Your AI traffic didn't drop. Your measurement did.

For about a week this month, a large share of ChatGPT's traffic to your site arrived in disguise.

Starting around October 9, 2026, links in GPT-6 answers began opening without the utm_source tag or a referrer, so analytics tools filed those visits as Direct rather than as referrals from ChatGPT. One analysis of roughly 500 pages found ChatGPT visits down about 31,000 a day while Direct rose about 26,000 a day. By October 15, Adobe's Cedric Huesler reported the tags were back and called it a bug. OpenAI has not commented publicly.

The bug is fixed. The lesson is not.

What happened, October 9 to 15

The change tracked the GPT-6 rollout. For several days the links the assistant placed in its answers carried nothing that identified their source, so every tool downstream made the same mistake: it counted a visit that ChatGPT sent as a visit that came from nowhere. The totals did not move. The labels did. When the tags returned on October 15, the lost traffic reappeared, which is how everyone knew where it had gone.

Why counting clicks undercounts AI

A referral click is the last mile of an AI-assisted purchase, not the whole trip. B2B buyers now build their shortlists inside an assistant before they ever touch your site. By the time a click arrives, the assistant has already decided whether to name you. If your only measure of AI visibility is the tagged click, you are grading the one step that happens after the decision is made, and you are trusting a platform to label it correctly.

A three-layer way to measure

We track AI visibility in three layers, none of which a missing tag can erase.

• Citations: whether you are named when buyers ask the questions that decide deals, across each assistant.
• Demand: branded search and direct visits, especially to the pages assistants tend to cite.
• Outcomes: total conversions, rather than channel-by-channel credit that a tag can quietly reassign.

Audit your own October data

Pull Direct traffic for October 9 to 15 on the landing pages ChatGPT tends to cite. A spike there, with no matching campaign, is most likely the mislabeled AI traffic from this window. Note it now, while you can still explain it.

Protect the budget from a glitch

Mark October 9 to 15 in your reports before a quarterly review turns a tracking bug into a budget cut. The sentence that prevents it is simple: AI referral traffic in this window was undercounted because of a platform tagging error, and total conversions did not fall. A number that drops for a technical reason should never be the reason a working channel loses funding.

The click was never the scoreboard. Measure whether the assistant chose you, and whether the business grew, and the next platform hiccup becomes a footnote instead of a fire drill.

Sources: Search Engine Roundtable, LLM Pulse.