ChatGPT Stopped Tagging Its Clicks. Your AI Traffic Now Shows Up as “Direct.”
On October 9, an analytics lead at Adobe opened his dashboard and watched a month of ChatGPT referrals evaporate. Not gradually. Overnight, across the ~500 pages where ChatGPT had been driving 5% or more of visits, roughly 31,000 ChatGPT-attributed visits per day disappeared. On the same pages, “Direct” traffic rose by about 26,000 a day. Search traffic moved by −7%, the same as everywhere else on the site.
The traffic did not leave. The label did.
That is the thing to understand about the week of October 9–13, 2026: the most consequential change to AI referral measurement in the last two years happened as a side effect of a model release, and nobody announced it. GPT-6 shipped, the links inside ChatGPT results changed how they open, and a channel that had taken three years to become visible went dark again.
What actually changed
In June 2026, OpenAI added UTM parameters to links inside ChatGPT results. For the first time, a site owner could filter a report by utm_source=chatgpt.com and see the channel as its own line. That was a meaningful upgrade over guessing from referrer strings, which AI clients had already started stripping.
With GPT-6, that tagging stopped. Links now open with window.open(url, "_blank", "noopener,noreferrer") and no UTM parameters. No tag, and no referrer either. Analytics tools see an anonymous, unadorned visit and file it under Direct.
The tell is in the comparison. In the week before the change, whether Direct rose on a page was close to a coin flip. In the observed window, Direct rose on two out of every three pages that lost ChatGPT traffic, and those new Direct visitors behaved like the ChatGPT visitors they replaced. On mobile, the session shape matched almost exactly. That is not a coincidence in the data; that is the same people arriving through a door whose sign got taken down.
Why the numbers still look fine, and why that’s the problem
Here is the uncomfortable part. A site whose ChatGPT share was 6% now shows 0% ChatGPT and 6% more Direct. Total sessions are unchanged. Nothing in a standard dashboard flags a problem, because from the dashboard’s point of view nothing happened.
What changes is your ability to answer questions that decide budget:
- Is the AI channel growing or flat this quarter?
- Which pages earn citations and which get ignored?
- Did the rewrite of your comparison page help, hurt, or do nothing?
Every one of those questions used to be answerable with a filter. Now it has to be answered with inference, and inference does not survive a budget meeting.
There is a second-order effect that matters more. When a channel becomes unmeasurable, it becomes invisible to the people who allocate money. The channels that keep their attribution keep their funding. This is how a channel that is genuinely growing gets cut, and it has happened before — dark social, WhatsApp shares, podcast mentions, and every app that stopped passing a referrer.
Three measurement approaches that still work
None of these replace clean attribution. All three survive it being taken away.
Self-reported attribution at the point of signup. Add a one-question field to your signup or demo form: “How did you hear about us?” with an option for AI assistants. Response quality is mediocre on free signups and much better on demo requests and paid conversions. It is the only method that tells you what a human believed, rather than what a browser passed.
Brand and direct demand as a proxy. Track branded search volume and direct visit trends monthly, and compare them against the weeks when your pages were being cited. When an AI assistant names your product, a slice of the people who see it will type your name rather than click the link. That shows up in branded search, not in referrals, and it always did.
The link and citation layer. This is the one that does not depend on any client’s decisions. A citation in an AI answer is generated from sources on the open web — pages an assistant can read, quote, and attribute. Those sources are inspectable from outside, by anyone, with or without platform cooperation. If a listing on an industry directory, a comparison article, or a review page is what put your product in front of an assistant, that listing exists whether or not a click ever gets tagged.
A practical note on the third one: it argues for looking at corroboration, not volume. Ten directories that all copy the same 40-word description from your homepage create one claim repeated ten times. Ten properties that each describe what you do in their own words create ten independent claims. Only the second kind is useful when an assistant is trying to decide whether your product is real.
What to do this week
| Task | Why now |
|---|---|
| Add a “How did you hear about us?” field | Historical data can’t be recovered; start collecting immediately |
| Split Direct into “Direct — new” and “Direct — returning” | New Direct visits are where the disguised AI traffic lands |
| Pull branded search volume month over month | Separates real demand growth from attribution loss |
| Inventory where your product is described off-site | This is your AI visibility surface, and it’s measurable |
The fourth row is the one most teams have never done. Write down every third-party page that describes your product in its own words: directories, comparison posts, review sites, community threads, partner pages. Then check how many of them are indexed, how many are still live, and how many were written by someone who actually looked at your product. Forty real ones beat four hundred templated ones, and the forty will outlast every attribution change OpenAI makes this year.
The uncomfortable takeaway
ChatGPT removed its tracking tags the same week it started selling ads. Read that pairing however you like, but the practical consequence is the same: the free, measurable slice of AI attention is shrinking, and the part you can verify from the outside is the part you own.
If you have been treating directory listings and citations as a low-value chore you’ll get to eventually, the last week was a reminder of the trade-off. Referral data will keep being taken away and given back on someone else’s schedule. A citation on a page that a search engine and a language model can both read is yours, and it keeps working whether or not it’s tracked.
That layer is the one you can act on while the measurement question stays unresolved, and the part that benefits from being run on a schedule rather than in bursts: https://backlinksubmit.org matches your product to relevant directories and communities from a handpicked list, submits manually, and reports in a console which platforms have been submitted and how each is progressing.
FAQ
Is my ChatGPT traffic actually down, or just relabeled?
Relabeled, in almost every case we’ve seen. If Direct rose on the same pages that lost ChatGPT traffic in the same week, and the sessions behave the same way, you are looking at one stream with two labels. Compare page-level deltas, not site-wide totals — the site-wide numbers hide it.
Will OpenAI bring the UTMs back?
Unknown, and planning around it would be a mistake. The tagging existed for about four months. Treat AI referral data as best-effort telemetry, not as a system of record.
Do directory submissions still make sense in this environment?
They never depended on referral tracking. A directory listing earns its value by telling search engines and language models that a business with a specific description exists at a specific domain. That value does not appear in a referral report whether or not UTMs exist, which is why teams that only measure referrals consistently undervalue it.
What’s the single highest-value action today?
Add the self-reported attribution field. It costs an afternoon, it survives every future platform change, and the data starts accumulating the moment it ships.
Share Article
