If your analytics showed ChatGPT traffic collapsing in the second week of October, there is a good chance your visitors did not go anywhere. Your measurement of them did. Between roughly October 8 and 15, ChatGPT stopped tagging the links it shows in answers, and four independent datasets traced exactly what happened next. The story, pulled together by Search Engine Journal's Matt G. Southern on October 15, 2026, is the cleanest demonstration yet that AI traffic reporting is one quiet product change away from misleading your entire quarter.
The timeline
GPT-6 began rolling out to paid plans on October 7, 2026, and to Free and Go users on October 8. From October 9, ChatGPT began opening inline links in answer text in a new tab, with a browser setting that withholds the referrer, and without the utm_source=chatgpt.com tag its citations had carried since search launched in 2024. Links in the Sources panel and shopping links kept their tags. On mobile, 95 percent of ChatGPT visits to the sites tracked by Graphite relied on that tag alone, making app traffic the most exposed.
What the datasets said
Four teams caught it from four different angles, which is what makes this incident so instructive.
| Source | What they measured | What they saw |
|---|---|---|
| Graphite (CEO Ethan Smith, LinkedIn) | GA4 ChatGPT sessions on sites with meaningful AI Assistant traffic | Down 22 to 57 percent starting October 9, roughly 45 percent below the prior baseline |
| LLM Pulse | ChatGPT sessions across monitored sites, October 9-11 vs prior two weeks | About 46 percent fewer ChatGPT sessions, with total site traffic relatively constant |
| Adobe (Cedric Huesler, LinkedIn) | About 500 pages where ChatGPT drove at least 5 percent of visits, October 9-13 | ChatGPT visits down about 31,000 a day; Direct up about 26,000 a day; Direct rose on two of every three affected pages |
| Profound | Cited ChatGPT answers it collects | Share of answers with at least one tagged source link fell from about 98 percent on October 8 to 66 percent by October 14; the Intelligent UI answer format went from 2 percent to about 35 percent |
| Peec AI (Tomek Rudzki) | The same 1 million prompts re-run October 2-7 and October 9-13 | Direct URLs for webpages in 42 percent of responses, versus about 8-10 percent before; 54 percent of them pointing to homepages; no tags, no referrer |
The behavioral fingerprint filled in the picture. Huesler reported that the new Direct visitors behaved like ChatGPT visitors: on mobile, the share of Direct visitors seeing a consent banner, his marker of a first visit, rose from 36 percent to 44 percent - moving toward the 55 percent he observes among ChatGPT visitors. Adobe called the findings early, based on four days of sampled data. LLM Pulse could not confirm a matching rise in Direct, because Direct is a much larger pool. The sources sell SEO and AI-search tools and Adobe employs one of the analysts, a disclosure worth keeping in view as with any vendor benchmark.
It was a bug, and it is being fixed
OpenAI confirmed to Huesler on October 14 that the change was a bug, according to Search Engine Journal. A fix began rolling out early on October 15, and by 9 a.m. Pacific, tagged ChatGPT clicks were back to roughly 95 to 100 percent of the same hour a week earlier across the thousands of domains Huesler tracks. Not everyone saw it immediately: SEO consultant Glenn Gabe reported he was not yet seeing the fix on some accounts, Profound said it had seen no official word, and OpenAI's release notes for October 1-15 do not mention link tracking.
Two footnotes matter for anyone building AI analytics. First, most ChatGPT clicks still arrive without a referrer, so the UTM tag remains the thing to count. Second, the direction of travel is not uniform across the industry: Gemini added UTM parameters to its links two weeks earlier, and custom UTM parameters set inside ChatGPT Ads clicks were unaffected.
What to do with October's data
- Annotate the window. Huesler and Smith both recommend marking October 8-15 in reporting so nobody interprets the dip as a loss of interest. If a stakeholder asks why AI traffic fell in mid-October, your annotation is the answer.
- Re-read your Direct channel. Compare ChatGPT and Direct side by side, page by page, for every page ChatGPT referred visitors to - and your homepage, where Rudzki saw most of the raw links land.
- Count the tag, not the referrer. If your dashboards key on referrer for AI sessions, switch to utm_source before comparing weeks.
- Keep a second signal. Whether it is a prompt panel you re-run monthly or a log of sampled AI answers about your brand, triangulate against something ChatGPT cannot silently switch off. We covered the same principle when rankings stopped guaranteeing AI citations, and it is why our measurement work always pairs traffic data with answer-side tracking.
The bigger lesson: October broke trust in every measurement layer at once
Step back and the month is almost comically unkind to anyone reporting AI performance. Google reworked links in AI Mode and AI Overviews while giving publishers no click data, and rank trackers have spent weeks complaining Google is feeding them distorted data, as Barry Schwartz has documented. Then ChatGPT, the largest single source of AI referral traffic, silently dropped its tracking tags for a week. When one number in your AI report can move 45 percent because a vendor changed a link format, the report is not a measurement - it is a rumor with a chart.
The workarounds are not exotic: annotate vendor changes, keep raw first-party logs where feasible, define AI segments on parameters you control, and treat any single-source benchmark as directional at best. That is the difference between a dashboard and an instrument, and it is the layer we build for clients when the dashboard stops agreeing with reality. If this month has you questioning your AI analytics, bring the question to the contact page - it is a better conversation to have before the board sees the December slide.
