About 24% of crawlable web pages contained AI-generated or AI-edited text in August 2026, according to a paper posted on arXiv on 27 October by researchers led from UC Berkeley and the University of Washington. The paper says the share “rose from near zero before 2022 to nearly 25% by August 2026”.
The rise is uneven. The paper estimates that 29.5% of .com pages contained AI-generated text, against 5.8% of .gov pages; by our arithmetic, that is about 5.1 times the rate.
Short explainer · AI text on the web
A vertical news Short narrated from this piece’s headline and opening line.
Business Pill · Data Flywheel
An evergreen Business Engineer explainer on data flywheels. It explains the general idea only and does not discuss this paper.
The key insight: A paper led from UC Berkeley and the University of Washington estimates that about 24% of crawlable web pages held AI-generated or AI-edited text by August 2026; by our arithmetic, the .com rate of 29.5% is about 5.1 times the .gov rate of 5.8%.
The authors say they analysed 5,118,347 eligible pages from 505,974 domains across 71 Common Crawl snapshots spanning January 2019 to August 2026, in every language. They scored pages with more than one AI detector and used a Bayesian model that corrects for each detector’s false positives and false negatives.
The paper puts the August 2026 estimate at about 24 percent, with a 95% credible interval of 22.9 to 24.4. It says the yearly rate of increase “has nearly doubled, from 4.6 percentage points per year between early 2023 and early 2025” to 8.1 points per year since.
Estimated share of crawlable pages with AI-generated or AI-edited text in August 2026, by domain type: .com 29.5%, .net 20.4%, .org 14.6%, .edu 6.1%, .gov 5.8%; about 24% across all pages. The about 5x ratio is our arithmetic. Source: Visokay, Pan et al., arXiv 2610.36631, 27 Oct 2026.
Where the AI Text Is
The growth is concentrated in small, little-linked sites. The paper says the two least-central tiers of the web, which hold 70.5% of eligible pages, “account for 78.8% of the net increase”. Examples it gives of domains in those tiers include “a Swedish rabbit breeder” and a pizza restaurant.
At the other end, the paper estimates 11.6% of pages in the most-central tier contained AI text, against 27.5% in the least-central tier. It names WordPress.org, Wix, Vimeo, Mozilla, the IRS, Harvard and the Financial Times among the most central domains.
Language matters too. The paper estimates that 32.8% of English-only pages contained AI-generated or AI-edited text in August 2026, against 17.0% of pages with no English detected.
Rohan Paul’s post on X, above, summarises the paper’s headline figures; it had about 2,300 views when we read it on 30 October.
Why It Matters for Models and Publishers
The authors note that “the crawlable web is a main input for training future language models”, so “earlier language models increasingly shape the conventions, style, and norms of newer ones”. Our recent coverage of Reddit’s falling search traffic and its user warning tracks how the same web is changing for publishers.
The Structural Read
As we read the paper, the growth is a long-tail story: it says the two least-central tiers of the web hold 70.5% of eligible pages and account for 78.8% of the net increase.
As we read it, the most-linked sites changed least: the paper estimates 11.6% in the most-central tier against 27.5% in the least-central one.
As we read the authors, the stakes run past readers: they write that the crawlable web is a main input for training future language models.
Visokay, Pan et al., arXiv 2610.36631, 27 October 2026
“earlier language models increasingly shape the conventions, style, and norms of newer ones”
Three Implications
FOR MODEL BUILDERS The authors write that the crawlable web is a main input for training future language models, and estimate about 24% of its pages now hold AI text.
FOR PUBLISHERS AND BRANDS The paper estimates 29.5% of .com pages held AI-generated or AI-edited text in August 2026, against 6.1% of .edu and 5.8% of .gov pages.
FOR SEARCH AND DISCOVERY The paper estimates that pages on the most central domains are the least likely to contain fully AI-generated or AI-edited text.
The framework describes “The Corpus Moat โ built by pretraining, eroding as the public web is exhausted and synthetic data closes gaps.”
As we read it, the paper puts a number on that public web: about 24% of crawlable pages held AI-generated or AI-edited text by August 2026, by the authors’ estimate. The framework was published in July 2026 and does not discuss this paper.
What Is Not Established
The figures are model estimates built on AI detectors, which the authors treat as imperfect; the paper corrects for detector error but does not observe how each page was written. The study covers crawlable pages in Common Crawl, not the whole web or pages behind logins. The 5.1 times ratio is our arithmetic from the paper’s .com and .gov estimates.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
A paper led from UC Berkeley and the University of Washington estimates that about 24% of crawlable web pages contained AI-generated or AI-edited text by August 2026, with the growth concentrated on small, little-linked sites and on .com pages.
A note on sourcing. Figures and quotes come from the paper “The uneven rise of AI-generated web content” (arXiv 2610.36631, 27 October 2026), read in full on 30 October; it is a preprint. The 5.1 times ratio is our arithmetic. We did not contact the authors.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
