A Pew Research study suggests generative AI has become a substantial contributor to the public web. Among pages published after ChatGPT’s November 2022 release, 35 percent showed significant signs of AI authorship or heavy AI editing, according to the analysis reported by TechCrunch.
That headline figure does not mean one-third of the entire web was written by machines. It describes a filtered group of newer English-language pages and rests on an automated detector that can make mistakes. Even with those qualifications, the research offers a large-scale view of how quickly synthetic writing has entered online publishing.
A half-million-page view of the changing web
Pew used the Common Crawl web archive to assemble nearly half a million English-language pages spanning roughly five years. The period began before ChatGPT appeared, giving the researchers material from both before and after widely accessible generative writing tools became available.
The researchers then applied detection technology from Open Pangram to estimate which pages were likely written or substantially edited by AI. In a random sample of 10,000 pages collected in July 2026, approximately 10 percent displayed significant signs of AI authorship.
That broad snapshot mixes recent pages with older material published before modern AI writing services existed. Pew therefore filtered the collection to pages posted after ChatGPT’s launch. Within that newer subset, the detected share rose to 35 percent.
The distinction between the two figures is essential. Ten percent describes a mixed-age snapshot of web pages, while 35 percent applies only after older pages were removed. Neither number counts authors directly, nor does it establish how much of any individual page came from an AI system. The detector identifies patterns consistent with machine generation or substantial editing.
Commercial domains show the strongest signal
The study also found large differences between domain types. Pages on .com domains showed AI-authorship signals at around ten times the rate recorded for .edu and .gov domains. Both educational and government domains were near 1 percent, while .org pages registered 4.6 percent.
Those results describe correlation by top-level domain, not the motives behind publication. The source does not establish whether commercial publishers use AI more often because they produce larger volumes, face different economic incentives or publish different kinds of material. It also does not show that every detected use was fully automated; substantial AI editing falls within the study’s measure.
Pew observed that several stylistic features commonly associated with generative systems have also become more frequent. These included em dashes, Oxford commas and contrast constructions built around the pattern “it is not X, it is Y.” Such changes may support a broad trend analysis, but none is proof that a particular writer used AI. Human authors used all of those forms long before ChatGPT.
The report arrives as automated activity is reshaping both sides of the web. TechCrunch notes a separate Cloudflare report saying bot traffic has overtaken human traffic. Pew’s work focuses on content rather than visitors, but together the findings raise the prospect of automated systems increasingly retrieving pages that other automated systems helped create.
Detection limits make the percentage an estimate
Pew explicitly acknowledged that Pangram, like other AI detectors, can classify human writing as machine-generated. A large dataset can still reveal directional changes even when individual classifications are imperfect, but the reported 35 percent should be treated as an estimate rather than a census.
The methodology also covers English-language pages available through Common Crawl. It therefore cannot automatically represent other languages, content excluded from the archive or private platforms. Publication date filtering improves the comparison with the post-ChatGPT period, yet it does not reveal which model was used or whether a human meaningfully revised the result.
What the study does establish is the scale at which machine-like writing patterns now appear in newer web content. The strongest conclusion is not that authors have disappeared. It is that AI assistance has become common enough to be measurable across hundreds of thousands of pages, while the tools used to measure it remain probabilistic and vulnerable to error.
