In April I asked ChatGPT about a site I run, sometimes by product name and sometimes with a link, and watched the server log. ChatGPT sent ChatGPT-User, the bot it uses to read a page while it writes a reply, often to several pages at once while it decided which to cite. I wrote that up in what nginx logs prove about AI traffic.
This week I asked about my products by name again, and ChatGPT-User barely showed up. Cloudflare’s public data shows the same thing across its network: ChatGPT fetches far fewer pages while it answers than it did three months ago.
The test
I run two production sites, one selling software products and one publishing articles. I tailed the access logs on both and asked ChatGPT ten questions in temporary chats. A temporary chat is unpersonalized: it has ChatGPT’s standard tools, including web search, but none of the apps or connectors I’ve added to my account. I didn’t force a web search; ChatGPT decided when to use one, the way it does for anyone.
They were buyer questions. Some named a product and asked whether it works and what it costs. Others described a problem our products solve without naming anyone. I aimed them at pages whose crawl history I already had. OpenAI’s bots had last requested most of those pages weeks earlier, and one more than three months earlier.
I watched for all three of OpenAI’s bots: ChatGPT-User, OAI-SearchBot, which crawls pages for ChatGPT’s search index, and GPTBot, which collects training data. Every request was checked against the IP ranges OpenAI publishes, so a bot using a fake name couldn’t count.
9 answers, 1 page fetched
Our sites were a source in 9 of the 10 answers. On the questions where I didn’t name a brand, ChatGPT still recommended our products by name and linked to us. The one miss was a broad industry question, which it answered from other sites.
Across those ten answers, ChatGPT fetched one page from us: a homepage, once, from an address inside OpenAI’s published range. It didn’t fetch a single product page, article or pricing page, and neither OAI-SearchBot nor GPTBot showed up. ChatGPT recommended pages it hadn’t read in weeks.
Cloudflare shows the same drop
Two sites on one afternoon could be a fluke, so I checked a much bigger sample. Cloudflare sits in front of a large share of the web and publishes weekly request volumes for each verified bot on Cloudflare Radar.
ChatGPT-User requests peaked in the week of June 29, 2026. By the week of September 14 they were at 30% of that peak, the lowest point in the past twelve months. The previous low, in March, was 53%.
OAI-SearchBot went the other way. It had its busiest week in twelve months in late August.
Here are the first six weeks of the summer (June 1 to July 12) against the two most recent weeks on Radar (September 7 to 20):
ChatGPT is fetching fewer pages while it answers and crawling more in the background. My reading is that it now answers from its search index more often and reads live pages less. OpenAI hasn’t announced a change, so that is an interpretation of the data, not a confirmed fact.
Two other things could add to the drop. More sites may be blocking OpenAI’s bots, and Radar only sees traffic that passes through Cloudflare. GPTBot fell too, so part of the decline may affect every OpenAI bot. But a decline across the board would have pulled OAI-SearchBot down as well, and it went up.
Why ChatGPT would change this
OpenAI’s crawler documentation has always allowed for it. It says ChatGPT-User “may visit a web page” when someone asks a question, which is not the same as always visiting. The same page says ChatGPT-User “is not used to determine whether content may appear in Search.” That job belongs to OAI-SearchBot: “sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers.”
My guess at the motive is cost and speed. Downloading and reading pages during every answer, for hundreds of millions of people, is slow and expensive. Answering from a stored copy is neither. OpenAI hasn’t said so, but it fits both my logs and Cloudflare’s numbers.
The stored copy could come from OpenAI’s own crawl, from the Bing index we covered in which search index each AI assistant reads, or from both. The effect on your site is the same either way. The single homepage fetch in my test may have been a quick check that the site was still up, but that’s also a guess.
So the ChatGPT I tested in April isn’t the ChatGPT I tested this week. These systems change without notice, and any measurement of them is only true for the date it was taken.
What this means for your site
ChatGPT-User hits no longer tell you whether ChatGPT uses you. If your AI visibility report counts ChatGPT-User requests, it’s counting a bot ChatGPT sends less and less. That count can fall while ChatGPT recommends you more often. What still reflects ChatGPT’s use of your site is what its answers say about you, and the people who click through from them, who arrive as normal browser visits with chatgpt.com as the referrer.
ChatGPT may be quoting an old version of your page. If you change a price, a plan or a feature, ChatGPT can keep repeating the old one until its crawler visits again. Updating the page isn’t enough; it has to be recrawled.
The crawl is now the moment that counts. If ChatGPT answers from what it stored, the visit from OAI-SearchBot is when your page actually gets read, so make sure it can get in. Our robots.txt checker shows which AI crawlers your site allows, and this guide covers the firewall and CDN blocks robots.txt won’t show you.
Try it on your own site
Pick a page ChatGPT should know about and watch your log while you ask about it:
tail -F /var/log/nginx/access.log | grep --line-buffered -iE 'chatgpt|oai-searchbot|gptbot'Ask by product name instead of pasting the URL, in a temporary chat so your own apps and connectors stay out of it. On my first two questions I used a normal chat, and ChatGPT pulled its answer from an app connected to my account instead of the web. If a CDN or page cache serves your pages, some requests never reach your server log, so check the CDN’s logs too.
Check the IP address on anything that matches. While I was testing, one Google Cloud address cycled through 12 AI crawler names in 81 seconds, from GPTBot and ChatGPT-User to ClaudeBot and Googlebot. It wasn’t reading pages. It was looking for .env files, SSH keys and Kubernetes tokens. We found the same pattern at scale in fake AI crawlers scanning for API keys.
Limits
The test was one assistant, ten questions, two sites and one afternoon in late September 2026. I didn’t save the source links from each answer, so I can say which pages the questions targeted and that none of them was fetched, but not the exact URL each answer linked. The Cloudflare figures are request volumes relative to each bot’s own peak, not absolute counts, and they only cover traffic through Cloudflare. Claude, Perplexity and Gemini are next.
