Study Finds 35% of New Websites Are Now AI-Generated

A joint study by researchers from Imperial College London, the Internet Archive, and Stanford University offers the first rigorous empirical measurement of AI-generated text's actual impact on the web. Using representative sampling via the Wayback Machine and benchmarking four detection methods (Binoculars, Desklib, DivEye, Pangram v3) to select the most reliable one, the researchers found that the share of AI-generated or AI-assisted websites rose from zero before ChatGPT's late-2022 launch to roughly 35% by mid-2025.

The team also surveyed 853 US adults about their beliefs regarding six hypothesized negative effects of AI text on the internet. The results reveal a stark gap between perception and evidence: while most respondents believed in all six hypotheses, only two were statistically confirmed. Semantic diversity does shrink as AI content grows (AI-generated sites showed 33% higher textual similarity), and positive sentiment increases sharply (107% higher). However, claims about rising factual errors, declining outbound links, content bloat, and stylistic homogenization found no significant support in the data.

For engineers working on content moderation, SEO systems, AI-detection tools, and digital trust infrastructure, the study is a reminder that public intuition about AI's effects on the web doesn't always align with measurable reality—underscoring the need for evidence-based rather than assumption-driven approaches.