A new study released Thursday by Pew Research indicates that more than a third of web pages published since ChatGPT’s launch exhibit characteristics of AI authorship. This report corroborates other research detailing the extent to which newer web content is now either generated or “substantially edited” by AI, as stated by the firm.
This finding emerges shortly after internet infrastructure provider Cloudflare reported that bot web traffic had surpassed human web traffic—a benchmark achieved earlier than the company had anticipated.
Pew’s data, however, focuses not on who or what is accessing the web, but on the nature of the content being viewed. It appears that a significant portion of this content consists of AI-generated pages being read by other bots.
To compile their report, Pew stated they used the Common Crawl web archive to gather nearly half a million English-language web pages from the past five years, starting a couple of years before ChatGPT’s November 2022 release. Pew then employed Open Pangram’s technology to identify how many were likely written or substantially edited by AI.
From a random sample of 10,000 web pages collected in July 2026, approximately 10% displayed “significant signs of AI authorship,” Pew noted.
However, Pew highlighted that such a random sample would invariably include older web pages—those published before AI writing tools existed, and thus could not have been AI-generated.
To gain a clearer understanding of the current prevalence of AI-written content on the web, Pew filtered out these older pages, focusing exclusively on those published after ChatGPT’s release.
Within this refined dataset, with older pages removed, signs of AI authorship were present in over one-third (35%) of the content.
Delving into specific domains, Pew discovered that URLs with a .com domain exhibited signs of AI authorship at roughly ten times the rate of .edu or .gov domains (both of which showed about 1% AI authorship). Additionally, .org domains had an AI authorship rate of only 4.6%.
Naturally, this analysis is not infallible. Pangram, like other AI-detection tools, can misclassify pages as AI-written when they are not. Nevertheless, when applied at scale, the data likely reflects a generally correct trend.
Furthermore, Pew observed an increase over the years in other potential indicators of AI authorship, such as the use of em dashes, Oxford commas, and specific phrasings like “it’s not X, it’s Y,” among other stylistic elements.