New research published by the Pew Research Centre shows that since ChatGPT came online in November 2022, about 35 per cent of the new English-language Internet page has been marked with AI writing or heavy editing. Generating AI ' s impact on content production is expanding from individual scenarios to a wider web publishing process.

The sample covers nearly 500,000 pages.

Using the Common Crawl web file, the study collected approximately 500,000 pages in English over the past five years, covering the period before and after ChatGPT's release. Subsequently, researchers used Open Pangram's testing techniques to identify whether the web page could be written by AI or substantially modified by AI.

Of the 10,000 random samples taken in July 2026, about 10% showed clear signs of an AI author. According to Pew, this caliber is mixed into more old pages, and when these pages are published, the relevant AI writing tools are not widely available and therefore the overall ratio is lowered.

Up to 35% after removing old pages

In order to observe more accurately the impact of the generated AI on new content, Pyo further removed the pages before ChatGPT was launched, and only the pages that were on line since then were counted. Within this range, the proportion of pages with AI writing marks rose to 35 per cent.

  • Random sample of 10,000 pages
  • The total sample is close to 500,000 English pages.
  • 35% of new pages with AI marks

This means that more than one third of the pages published after the generation of AI may have been written by AI. The study does not define these pages in their entirety as pure AI generation and also includes material that is substantially edited by AI.

Faster use of commercial domain names

When split by domain name type, the ratio of the.com domain name web page showing an AI writing sign is about 10 times that of.edu and.gov domain names. The latter share is around 1 per cent, while.org domain names are about 4.6 per cent.

The results show that commercial websites generate or edit content at a much faster pace than those of educational, governmental and non-profit institutions. There was no further industry-specific breakdown in the study, but differences in domain names already reflect the fragmentation of content production methods across different sites.

There are still errors in detection.

Pew also mentioned that such analysis was not perfect. Like other AI testing tools, Open Pangram may misjudge content not written by AI as AI generation. However, the study found that overall trends were still of reference value under large sample conditions.

In addition to the direct test results, studies have found that some expressions that are often seen as AI writing features have also increased in recent years, such as dashes, Oxford commas and the phrase “not X, but Y”. These changes in language habits also reflect side by side that AI is affecting the style of Internet text.