Approximately 10% of web pages on the internet show signs of being created or significantly edited by AI. Among materials published after the release of ChatGPT, this figure exceeds one-third, according to researchers at the Pew Research Center.
AI Gradually Fills the Internet
Researchers analyzed a random sample of 10,000 English-language web pages collected in July 2026 via Common Crawl. A machine learning model was used to identify likely authorship by detecting linguistic features typical of generative AI.
About 10% of all sampled pages showed significant signs of AI authorship. While this figure may seem relatively small, the sample included a lot of older content created before the advent of modern generative models.
Source: Pew Research Center.
Excluding older pages and focusing only on materials published after the launch of ChatGPT in November 2022, the picture changes dramatically: signs of AI authorship are found on more than a third of these pages.
Thus, the share of synthetic content is growing alongside the spread of ChatGPT and its competitors, such as Claude, Gemini, and other models.
Most AI Content Found on Commercial Sites
The distribution of AI content across domain zones is uneven.
In 2026, signs of AI generation were found on about one in ten pages in the .com zone. This is roughly double the rate in .org, where the figure was 4.6%.
On sites with .edu and .gov domains, the share was about ten times lower—around 1%. Researchers note that at the beginning of ChatGPT’s spread, differences between major domain zones were much less pronounced.
The difference may reflect the nature of the content itself. Commercial sites more frequently publish large volumes of text materials that can be relatively easily automated: product descriptions, reference articles, SEO content, and news.
Study Does Not Claim Every Such Text is AI-Written
An important caveat concerns the methodology. Pew did not verify the creation history of each page or ask authors if they used ChatGPT or other tools.
Researchers analyzed linguistic features statistically associated with AI generation. Therefore, the result should be interpreted as an estimate of likely authorship, not an exact count of pages written by neural networks.
Moreover, the focus is on content that is “written or significantly edited” by AI. Using a neural network for minor text edits does not necessarily fall into this category.
In August, Anthropic introduced global labeling for content created by Claude models.
