One Crawlable Page in Four Has AI Text. At the Center of the Web It Is One in Nine.

A paper posted to arXiv late on 26 October, by Adam Visokay and colleagues, puts a number on a question everyone has an impression about: how much of the web is now written by AI. Using “a sample of more than 5 million crawlable webpages spanning 2019 to 2026,” they estimate that “in August 2026 about 24 percent (95% credible interval 22.9-24.4) of crawlable webpages contained AI-generated or AI-edited text,” up from near zero before 2022.

That is the number that will travel. The paper’s more useful finding is how unevenly that quarter is spread, because the average sits in places most readers rarely visit.

The method is the careful part

AI detectors make mistakes in both directions, and the paper builds that in. It uses several detectors and a Bayesian model that corrects for each one’s false positives and false negatives, and lets detector performance differ across groups. The correction is not cosmetic. On English pages from crawls made before capable language models existed, where every flag is by construction a false positive, the open Pangram detector still flagged 4.5% of 1,336,683 pages. Read without correction, a detector that flags 4.5% of pre-2022 pages would inflate every later estimate.

Language is where this matters most. A 2023 study in Patterns found that GPT detectors frequently misclassify non-native English writing as AI-generated. The paper sets its detector thresholds separately by language, and still reports that 55.5% of its non-English pages are in languages that appear in none of its validation sets, so the non-English estimates rest on less calibration than the English ones.

Where the quarter is

The paper ranks pages by how central their domain is in the web’s link graph. The most central tier, “the platforms and infrastructure of the web itself, such as WordPress.org, Wix, Vimeo, Yahoo, Mozilla and the Internet Archive,” made up 0.001% of domains and 2.7% of eligible pages. Its estimated prevalence in August 2026 was 11.6%. The least central tier was at 27.5%. The two least central tiers together hold 70.5% of eligible pages and “account for 78.8% of the net increase.”

AI-generated text by position in the web's link graph, August 2026 Two comparisons. Prevalence: 11.6 percent in the most central tier against 27.5 percent in the least central tier. Growth: the two least central tiers hold 70.5 percent of eligible pages and account for 78.8 percent of the net increase since December 2021. CENTRALITY IN THE LINK GRAPH · AUGUST 2026 · VISOKAY ET AL. Share of pages with AI text Bottom and Tail tiers together Most central tier (Top) Least central tier (Tail) Share of eligible pages Share of the net increase 11.6% (11.2–11.9) — Visokay et al., arXiv, 26 Oct 2026 27.5% (26.5–28.4) — Visokay et al. 70.5% of eligible pages — Visokay et al. 78.8% of the net increase — Visokay et al. 11.6% 27.5% 70.5% 78.8% The Top tier is 0.001% of domains and 2.7% of eligible pages.
The web's center has about one AI-text page in nine. The periphery, where most pages are, has more than one in four and produced most of the growth.

The tier gap is not an artifact of what kind of pages each tier holds. When the authors give every tier the same mix of domain type, language and length, the estimate “still increases monotonically across tiers, from 11.1% in the Top to 26.7% in the Tail.”

Domains and languages

The split by top-level domain is sharper. In August 2026, 29.5% of .com pages carried AI-generated text, against 20.4% of .net, 14.6% of .org, and 6.1% and 5.8% of .edu and .gov. The institutional domains also started later: .edu and .gov were at 2.1% and 2.5% as late as February to March 2024.

Share of pages with AI-generated text by top-level domain, August 2026 Horizontal bars on a 0 to 40 percent scale. .com 29.5 percent, .net 20.4, .org 14.6, .edu 6.1, .gov 5.8. PAGES WITH AI-GENERATED TEXT BY TOP-LEVEL DOMAIN · AUGUST 2026 .com .net .org .edu .gov 29.5% (28.7–30.2) — Visokay et al. 20.4% — Visokay et al. 14.6% — Visokay et al. 6.1% — Visokay et al. 5.8% — Visokay et al. 29.5% 20.4% 14.6% 6.1% 5.8% Scale runs to 40%. The .edu and .gov estimates lean more on the model for rare page types.
Commercial pages carry AI text about five times as often as government and university pages.

Language splits the same way. In August 2026, 32.8% of English-only pages carried AI-generated or AI-edited text, against 17.0% of pages with no English. Holding the mix of tier, domain and length fixed, the figures are 32.5% for English-only pages, 21.4% for bilingual pages and 16.8% for non-English pages, so the gap is not explained by where English pages sit.

AI-generated text by page language, August 2026 Horizontal bars on a 0 to 40 percent scale at a common composition of tier, domain and length: English-only 32.5 percent, bilingual 21.4 percent, non-English-only 16.8 percent. Actual-composition values: English-only 32.8 percent, non-English-only 17.0 percent. AI TEXT BY PAGE LANGUAGE · COMMON COMPOSITION · AUGUST 2026 English only Bilingual No English 32.5% at common composition (32.8% actual) — Visokay et al. 21.4% at common composition — Visokay et al. 16.8% at common composition (17.0% actual) — Visokay et al. 32.5% 21.4% 16.8% Same mix of centrality tier, domain and length for every group. Actual figures: 32.8% and 17.0%. 55.5% of non-English pages are in languages none of the validation sets cover.
English pages carry AI text about twice as often as pages in other languages, even with the page mix held fixed.

What “one in four” describes

I think the honest reading of the headline is narrower than it sounds. One in four is the share of crawlable pages, and most crawlable pages sit at the edge of the link graph, on domains few other sites point to. The pages at the center of the web, which a person or a search engine is more likely to reach, are at about one in nine. The growth is real and fast, with the rate of increase rising from 4.6 percentage points a year between early 2023 and early 2025 to 8.1 since. But most of the growth is at the thin edge of the web. Prevalence is highest on .com pages in English, and far lower on the government and university sites people go to when the answer has to be right.

References

  1. Visokay, A., Pan, S., Griston, M., Iasakova, P., Nepf, M., Peskoff, D., Sinha, R., Sun, J. et al. (2026). The uneven rise of AI-generated web content. arXiv:2610.36631, 26 October 2026.
  2. Liang, W., Yuksekgonul, M., Mao, Y., Wu, E. & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns 4(7), 100779.