ChatGPT didn’t stop sending visitors.It stopped telling you it sent them.
From 9 October 2026, the typical small business website in our referral and server logs lost 62% of its visits tagged as coming from ChatGPT, while ChatGPT’s own fetcher, which reads a page when someone asks it a question, dipped far less. At 14:00 UTC on 15 October the tags came back, and tagged visits went straight back to their usual level. That points to a missing label, not missing visitors.
ChatGPT’s tags came back at 14:00 UTC on 15 October
Until 13:59 UTC on 15 October, ChatGPT-tagged visits stayed as low as on the gap days. From 14:00 they were back to their usual level on sites across the logs at once, and they have stayed there: 681 tagged visits from 00:00 to 15:59 UTC on 16 October, against an average of 679 for the same hours from 30 September to 8 October.
A jump that size, in the same hour across the sites, is a switch on ChatGPT’s side. OpenAI has said nothing publicly. Adobe’s Cedric Huesler reported on 15 October that OpenAI had confirmed the missing tag to him as a bug, with a fix rolling out that day, as Search Engine Journal reported. The fix did not reach back: the days without a tag stay without one.
That is the median across the 184 sites that averaged at least one ChatGPT visit a day from 24 September to 8 October 2026, each comparing itself over 9 to 14 October. All sites combined fell less, 39%, from 1,034 to 629 tagged visits a day, because the busiest sites lost proportionally less. Over the same days ChatGPT read pages on these sites 7,842 times a day and then 6,712, a much smaller dip. By Wednesday 14 October it was reading as much as the Wednesday before: 7,845 reads against 7,792, while tagged visits had halved, 1,094 to 527.
For every 1,000 times ChatGPT read a page, fewer visits followed
Tagged visits per 1,000 ChatGPT reads were 132 across 24 September to 8 October 2026. They were 102 on Friday 9 October, then lower again on each weekday after: 89 on Monday 12, 81 on Tuesday 13 and 67 on Wednesday 14 October, which fits a change reaching more ChatGPT users each day. On Thursday 15 October, with the tags back from 14:00 UTC, it rose to 86.
The other assistants did not fall
The same tracker, on the same sites, over the same days, saw no fall from Gemini, Perplexity or Claude. If visit tracking had broken, every source would have dropped together. Their numbers are small, but they point the same way.
Of the 184 sites that averaged at least one ChatGPT visit a day from 24 September to 8 October 2026, 153 fell by more than a fifth over 9 to 14 October and 11 rose.
All tagged visits on Wednesday 14 October 2026 against Wednesday 7 October, the same weekday a week apart: 1,094 to 527. ChatGPT's reads on the same two days: 7,792 to 7,845.
Computers fell −38% over the same days, 24 September to 8 October against 9 to 14 October 2026, so this is not one app or one browser. Device is read from the visitor's user agent.
- Mark 9 to 15 October 2026 in your analytics as a gap in ChatGPT's tagging. Your ChatGPT line dips over those days, and the missing visits most likely sit in direct traffic. Read it as a fault in the label, not a change in demand.
- The fix did not reach back. A visit that arrived untagged was recorded as direct when it happened and stays that way, so leave 9 to 15 October out when you compare October with September or November.
- Count ChatGPT by its tag, utm_source=chatgpt.com, not by the referrer. Since the fix, 99.8% of ChatGPT visits on these sites carried the tag and 2.5% a chatgpt.com referrer.
- The tag never changed form. utm_source=chatgpt.com was on 99.1% of ChatGPT visits from 24 September to 8 October 2026, on 99.2% of those that still arrived tagged from 9 to 14 October, and on 99.8% since the fix. No new tag value appeared.
- The referrer did not come back. A chatgpt.com referrer was on 9.3% of ChatGPT visits from 24 September to 8 October 2026, 6.2% from 9 to 14 October, and 2.5% from 14:00 UTC on 15 October to 15:59 on 16 October.
- Our logs cannot see direct traffic, so on their own they could not separate fewer clicks from unlabelled clicks. The return mostly settles it: tagged visits went back to their usual level within the hour, which a missing label explains and a change in how people use ChatGPT does not.
One row per person arriving from an assistant, from the small business websites' own referral logs. A ChatGPT visit carries utm_source=chatgpt.com or a chatgpt.com referrer. Gemini, Perplexity and Claude are matched on their own domains.
ChatGPT-User, the agent ChatGPT sends to fetch a page when someone asks it a question, counted from the same sites' server logs. Our tracker also writes one ChatGPT-User line for every ChatGPT visit; those are removed, or a fall in visits would drag the reads down with it.
24 September to 8 October 2026 against 9 to 14 October 2026, in UTC days. Daily averages, so the windows' different lengths do not matter.
Hourly ChatGPT-tagged visits on 15 and 16 October 2026 against the same UTC hours averaged over 30 September to 8 October, the nine days before the break, and over 9 to 14 October. Re-counted on 16 October, the 9 to 14 October figures had not changed.
A fixed set: the 184 sites that averaged at least one ChatGPT visit a day before 9 October 2026, each compared with itself. The totals cover every site in the logs.
Resolved. First published 15 October 2026, updated 16 October 2026 after the tags came back.