Research institute · Pew Research Center
35% of Recent Web Pages Show Signs of AI Writing
Pew Research Center ran 490,000 web pages through an AI detection model and found signs of AI authorship on 10% of the English-language web. Among pages published since ChatGPT launched, the figure is 35%.
Source published Summarised
Key findings at a glance
- 10% of all English-language pages sampled in the July 2026 crawl showed meaningful signs of AI authorship or editing.
- 35% of July 2026 pages carrying a publication date after ChatGPT's release showed the same signals.
- 490,000 English-language web pages analysed, drawn from 49 Common Crawl snapshots between January 2021 and July 2026.
- 9.35% of .com pages in the 2026 samples showed signs of AI authorship, the highest of the four domains measured.
- 4.59% of .org pages showed AI signals, roughly half the .com rate.
- 1.03% of .edu pages showed AI signals, about a tenth of the commercial web's rate.
- 0.76% of .gov pages showed AI signals, the lowest reading of any top-level domain in the study.
- 1.08% was the .com rate in January 2022, before ChatGPT existed, when all four domains sat at a similar low level.
- 11.19 em dashes per 10,000 words in the January 2026 sample, up from 5.79 three years earlier.
- 63% increase in Oxford comma usage per 10,000 words between the 2023 and 2026 samples.
- 26.02 uses per 10,000 words of AI-favoured vocabulary such as delve, tapestry and testament, more than double the 2023 rate.
- 96% agreement between the open detection model and Pangram's commercial model across 62,370 validation pages.
What the report covers
Pew Research Center's Data Labs team set out to measure something that has been asserted loosely since late 2022 and rarely quantified: how much of the web is now written by machines. The team randomly sampled 10,000 English-language pages from each of the 49 Common Crawl snapshots taken between January 2021 and July 2026, giving 490,000 pages in total, then scored every page with an open-weight AI detection model.
The headline number is 10%. That is the share of all pages in the July 2026 sample showing meaningful signs of AI authorship or editing. It sounds modest until you remember that most of the web is old, and a large share of any random sample could not have been AI-written at all. Filter down to pages carrying a publication date after ChatGPT's release on 30 November 2022 and the figure rises to 35%.
The breakdown by top-level domain is where the study gets interesting. Around one in ten .com pages in the 2026 samples carry AI signals, against 4.59% on .org and roughly 1% on both .edu and .gov. Before ChatGPT existed all four sat at a similar low level near 1%. The shift has happened almost entirely in commercial publishing, not in academic or government material.
What this means for UK marketers
For UK marketers the useful finding is not that AI writes a lot of the web. It is where. Commercial .com publishing has moved to roughly one page in ten, while .gov and .edu have barely moved. That splits the web into two tiers: a fast-growing layer of machine-assisted commercial content, and a slower layer of primary sources that both AI systems and human readers still treat as authoritative.
The practical consequence is that competent prose about a topic is no longer evidence of expertise, because it now costs almost nothing to produce. What stays scarce is the input a language model cannot generate: your own customer data, your own pricing, named people with a track record, and original measurement. UK B2B and ecommerce brands that publish first-party numbers stand out by arithmetic alone, and that is the argument for treating research as an SEO asset rather than a nice-to-have.
It also reframes the objective. When the answer layer is assembled from web text, the goal shifts from ranking a page to being the source worth citing, which is what our generative engine optimisation work targets and the same logic behind the content formats that hold up in AI search.
One warning. Pew's own methodology stresses that page-level detection is probabilistic. The two models it compared agreed 96% of the time but returned a Cohen's kappa of 0.61, which is decent agreement and well short of interchangeable. Do not run an AI detector over a freelancer's copy or a competitor's landing page and treat the score as a verdict. These tools are dependable in aggregate across hundreds of thousands of documents and unreliable on any single one.
Finally, read the stylistic tells as an editorial checklist rather than a scandal. Em dash frequency has roughly doubled, Oxford commas are up 63%, and words such as delve, landscape and testament have more than doubled in use. None of these prove anything individually. But copy carrying all of them reads as generic to an experienced reader whatever produced it, which is why we banned em dashes from Peak Digital copy long before this study quantified it. The UK context is not distant either: ONS tracking of AI adoption in UK businesses shows the tooling is already in widespread commercial use, and our earlier report on how much of an average job AI actually touches covers where that effort is landing.
How they got the data
Pew sampled 10,000 English-language pages at random from each of the 49 Common Crawl snapshots between January 2021 and July 2026, collecting both the raw HTML and the extracted body text. Only publicly accessible pages appear in Common Crawl, so paywalled and login-gated sites are underrepresented. Roughly 10% to 15% of sampled pages carry a publication date in their HTML, and the 35% post-ChatGPT figure is drawn from that dated subset rather than from the web as a whole, a limitation Pew states plainly.
Detection used an open-weight model from Pangram that scores text from 0 (fully human) to 1 (fully AI-generated). Pages scoring 0.2 or above were counted as showing meaningful signs of AI authorship or editing. To test that threshold, Pew re-ran 62,370 pages from seven crawls through Pangram's commercial model and found 96% agreement. Pew also flags that the open model appears to over-report on pre-ChatGPT pages, so the roughly 1% readings from 2021 and 2022 should be treated as a false-positive floor rather than as early evidence of AI writing. An independent analysis of Internet Archive data, cited in Pew's methodology, reached the same 35% figure by a different route.
Research institute
Read the full report
From Pew Research Center · published 20/08/2026