Pew Research Center ran nearly half a million webpages through an AI detector and found signs of AI authorship or editing on about 10% of them. Among pages published after ChatGPT’s launch specifically, that share jumps to 35%. Pew’s Data Labs team published the analysis on August 20, less than a year after two other major estimates of AI-generated text on the open web, and the numbers add hard data to a trend already visible in how AI-generated content keeps recirculating across the web.
What the data shows
The clearest pattern in Pew’s data is where AI text concentrates. Pages on .com domains show signs of AI authorship at roughly ten times the rate of pages on .edu and .gov domains. In a six-month average, the detection rate ran 9.35% on .com, 4.59% on .org, 1.03% on .edu, and just 0.76% on .gov.
All four domain types sat at or below 1% in the sample Pew collected before ChatGPT launched. They’ve separated sharply since then, with .com climbing the fastest while .edu and .gov have stayed close to that original 1% baseline.
The stylistic tells are spreading
Pew also tracked specific writing markers commonly associated with AI-generated text in pages published after ChatGPT’s release. Em dash usage rose from 5.79 per 10,000 words in early 2023 to 11.19 in early 2026. Oxford comma usage rose 63% over the same period. Words common in AI writing — “delve,” “interplay,” “testament” — more than doubled in frequency. The “it’s not just X, it’s Y” negative-parallelism construction rose from 0.87 to 2.36 uses per 10,000 pages, though it remains a comparatively rare pattern overall.
Pew is careful to note that none of these markers can identify a single document as AI-written on their own, since human writers use all of them too — the claim is about rates across large sets of text, not about any individual page. A separate analysis of Ahrefs’ detector data published in July raised a related open question: how useful a detector score actually is as AI-assisted editing becomes a normal part of everyday writing tools, rather than a marker of wholly AI-generated content. Pew’s data shows these stylistic markers appearing more often over time; whether that makes any single marker more or less reliable for flagging AI involvement is a separate question Pew’s report doesn’t attempt to answer.
Other estimates disagree — but land in a similar range
Pew’s numbers aren’t the only ones in circulation, and they don’t all agree. SEO firm Graphite estimated that 49.9% of newly published English-language articles in the first quarter of 2026 were primarily AI-generated — roughly five times Pew’s overall detection rate, though Pew’s figure spans a broader domain mix, not just newly published articles. A preprint from Imperial College London, the Internet Archive, and Stanford (not yet peer-reviewed) found that by mid-2025, 35% of newly published websites were detected as AI-generated or AI-assisted — which lines up closely with Pew’s 35% figure for post-ChatGPT pages specifically, even though the two studies measure somewhat different things. All of these estimates lean on the detector Pangram in some form; Graphite also incorporated Copyleaks and GPTZero.
Why this matters for SEO work
The concentration pattern is the part worth sitting with: signs of AI text are heaviest on the commercial web, and that’s precisely the part of the internet most SEO and content work touches. Pew’s .com detection rate has climbed in every reading taken since ChatGPT launched, while .edu and .gov have stayed under 2% throughout. That gap is a reasonable proxy for where content quality pressure is actually building — a dynamic that lines up with reporting that Google may already be treating some AI-generated content as thin content.
It’s also worth being precise about what Pew’s threshold captures. It flags AI editing as well as AI authorship, so a page a person wrote and then ran through a cleanup tool lands in the same bucket as a page an AI produced start to finish. That distinction matters for interpreting the headline numbers, and it’s part of why AI-generated content isn’t really the problem on its own — the strategy behind it is.
Looking ahead
AI-assisted editing is now a native feature in both Google Docs and Microsoft Word, which means the line between “AI-assisted” and “AI-generated” is only going to blur further. Pew’s threshold already folds AI editing into its count, but none of the three studies discussed here distinguish lightly AI-edited human writing from text generated by AI from scratch — a gap that matters more every quarter these tools get more deeply embedded in ordinary writing workflows, and one that’s closely tied to why platforms are already struggling to separate genuinely low-quality AI slop from AI-assisted content that’s actually fine.
None of the detection percentages settle the question that actually matters for readers, or for Google: whether a given page is accurate, useful, and worth publishing. That’s still decided the same way it always has been — by reading it.