If you read ChatGPT-User hits in your server logs as a count of people asking ChatGPT about you, one site just gave you a reason to doubt that.
The consultancy Evil Martians re-measured its own logs for September and October 2026. In its published study, ChatGPT-User, the fetcher that pulls a page during a conversation, made 16,445 requests, down 92 percent from May to July. The company says it changed nothing on its side. Claude Code, a coding agent, became the largest AI reader of the site with 36,691 requests.
My read is that a fetch count cannot tell you whether people want your pages or whether ChatGPT changed how it gets them.
Evil Martians tracks AI traffic on the server, before any JavaScript runs. It sorts that traffic into fetchers that act on a person's request and crawlers that fill an index, and it counted only the fetchers. In those two months they made about 62,000 requests, against 84,000 human pageviews.
The 92 percent compares two periods of unequal length, about three months against two. In May to July, the study says, ChatGPT-User made 196,973 requests, 90 percent of everything agents fetched for a person. The study also says its logs cannot tell whether OpenAI changed how often ChatGPT fetches live pages or whether people asked about the company less.
Who read the site instead
Claude Code made 36,691 requests in September and October. Claude Code asked for Markdown 88 percent of the time. ChatGPT-User read rendered HTML on 16,326 of its 16,445 requests. DuckDuckGo's DuckAssistBot came third with 3,179, then DeepSeek with 483 and Exa with 456.

The ranking changed, though the study does not say Claude Code made up for the lost ChatGPT-User traffic.
Claude Code accounts for 94 percent of the site's Markdown reads, and the other named agents asked for Markdown 6 percent of the time. That is one client, not a trend among assistants.
A second set of logs
SEO consultant Carl Hendy wrote on X on October 23 that verified ChatGPT-User requests had fallen by about 80 percent since mid-August in a far larger log set, while verified OAI-SearchBot requests more than doubled. By verified he means requests he matched to the bot. He reads the pattern as a shift toward crawling content in advance instead of fetching pages during conversations. He has not published his sample, and OpenAI has not said anything I could find. His explanation is a hypothesis.
OAI-SearchBot is the crawler OpenAI uses for search results. If ChatGPT answers more from an index it built earlier, fewer ChatGPT-User hits would not mean fewer people were asking about your pages.
Why llms.txt is not the story
The study moved llms.txt from the top tip to near the bottom of Evil Martians' guide. Of 593 direct fetches of the two files, 47 came from named AI assistants, and 43 of those were Claude Code. The study cites two 2026 log studies, including one by Ahrefs across 137,000 domains, that each traced only 1.1 percent of requests for the file to AI bots. The fetcher result is the newer part.
Microsoft announced an Agent Activity dashboard in Clarity on November 4, for agents from Google, OpenAI and Anthropic operators. It is a useful addition to raw logs, and I would still look at the logs. Disclosure: my agency, ViewEngine, sells AI visibility work and builds its own tracking tools.
My read
A user-triggered fetch is evidence that ChatGPT pulled your page. It is not evidence that the page entered an answer, and it says nothing about whether anyone recommended you or searched for your brand afterward. When the count for one agent falls 92 percent on a site that did not change, I blame how the vendor retrieves pages before I blame interest.
I would watch ChatGPT-User and OAI-SearchBot week by week next to ChatGPT referral sessions, branded search and the answers themselves. If the fetches fall while those hold, the fetch count was never measuring demand.
Free newsletter
The Sherpa Brief
One email a week on what changed in AI search, what it means, and what to do about it. Written for marketers, SEOs, and publishers.
Free. Unsubscribe in one click.



