ChatGPT Dropped Its UTM Tags for a Week

On Thursday 9 October 2026, the ChatGPT row in a great many GA4 reports shrank, and the Direct row grew by roughly the same amount. The visitors had not gone anywhere. Links in ChatGPT answers had started arriving without utm_source=chatgpt.com, the ChatGPT UTM tag that OpenAI's own help centre says is added to referral URLs, and most of those clicks carried no referrer either. A visit with neither has no name, so analytics filed it under Direct. The tags came back between 15 and 16 October. The sessions recorded in between stayed where they had landed.

This is a short note on what is known, how firmly it is known, and what the episode says about counting AI traffic from analytics alone.

The missing ChatGPT UTM tags: what is confirmed

The facts sit on three levels, and they are worth keeping apart.

• Documented by OpenAI. ChatGPT appends utm_source=chatgpt.com to the links it sends. This is written in OpenAI's publisher FAQ, and it is the basis of nearly every GA4 setup for AI traffic.
• Reported by the trade press. Search Engine Roundtable described the missing tags on 14 October and Search Engine Journal on 15 October, both drawing on practitioners' data. The pattern is consistent across their accounts: links inside the answer text lost the tag, links in the sources panel mostly kept it, and Direct rose while ChatGPT fell. The start falls between 8 and 9 October, the days GPT-6 reached every plan and brought a new answer layout with it.
• Not confirmed in public. That it was a bug. Someone at Adobe who follows this traffic wrote on 15 October that OpenAI had told him so privately and that a fix was rolling out. OpenAI has said nothing in public, and its release notes for the first half of October do not mention link tracking.

So the dates are solid enough to annotate a report. The cause is not. Anyone writing that OpenAI changed its policy is going beyond the evidence, and so is anyone writing that nothing happened.

Behind the dates sits a structural weakness. A UTM parameter is a courtesy. It is text the sender chooses to append to a URL, and nothing obliges the sender to keep doing it. The referrer is no sturdier: the browser sends it or withholds it depending on how the link is opened. When both disappear together, analytics has nothing to reason from. It does not guess. It writes Direct.

Three consequences follow.

• The gap is permanent in your history. Tags returning on 16 October do not relabel the week before. Next October, a year-on-year comparison will show a dip that is an accounting artefact. Annotate the period in GA4 now, while the reason is still fresh.
• Direct was never a neutral bucket. It already holds bookmarks, typed addresses, and links opened from email clients and PDFs. A week of misfiled ChatGPT visits dissolves into it, and nothing lets you separate them afterwards.
• It can happen again, quietly. This episode was noticed because the drop was sharp and a lot of people watch ChatGPT traffic. A slow erosion, or the same change in a smaller assistant, would read as ordinary decline.

None of this makes the GA4 setup pointless. The method in How to Track ChatGPT and Perplexity Traffic in GA4 still holds, and it started working again the moment the tags returned. It is a sensible floor. It is not a measure of presence.

The deeper limit has nothing to do with tags. Sessions exist only when somebody clicks. An assistant can name your company, describe what you make and recommend you to a buyer who then phones, writes to your distributor or types your address a day later. None of that reaches the ChatGPT row in any week, tagged or not. Appearing in an answer and receiving a visit from it are different events, and the difference between being mentioned, cited and used as a source is set out in Mention, citation, source.

The first event is measured by reading the answers themselves. That is what PSentry does: it puts the questions of your sector to ChatGPT, Claude, Gemini and Perplexity twice a month, in each language you sell in, and records whether your brand is named. It uses no tags and no referrers. It also cannot tell you how many visits an answer produced; for that you still need GA4. The two measures answer different questions, and the week of 8 to 16 October is the plainest argument for keeping both.