Pew Research just quantified how much of the web is now written by machines, and the number is large enough that it should change how you read almost anything online.
The study used the Common Crawl web archive to collect nearly half a million English-language web pages from the past five years, starting a couple of years before ChatGPT’s release. Pew then ran the text through Open Pangram, an AI detection tool, to see how many pages were likely written or substantially edited by AI.
In a random sample of 10,000 pages collected in July 2026, around 10% showed significant signs of AI authorship. That is the whole-web number, but it includes older pages that predate AI writing tools entirely, so it understates what is happening on new content. When Pew filtered to pages published after ChatGPT’s November 2022 release, signs of AI authorship appeared in over one-third, 35%, of them.
The domain breakdown is where it gets stark. URLs with a .com domain showed signs of AI authorship at roughly ten times the rate of .edu or .gov domains, both of which were around 1%. .org domains came in at 4.6%. The commercial web is the most AI-written corner of the internet, and the institutional and academic corners are nearly untouched.
Pew also found that the supposed tells of AI authorship have increased over the years: the use of em dashes, Oxford commas, and phrasing patterns like “it’s not X, it’s Y.” Those are the same stylistic markers that AI detection tools flag, which means the web is not just being written by AI, it is being written in a style that is measurably more machine-like than before.
The study corroborates other work on the same question, and it lands right after Cloudflare reported that bot web traffic had overtaken human web traffic, a milestone reached sooner than the company had estimated. Pew’s data is focused not on who is browsing the web but on what is being browsed, and the picture is a lot of bots reading pages written by other bots.
There is a caveat that should be stated plainly. Pangram, like other AI detection tools, can misclassify pages as AI-written when they were not. At the scale of nearly half a million pages, the data is likely at least directionally correct, but the 35% figure is an estimate with real error in it, not a precise count.
The durable point is not the exact percentage. It is that the default assumption about new web content has flipped. A decade ago, the reasonable assumption was that a page was written by a person. Now, for the commercial web, the reasonable prior is that new content might well be machine-generated, and the burden of proof has quietly moved. For anyone who builds on the web, including people who write about AI for a living, that shift is the story.
Sources: Pew Research, TechCrunch, The Next Web