A preprint posted to arXiv on 27 October 2026 estimates that about 24% of the crawlable pages it sampled contained AI-generated or AI-edited text by August 2026. Pew Research Center, working from the same Common Crawl archive, reports 10%. A team using the Internet Archive puts the figure at roughly 35% of newly published websites. A marketing analytics firm says about half of published articles.
The four numbers are not competing answers to one question. Each rests on its own sample, its own detector setup and its own denominator. Only the newest tries to estimate total prevalence across the crawlable web while correcting for detector error.
What the 5m-page crawl counted, and what it left out
“The uneven rise of AI-generated web content” is a preprint, arXiv:2610.36631v1, submitted on 27 October 2026 by thirteen authors from the 2026 Summer Institute in Computational Social Science at the University of Washington. It has not been peer-reviewed, and no independent methodological review had appeared as of 29 October.
The sample is 5,118,347 pages from 505,974 domains, drawn from 71 Common Crawl snapshots between January 2019 and August 2026. A page counts as eligible when the crawler captured it and extracted at least 500 characters of main text; navigation and boilerplate are removed, and pages behind logins or inside apps, or on sites that block crawlers, are absent. Everything the paper reports is a share of eligible pages.
Domains are sorted into six tiers by PageRank on Common Crawl’s hyperlink graph, where domains are nodes and links between them are edges. The most central tier was taken in full, pages were sampled within each selected domain, and the results were reweighted by language, length, tier and top-level domain to match the eligible-page population.
Language comes from Common Crawl’s per-page flag, and in the August 2026 snapshot 35.9% of eligible pages are English-only. Bilingual pages are scored as English and unknown-language pages as non-English.
About 24% of eligible pages, with a credible interval of 22.9 to 24.4
For August 2026 the authors estimate that about 24% of eligible pages contained AI-generated or AI-edited text, with a 95% credible interval of 22.9 to 24.4. The abstract phrases the same result as “nearly 25%”. This is the authors’ own model output, not a citation of anyone else’s work.
The measure counts pages containing text “generated with or revised by” a language model, so an AI-edited page counts. It is not a share of words, and it does not mean the whole page was machine-written.
Before the commercial language-model era, prevalence was low but not zero: between 0.8% and 4.9% across all tiers in 2019. Average growth ran at 0.2 percentage points per year from January 2019 to February 2023, 4.6 points per year from February 2023 to January 2025, and 8.1 points per year from January 2025 to August 2026, a difference of 3.6 points (credible interval 3.3 to 3.9) between the last two rates. These are averages over fixed periods, not an estimated turning point.
Detector error varies by language and tier, and the model corrects for it
Two detectors did the scoring. Open Pangram, a supervised classifier fine-tuned from Llama-3.2-3B, read every page in every language. Binoculars with the Falcon-7B pair, a zero-shot perplexity-ratio detector, read only English pages: its sensitivity on non-English text was 0.151, which the authors say did not justify scoring other languages at scale. Pangram 4 was used only to check the deployed detectors’ error rates, because its mistakes track Open Pangram’s too closely to count as an independent reading.
Those raw scores are not the published estimate. The authors treat a detector flag as an imperfect indicator of whether a model was involved, then estimate how imperfect it is. Sensitivity is allowed to vary by detector and language. Specificity is allowed to vary by detector, language, centrality tier, top-level domain and text length. Specificity is anchored on pre-2022 control crawls, when prevalence was low, and sensitivity on four labelled benchmark datasets, one containing text in nine languages. Error rates may drift after the control period, and uncertainty in them is propagated into every prevalence figure and comparison.
Pew reports a raw flag rate with no correction, from the same Open Pangram model. A detection result is not a prevalence measurement, and the paper’s own chart shows why: its uncorrected flag share sits well above the corrected, poststratified estimate, because a raw flag rate carries both detector error and the composition of the sample.
The gradient: 11.6% in the most central tier, 27.5% in the least
The uneven spread is the part no other study in this comparison provides. In August 2026 the most central tier sat at 11.6% (credible interval 11.2 to 11.9) and the least central tier at 27.5% (26.5 to 28.4), with every tier in between.
The two least central tiers hold 70.5% of eligible pages and account for 78.8% of the net prevalence increase between December 2021 and August 2026, with a conservative one-sided 95% lower bound of 77.9%. Holding the page distribution fixed at the average of the two endpoints, prevalence still rose 24.0 percentage points there against 16.0 elsewhere, so the pattern is not an artefact of the web drifting toward peripheral pages.
English-only pages carry more AI text than the rest: 32.8% (32.2 to 33.4) against 17.0% (15.4 to 18.4) for non-English-only pages. The prevalence ratio between the two groups is 1.93 (1.78 to 2.13), a relative figure rather than a percentage-point gap. Top-level domains differ too: .com at 29.5% (28.7 to 30.2), .org at 14.6%, .edu at 6.1% and .gov at 5.8%.
Pew, Dolezal and Graphite measure different populations
Pew’s 10% is the share of its July 2026 sample that Open Pangram flagged at a score of 0.2 or higher. That sample is 10,000 pages per crawl across 49 Common Crawl crawls from January 2021 to July 2026, English only, drawn at random and scored without correction. Pew also flags over one third of pages with post-ChatGPT publication dates, while noting that only 10 to 15% of pages carry a publication date at all and that subset is not random.
Dolezal et al. (arXiv:2604.26965) used the Internet Archive’s Wayback Machine across 33 monthly intervals from August 2022 to May 2025, English only, scored with the commercial Pangram v3 API, and treated a newly published website as the unit. Their headline, roughly 35% of newly published sites classed as AI-generated or AI-assisted by mid-2025, describes new material rather than the crawlable web as a whole.
Graphite Growth self-publishes its research. Its May 2026 study draws 55,400 random English URLs from Common Crawl, keeps pages with article schema and at least 100 words, averages three detectors, and reports that about 50% of published articles have been AI-generated, a share it says has plateaued since the first quarter of 2025.
One study estimates total prevalence, two measure new material, and one measures the flagged share of a mixed-age sample. One corrects for detector error and three do not. Eligibility differs: no character floor at Pew, a longest-paragraph rule at Dolezal et al., article schema at Graphite. Three of the four are English-only, a subset where prevalence is higher. Pew and the new paper share the same underlying open Pangram model, so they are less independent of each other than the numbers suggest.
Lower .gov and .edu prevalence survives adjustment, with one caveat
Government and university domains look different from commercial ones under every cut in the paper. .edu sits at 6.1% and .gov at 5.8% in August 2026, against 29.5% for .com. Standardising to a common composition of centrality, language and length does not remove the ordering: .com falls to 26.4%, while .gov and .edu rise to 7.5% and 7.0%.
Those standardised estimates rest on model predictions for sparse cells. The cells the model represents cover 22.5% of the reference distribution for .gov and 16.6% for .edu, a substantial extrapolation the authors acknowledge. Detector behaviour varies by domain type as well: Open Pangram has fewer false positives than Binoculars-Falcon on .gov and .edu but more on .com. The lower institutional share is a measured difference under this design, and it is not proof that government and university publishing is untouched by AI.
What these numbers do not cover
Every figure here is a property of one corpus, one detector and one denominator, and none is a share of “the internet”. The crawl sees crawlable pages a detector can classify, each with at least 500 characters of main text; logins, apps, crawler-blocked sites and short standalone text such as comments and reviews sit outside the measure. Prevalence is descriptive: the paper does not claim to explain why peripheral pages carry more AI text.
Prevalence is also not visibility. Graphite’s October 2025 study looked at articles ranking in the top two Google results for 31,493 keywords collected in June 2025 and found 14% of them classified as AI-generated in its own sample, against the roughly 50% the same firm reports among published articles.
No independent group has yet re-run this pipeline. Anyone quoting a share of the web should carry three facts alongside the number: which pages counted, which detector scored them, and whether the figure describes all pages or only new ones.
