More Than One-Third of Post-ChatGPT Web Pages Show AI Traces

Introduction

Generative AI is changing not only how people consume information online, but also how the web is produced. A new Pew Research study found that more than one-third of pages published after ChatGPT’s November 2022 launch show significant signs of being written by AI or heavily edited with AI. The result suggests that a growing share of newly published online material may involve models somewhere in the editorial workflow.

What the study found

• The overall sample produced a lower figure. Pew used the Common Crawl archive to collect nearly half a million English-language pages published over roughly the past five years. In a random sample of 10,000 pages collected in July 2026, around 10% showed significant signs of AI authorship. That sample also included older pages, including pages published before modern AI writing tools were available.
• The post-ChatGPT share reached 35%. After filtering out older pages and focusing on pages published after ChatGPT’s release, Pew found AI authorship signals on more than one-third of pages. The measure does not mean every page was generated entirely by a model. It can also include material written by a person and substantially revised by AI.
• Domain types differed sharply. Pages using .com domains showed AI authorship signals at roughly 10 times the rate found on .edu and .gov pages, both of which were around 1%. The rate for .org pages was 4.6%. The supplied study summary does not establish why these differences exist, but they point to variation in publishing practices across the web.
• Style clues are becoming more common. Pew also tracked features often associated with AI writing, including em dashes, Oxford commas, and constructions such as “it’s not X, it’s Y.” These patterns may be useful as indicators, but none can prove that a specific page was generated by AI.

Why it matters

The study is notable because it measures the content being published, rather than simply counting model users or AI-generated requests. Cloudflare has separately reported that bot web traffic has surpassed human traffic. Put together, the two developments raise the possibility of a web in which automated systems increasingly read material produced with substantial automated assistance.

That does not make the 35% figure a definitive census. Pew relied on Open Pangram’s detection technology, and AI detectors can misclassify human writing. They can also struggle with pages that combine model output with extensive human revision. The researchers therefore present the results as directionally useful, not as page-by-page proof of machine authorship.

For publishers, search companies, and readers, the central issue is likely to be provenance rather than the mere presence of AI. As machine-assisted publishing expands, platforms may need stronger ways to evaluate sourcing, factual support, editorial review, and disclosure. Speed and volume can help fill the web, but trust will increasingly depend on whether publishers can show how content was produced and checked.

Source: TechCrunch AI