35% of post-ChatGPT webpages show AI authorship, Pew finds

Over one-third of webpages published since ChatGPT launched in November 2022 show significant signs of AI authorship, according to a Pew Research study released August 20. The finding lands just days after Cloudflare reported that bot traffic had overtaken human traffic on the web, a milestone the infrastructure provider had not expected so soon.

Pew analyzed nearly half a million English-language webpages from the Common Crawl archive, spanning roughly five years. It then ran the sample through Open Pangram's detection technology to flag pages likely written or heavily edited by AI. In a random sample of 10,000 pages collected in July 2026, about 10% showed clear AI authorship. But that figure includes pages published before ChatGPT existed.

When Pew filtered out older pages and focused only on those published after ChatGPT's release, the rate jumped to 35%. The implication: much of the web's new content is now either generated or substantially shaped by large language models.

Which domains show the most AI-written content?

Domain type makes a dramatic difference. Pages on .com domains showed AI authorship at roughly 10 times the rate of .edu or .gov domains, which both hovered around 1%. Nonprofit .org sites came in at 4.6%.

The gap is not surprising. Commercial sites have the clearest incentive to scale content production cheaply, while academic and government pages face tighter editorial oversight and, often, disclosure requirements. For decision-makers relying on web research, the takeaway is blunt: treat .com sources with extra scrutiny when the writing feels generic or the citations are thin.

How reliable is AI detection at scale?

Pew acknowledges the limits. Pangram, like every other AI-detection tool, can misclassify human-written text as AI-generated, and vice versa. Individual page verdicts are noisy. At scale, though, the data is "at least directionally correct," the researchers say.

The study also tracked stylistic markers that have become associated with AI output: increased use of em dashes, Oxford commas, and constructions like "it's not X, it's Y." All have grown more common in post-ChatGPT pages. Those patterns are not proof of AI authorship on their own, but their collective rise aligns with the detection numbers.

Bots writing for bots

The Cloudflare traffic report adds a grimmer layer. If bots now account for more than half of web traffic, and more than a third of new pages are AI-written, a growing share of the web is machines talking to machines. For search engines, that creates a feedback loop: LLMs train on web text, then generate new text that future models will ingest.

Google and Bing have both tweaked ranking signals to demote thin AI content, but the volume keeps climbing. Publishers using AI writing tools like Jasper, Copy.ai, or Writesonic face a strategic choice: lean into speed and scale, or invest in human editing that clears the detection threshold and, more importantly, adds genuine insight.

What this means for content strategy

If a third of competitor pages are AI-generated, the bar for standing out is both lower and higher. Lower because much AI content is generic. Higher because readers and algorithms are learning to spot it. The winning play is not to avoid AI tools entirely but to use them for speed on drafts while investing editorial time in original reporting, data, and perspective that machines cannot replicate.

Source: TechCrunch / Sarah Perez