ChatGPT-User Loads Far Fewer Web Pages — Why That Does Not Mean Lower AI Visibility

In May, we reported that ChatGPT loaded our pages roughly 350,000 times in a single month, with a clearly rising tendency. That trend ended abruptly in August. The following table shows the page loads by ChatGPT-User together with a second quantity: the number of human visitors arriving from ChatGPT, i.e., people who clicked on a link to one of our pages in a ChatGPT answer, which brought them to our website. We recognize these by the URL marker utm_source=chatgpt.com, which ChatGPT attaches to such links, and count them with the same strict filtering against bots as for our other human traffic.

The two columns tell different stories. The page loads of September are at about a fifth of the level of the first half of the year. The visitors arriving from ChatGPT show the usual summer dip in July and August, and in September they are back at the level of spring.

The only explanation consistent with both columns is that ChatGPT continues to use our pages as before, but obtains them differently. Apparently, it now answers mostly from page copies stored in its own search index, instead of loading a page anew for each user question. That index is filled by a separate crawler (OAI-SearchBot), which is indeed increasingly active on our site.

Although OpenAI has not announced that change, we are confident in our interpretation of the data. It fits well with an analysis by Botify of several billion log entries of many websites, which found declining ChatGPT-User activity already in early 2026, while OpenAI's crawlers were more active than ever. It also makes sense technically: Answering from stored copies is faster and cheaper than fetching several pages for each question.

The uncomfortable consequence is that web server logs can no longer tell how intensively ChatGPT uses a website. When an answer is produced from a stored copy, no signal arrives at our server.

Some useful indicators remain:

• Visitors arriving from ChatGPT, identified by utm_source=chatgpt.com in the requested address. They are only a small fraction of the users who saw the content — in the first half of 2026, we registered around 100 to 250 page loads by ChatGPT-User for each such visitor — but their trend over the months is meaningful: confirming that ChatGPT continues using our pages on a similar level as before.
• Citations in Microsoft Copilot, as reported in the AI Performance report of Bing Webmaster Tools. These count how often a page was actually shown to a user as a source. For our website, they have risen from about 120,000 per month at the end of 2025 to over 300,000 in September 2026.

A practical point for your own website: Check that the robots.txt file on your web server admits OAI-SearchBot, not only ChatGPT-User. With answers coming from the search index, a page which that crawler may not read cannot appear in them.

For customers with our Advertising Package, nothing has changed in substance: The product descriptions published on our pages continue to reach users of AI systems — including the many who never visit our website. What has changed is that part of this use has become invisible in the page load numbers on the advertiser statistics pages, where we have added a corresponding explanation.

The Copilot citations are still rising, further confirming the popularity of our site with AI systems.

As always, we at RP Photonics analyze our data carefully and show the results openly, including the less convenient ones. (We are not aware of any other digital marketing offers in photonics where transparency is so high.) You find the monthly numbers on our public statistics page, and each advertiser has their own statistics page with specific data as relevant for them.