One in three web pages published since ChatGPT launched now shows signs of AI authorship

Key points

• Pew Research found that 35% of English-language web pages published after ChatGPT's November 2022 launch show signs of AI authorship, in a report released Thursday.
• A random snapshot of 10,000 pages collected in July 2026 showed around 10% with significant AI-writing signals, rising to 35% once pre-ChatGPT pages were removed.
• Commercial .com domains showed AI-authorship signals at roughly ten times the rate of .edu and .gov domains, both of which sat near 1%.
• Cloudflare reported that automated bot traffic has overtaken human browsing, a milestone the company said arrived sooner than it had expected.
• AI detection tools can misidentify human writing as AI-generated, so individual results carry real uncertainty, though the overall trend in the Pew data holds.

Most of us assume that when we load a web page, a person wrote it. That assumption is getting harder to hold.

Pew Research released a study Thursday showing more than one third of English-language web pages published after ChatGPT went public in November 2022 show signs of being written or heavily edited by AI. The study, first reported by TechCrunch AI, analysed nearly half a million pages pulled from the Common Crawl web archive, a freely available collection of billions of web pages. Pew used detection technology from Open Pangram to score each page for AI-writing signals.

A random sample of 10,000 pages collected in July 2026 showed about 10% with significant AI-authorship signs. Filter out pages that predate AI writing tools, ones that could not have been AI-written by definition, and that figure climbs to 35%.

Who is writing, and who is reading?

Bots, increasingly, on both counts. Cloudflare, which handles traffic for a large share of the world's websites, recently reported that automated bot traffic now outnumbers genuine human browsing, and that the crossover came sooner than the company had forecast. On 7 August we covered Cloudflare's Kitesurf browser, built specifically for AI agents rather than people. Pew's study does not cover who visits pages; it looks at what those pages contain. Combine the two findings and a strange picture forms: software browsing pages that software wrote.

Not all corners of the web look alike. Commercial .com domains showed AI-writing signals at around ten times the rate of .edu and .edu sites, both near 1%. Non-profit .org domains came in at 4.6%. AI writing tools lower the cost of producing content at scale, and commercial sites have the strongest financial motive to do exactly that.

Should you worry about AI detection errors?

Partially, but the direction is solid. Detection tools, including Open Pangram's, make mistakes in both directions. A human who favours short, clean sentences might score as AI-generated; a polished AI output could slip through. We looked at this closely on 9 August: AI writing detectors have suspended students and collapsed book deals, based on tools their own makers admit are unreliable. Pew acknowledged the same limits openly. At nearly half a million pages, though, individual errors average out.

Pew also found that classic AI signals grew more common over the study period: frequent em dashes, Oxford commas, and stock phrases like "it's not X, it's Y."

For ordinary readers, none of this means every page you visit was machine-written. But the odds have shifted. A product review, a how-to article, a news summary found via search: any of them may have been assembled by software.

Watch for pages that sound fluent but feel generic, that cover a topic without specific detail or personal experience, or that repeat the same few sentence structures throughout. Those are the fingerprints Pew is now counting at scale.