag and no referrer means your analytics has nothing to go on. Direct is the default.
Profound, which tracks ChatGPT answers for a living, logged the same curve from the outside. Before the change, about 99% of the answers it collected with cited links carried at least one tagged link. Across October 9 that fell to about 75%. Then, at 19:00 UTC on October 14, Intelligent UI went from roughly a third of answers to more than nine in ten within half an hour, and the tagged share collapsed to around 10%, then 5%. It stayed between about 5% and 12% for roughly 27 hours. Tags came back in stages, and by 06:30 UTC on October 16 Profound was seeing about 99% again.
The numbers your peers saw line up. Huesler compared October 9 to 13 with the week before on about 500 pages where ChatGPT drove at least 5% of visits. ChatGPT visits fell by about 31,000 a day. Direct rose by about 26,000 a day on the same pages. Search fell 7%, in line with everywhere else. Search Engine Land and Search Engine Journal reported the same figures. SEJ also gathered estimates from Graphite, which saw ChatGPT-attributed sessions in Google Analytics fall 22% to 57% across its tracked sites, and LLM Pulse, which saw about 46% fewer ChatGPT sessions from October 9 to 11 while total traffic held steady. As SEJ pointed out, several of those sources sell AI search tools. They still agree on the shape.
How you found out
Here’s the part that should bother you more than the bug.
According to Huesler, he asked OpenAI what was going on at 9 a.m. Pacific on October 14, and by 11 a.m. it confirmed a known bug with a fix due that week. The fix began rolling out early on October 15. That is a fast turnaround, and credit where it’s due. But as of SEJ’s reporting and AEO/GEO Radar’s October 17 roundup, OpenAI had made no public statement, and its ChatGPT release notes still said nothing about link tracking. I checked them again this morning. Still nothing.
So the official record of a week in which a large share of ChatGPT’s referral attribution vanished is a LinkedIn post by someone at Adobe, a blog post from a vendor, and trade press coverage built on both. Not everyone saw the fix land at once, either. SEO consultant Glenn Gabe replied to Huesler that he wasn’t seeing the change on his own accounts yet, and a commenter on the same post reported days later that tags were still missing in about half of his manual tests.
None of that is a scandal. It’s just what the AI channel actually is right now: a feature of someone else’s front end, maintained on someone else’s release schedule, with no service level and no changelog obligation to you.
Your AI channel is a courtesy
Most marketers assume “AI assistant traffic” is detected the way search traffic is, from the referrer. Mostly it isn’t. Huesler’s advice after the fix was blunt: keep counting the tag, because the referrer is still missing on most ChatGPT clicks. Graphite’s Ethan Smith said, according to SEJ, that 95% of mobile ChatGPT visits to its sites relied on the tag alone.
Now look at how Google Analytics builds its own AI Assistant channel, the one it added as a default in May. The rule is that the medium exactly matches “ai-assistant”, and Analytics sets that medium when the referrer matches its list of AI assistants. ChatGPT’s tag sets a source, not that medium. A visit that arrives tagged but with no referrer can end up outside the AI Assistant channel even on a normal day, which is why the people tracking this closely tell you to look at session source as well as the channel total.
Google’s own assistant only started adding tags to Gemini links around October 1, according to Search Engine Journal, and Google hasn’t documented when they appear. So ChatGPT and Gemini both tag links now, neither under any obligation to you, and one of them just showed what a bad deploy does to your numbers.
There’s an uncomfortable asymmetry here too. Smith also said, per SEJ, that UTM parameters you set yourself on ChatGPT Ads clicks should have been unaffected. So for a week, the paid side of ChatGPT stayed measurable and the free side went dark. Put those two lines next to each other in an October deck without a footnote and you have a tidy, wrong story: ads work, organic died.
Why this week matters for budgets
This would be a footnote if AI search were a rounding error. It isn’t, at least not in planning documents. In Digiday’s Media Agency Report, published October 5, 77% of agency respondents said they plan to increase budget allocation for AI search and GEO in the next 12 months, and 20% of those said the money would come out of traditional search.
Those decisions get made on numbers exactly like the ones that just broke. October is when a lot of teams pull month-to-date reports for Q4 reviews and 2027 plans. A 45% drop in AI referrals in the first half of the month is the kind of chart that kills a line item. Nobody in finance will ask whether it was a front-end bug in someone else’s product.
The error runs the other way too. The misfiled visits didn’t vanish. They landed somewhere, and that somewhere may be getting credit. Peec AI’s Tomek Rudzki ran the same million prompts before and after the change and, per SEJ, found ChatGPT putting a direct URL in 42% of responses, up from roughly 8 to 10%, with 54% of those pointing to home pages. Those are links shown, not clicks, but if your homepage Direct line jumped in mid-October, check before anyone attributes it to the brand campaign that happened to launch that week.
And there’s a quieter cost. Any model that learns from your channel data, from a media mix model to the budget pacing tool your agency runs, just ate a week of mislabeled history. Profound made the point plainly: the tags coming back doesn’t recover what was already lost.
Don’t rebuild the model, rebuild the dependency
Huesler’s advice was not to rebuild your attribution model around one bad week. He’s right about the model. The visits never went away, and the dip is one event, not a trend.
But the dependency deserves a rebuild. If one query string from one vendor is the only thing standing between your AI channel and your Direct bucket, you don’t have a channel. You have a convention. The fix is not a better dashboard. It’s a second signal you own, so that the next time a tag disappears you see it within a day and can still answer the budget question.
What to do this week
- Annotate October 8 to 16 everywhere the data lives. Google Analytics, Adobe Analytics, your BI dashboards, and the notes that travel with any data you feed to a mix model or an agency. Huesler puts the window at October 8 to 15. Profound saw tags still recovering until early on the 16th, so mark the wider window.
- Estimate the misfiled visits and write the estimate down. Take your top ChatGPT landing pages and your homepage. Compare Direct sessions for October 8 to 16 with the same days the week before. A Direct rise that mirrors the ChatGPT fall is your best estimate of what moved. Put it in the October report as an adjusted line, clearly labeled, instead of leaving the raw dip on the slide.
- Check where tagged visits actually land. Look at session source equal to chatgpt.com, not only the AI Assistant channel total. If tagged visits are falling into Unassigned or Referral, add a custom channel rule for known assistant sources and place it above Direct and Unassigned.
- Test it yourself, weekly, on a phone. Ask ChatGPT and Gemini three questions a buyer in your category would ask. Click the cited links inside the answer, in the source list and in any product panel, on mobile and on desktop. Check whether the landing URL carries a tag. It takes ten minutes and would have caught this on October 9.
- Set an alert you’ll actually see. Trigger it when ChatGPT sessions fall more than about 30% week on week while Direct on the same landing pages rises. That pattern is the signature of a tracking break, not lost demand.
- Add a signal no vendor can switch off. Put an AI assistant option in your “how did you hear about us” field on lead forms and at checkout. Track whether assistants name you at all, separately from clicks. Neither depends on a query string.
- Change the budget conversation before it happens. Tell finance and your agency now that AI referral numbers for October carry a known break, and agree that no GEO line gets cut or doubled on referral sessions alone this quarter.
- If you buy ChatGPT Ads, keep your own UTMs on every click, and don’t let a stable paid line next to a broken organic line become the story.
Nobody at OpenAI broke your reporting on purpose. That’s the point. The AI channel you report to leadership is held up by a convention someone else maintains, and it took a week and a stranger’s LinkedIn post to learn it had slipped. The brands that come out of planning season with their AI budgets intact won’t be the ones with the best dashboards. They’ll be the ones who noticed the gap, wrote it down and built a number of their own.
