The Week ChatGPT Lost Its Referrer

01Three products, one bad week

On 9 October the ChatGPT line fell on three of my products on the same day. Sign ups attributed to ChatGPT on the first went from about 78 a day to 32. Visits on the second went from about 83 a day to 21. The third, which only ever got a handful, went to zero. Nothing had been deployed, nothing had been blocked, and the live agent that reads pages for ChatGPT users was still fetching them. Three unrelated products losing the same channel on the same morning is not three problems. When a channel drops on every product at once, suspect the measurement before the market.

The label fell off

The cause, reported by Search Engine Journal and acknowledged second hand by OpenAI, was a tracking bug. A newer model had reached free ChatGPT users the day before, and links clicked in its answers stopped carrying the small tag that says where a visit came from. The visitors still arrived. Analytics filed them as direct traffic, or dropped them into whatever bucket catches visits with no source. A fix rolled out from 15 October, and by the following week the tagged line was back.

02The tag returned, the referrer did not

After the fix

From 16 to 21 October the first product recorded 120 to 126 ChatGPT sign ups a day, more than before the bug. The second came back to about 63 visits a day, below its September level. The third recovered to about one a day. I cannot tell how much of the first product's jump was real growth and how much was a backlog of people returning, and I will not guess. A recovery that overshoots after a tracking bug is partly people, partly arithmetic, and you rarely know the split.

What never came back

The tag recovered. The referrer did not. Before this, many ChatGPT clicks also arrived with the chatgpt.com address in the referrer header, which some analytics tools use to recognise the source. After it, they mostly did not. In the week I measured in the referrers that never clicked, 244 requests carried a ChatGPT referrer and not one of them was a page load. The tag in the address is now the only dependable signal. Any report that still counts ChatGPT by referrer is undercounting, and has been since the second week of October. Nobody can say by how much, because a visit that lost its referrer looks exactly like a person typing your address into the browser, and lands in the same bucket.

03How to measure a channel you do not control

Keep a second witness

A single tracking signal is a single point of failure. What made this readable as a bug, and not as ChatGPT abandoning my products, was a second witness. On another site where I count ChatGPT's live agent in the server log, its daily page fetches that week stayed in the same range as the week before. When the fetches hold and the visits vanish, the visits did not vanish. When one measurement drops and the other holds, believe the one that did not depend on a tag. That is the same reason I keep server logs beside analytics, the split described in nobody reads your blog.

Annotate, then compare like with like

Mark 9 to 15 October in every chart that includes ChatGPT traffic, and do not compare a week inside it with a week outside it. Count ChatGPT by the utm_source=chatgpt.com tag, check the user agent of tagged landings because bots copy the tag, and treat referrer based numbers from before October as a different series from those after. And remember what no tag can show. Most people who use a product through an assistant never visit at all, as the users who never open your website found. Referral traffic from an assistant is the smallest, noisiest part of what that assistant does for you. Measure it, but do not let it be the only number you look at.

Tomorrow: why every number I publish carries its source in the same sentence.