For one week in October, ChatGPT stopped telling your analytics it sent you. Outbound links in ChatGPT answers arrived with no utm_source=chatgpt.com tag and no referral information — so your analytics quietly filed those visitors under “direct.” Then, on October 15, it turned out to be a bug, and the tags came back.
The episode lasted roughly a week, and it taught the industry something worth more than the bug itself: AI referral measurement is fragile, built on signals that can vanish without warning. If your ChatGPT numbers cratered between October 9 and 15 — while direct traffic mysteriously rose — this is why. And the next time it happens, you will want a way to catch it before it reaches a client report.
Key Takeaways
- Around October 7–9, 2026, shortly after the GPT-6 rollout, ChatGPT’s outbound links stopped carrying the
utm_source=chatgpt.comparameter (introduced June 2026) and suppressed referral information — visits were filed as direct traffic. - Adobe’s Cedric Huesler, analyzing ~500 pages (Oct 9–13 vs. the prior week): ChatGPT-attributed traffic down ~31,000 visits/day, direct traffic up ~26,000/day on the same pages — two-thirds of pages losing ChatGPT referrals saw direct rise. An attribution shift, not a real traffic decline.
- On October 15, Huesler updated his analysis: it was a bug, and ChatGPT was tagging its links again. Per his post, OpenAI confirmed the issue on October 14 and the fix rolled out overnight.
- The playbook: check your data for the signature, report ChatGPT and direct together for that window, and build measurement that does not depend on a single tag surviving.
What Happened to ChatGPT’s UTM Tags in October?
In June 2026, OpenAI added a small gift to marketers: every outbound link in a ChatGPT answer carried a utm_source=chatgpt.com parameter, so analytics platforms could cleanly attribute sessions to ChatGPT. It was the reason the GA4 AI-traffic method worked so well for ChatGPT in particular — the tag was automatic and reliable.
Then GPT-6 rolled out in early October. Somewhere around October 7–9, links in ChatGPT answers started opening via window.open(url, "_blank", "noopener,noreferrer") — a method that both dropped the UTM parameter and suppressed the HTTP referrer. With no tag and no referrer, Google Analytics had nothing to classify the session with except “direct.” The clicks kept happening; the labels disappeared.
The pattern was spotted by Cedric Huesler at Adobe, who posted his analysis on LinkedIn. Search Engine Land reported the story on October 15, and Search Engine Roundtable covered it the same week. For roughly a week, anyone measuring ChatGPT referrals was flying blind.
Was It Intentional or a Bug?
Per Cedric Huesler’s update on October 15, it was a bug. His words: “Relieved. It was a bug. ChatGPT is tagging its links with utm_source=chatgpt(.)com again.” Per his LinkedIn post, OpenAI confirmed the issue on October 14, and the fix rolled out overnight — tagged visits returned to normal levels.
That is the confirmed endpoint of the story, and it matters for how you treat the data. This was not OpenAI changing its measurement philosophy or quietly deprecating referral labels. It was a defect in the GPT-6 link-opening path, caught by an analyst watching his dashboards, confirmed within days, and fixed within a week.
Two honest caveats remain. First, the confirmation came through Huesler’s reporting of OpenAI’s response — OpenAI did not, to my knowledge, publish a standalone public statement about the bug. Second, Huesler’s analysis itself does not establish that every missing referral on every site was caused by the same issue. The ~500-page sample is strong evidence of the mechanism, not a census of the whole web. Say “appears to have affected” rather than “definitely affected” when describing your own data.
How Do I Check Whether My Analytics Were Affected?
Look for the signature — the paired move that identifies the bug:
Compare October 9–13 against the week before. Pull ChatGPT-attributed sessions and direct-traffic sessions for the same landing pages. The bug’s fingerprint is ChatGPT down and direct up on the same pages in the same window. Huesler found direct rose on two-thirds of the pages that lost ChatGPT traffic — if your chart shows that shape, you were affected.
Check landing pages, not just channels. The sessions did not leave your site; they changed labels. The same deep pages ChatGPT used to send people to should show stable or rising pageviews while the channel mix shifts. If pageviews collapsed too, that is a different problem — real traffic loss, not misattribution.
Look at new-visitor behavior. Huesler’s data showed the newly classified “direct” visitors behaving like ChatGPT visitors — on mobile, the share seeing consent banners (new to the site) rose from 36% to 44%, versus 55% for known ChatGPT visitors. In your own GA4, if the “new direct” cohort in that window looks like your pre-October ChatGPT cohort (landing pages, engagement, device mix), that is corroboration.
Annotate the window. In GA4, drop an annotation on October 7–15 marking the ChatGPT UTM outage. Six months from now, nobody will remember the bug — but the annotation will stop someone from building a narrative on corrupted channel data.
What Should I Do When Attribution Breaks Like This?
The general playbook, for this bug and the next one:
Report ChatGPT and direct together for the affected window. The honest chart for October 9–15 is a combined line: ChatGPT-attributed plus direct on the affected pages. That is the number you can defend in a meeting. Reporting the ChatGPT line alone understates reality; reporting direct alone misattributes it.
Triangulate with the manual measurement method. Tags can break, but citations do not. The prompt-portfolio discipline — asking your target questions in clean sessions and logging whether you are cited — measures visibility independently of any tag. If citations held steady while tagged sessions collapsed, you have evidence the traffic was still there.
Use server logs where you have them. During the outage, the referrer was suppressed for most analytics, but server-side signals and user-agent data can corroborate patterns. Do not treat any single signal as proof — the point of server logs here is corroboration, not precision.
Resist the story before you verify the data. The most expensive mistake in this episode would have been telling a client “ChatGPT traffic is falling, our AI visibility work is failing” — and acting on it — before checking whether direct rose on the same pages. Attribution breaks downward first and gets fixed upward; narratives should move slower than dashboards.
How Do I Make AI Traffic Measurement More Resilient?
The bug was a stress test for how we measure AI referrals. The sites that diagnosed it fastest were the ones with layered measurement:
Do not depend on one signal. The GA4 method catches tagged sessions; the prompt portfolio measures citations; server logs catch the residue. Any single layer can fail — as this month proved. Two layers catch a failure; three let you say what the failure was.
Note the platform asymmetry. Gemini still tags its links (per the Oct 15 GA4 guide); ChatGPT’s tags briefly vanished. That asymmetry is itself the story for marketers: every AI surface has its own referral behavior, and your measurement should be per-engine, not blended into one “AI traffic” number. The AI visibility dashboard method is built exactly this way.
Budget for ambiguity in your reporting. The most honest line in an October report: “ChatGPT-attributed sessions fell ~X% and direct rose ~Y% on the same pages between Oct 9–15, consistent with the confirmed UTM tagging bug. Combined, the pages were flat.” That sentence contains the diagnosis, the evidence, and the restraint — and it is the sentence that keeps a client’s trust when the next bug arrives.
Note that referral attribution is only one half of the measurement picture on this platform — for the advertising side of ChatGPT traffic, see the ChatGPT ads guide.
Final Thoughts
The ChatGPT UTM bug was a week-long reminder of a structural truth: we measure AI referrals with tools we do not control, on signals that can change without notice. OpenAI fixed this one fast. The next one — from any platform — may not be a bug at all. Build your measurement so that a missing tag is a footnote in your report, not a crisis in your meeting. Check your October 9–15 data now, annotate the window, and layer your methods. The clicks are still happening. Make sure your dashboards can prove it.
People Also Ask
Did ChatGPT permanently remove its UTM tags?
No — per Cedric Huesler’s October 15 update, it was a bug in the GPT-6 link-opening path. ChatGPT is tagging links with utm_source=chatgpt.com again, and per his post, OpenAI confirmed the issue on October 14 with the fix rolling out overnight.
Why did my direct traffic spike in early October 2026?
If you saw direct rise while ChatGPT-attributed sessions fell on the same pages around October 9–13, the likely cause was this bug: untagged, referrerless ChatGPT clicks were filed as direct. Check the paired signature on your landing pages before drawing other conclusions.
How can I tell if AI traffic is real traffic or just mislabeled direct?
Compare the “new direct” cohort’s behavior (landing pages, device mix, engagement) with your known ChatGPT cohort from before the window, and run a prompt portfolio to verify your citation presence held steady. Citations measure visibility independent of tags.
Does Gemini have the same problem?
No — Gemini still tags its outbound links, per the site’s GA4 AI-traffic guide. The platforms behave differently, which is why per-engine measurement beats a single “AI traffic” number.
What is the difference between attribution and traffic?
Attribution is the label on the visit; traffic is the visit itself. This bug changed labels, not visits. Always verify whether a channel “decline” is fewer visitors or fewer correct labels before acting on it.
