A headline-grabbing 35% estimate rests on a much thinner slice of the web than it first appears: only 10% to 15% of the pages in each crawl carried a publication date that researchers could extract. That limitation does not make the result unimportant. It makes the real consequence more urgent for developers. The web is filling with machine-shaped text just as its provenance is becoming harder to establish, leaving search engines, archive projects, and AI teams to make consequential filtering decisions with incomplete metadata.
Pew Research Center analyzed 490,000 English-language pages collected from Common Crawl between January 2021 and July 2026. In the July 2026 snapshot, 10% of all sampled pages showed what Pew called meaningful signs of AI authorship or editing. Among pages carrying a detectable publication date after ChatGPT's public release on November 30, 2022, the estimate rose to 35%.
That is a large enough shift to affect more than debates over whether online prose sounds repetitive. It changes the raw material used to build search indexes, retrieval systems, research corpora, and future language models. But the finding is best read as a trend measured by a detector, not as a definitive count of pages written by machines.
What Pew actually measured
The researchers drew a random sample of 10,000 English-language pages from each of 49 Common Crawl snapshots. For every page, they collected the full HTML archive record and a text-only version. They then ran the body text through editlens_Llama-3.2-3B, an open-weight detector developed by Pangram.
The model assigns a score between zero and one, intended to represent a spectrum from human-written to AI-generated text. Pew classified a page scoring 0.2 or higher as showing meaningful evidence of AI authorship or editing. This matters because the study was not limited to pages supposedly generated end to end. A human draft substantially rewritten with an AI assistant could also cross the threshold.
The broadest figure is therefore the least dramatic and the easiest to interpret: one in ten pages in the July 2026 sample crossed that line. That sample mixes recently published material with pages created before consumer chatbots became widespread. It offers a snapshot of the crawlable English-language web, old and new together.
The 35% estimate answers a narrower question. Pew used publication dates embedded in page HTML to isolate material published after ChatGPT arrived. Its methodology notes that only 10% to 15% of pages in a typical sample exposed such a date, and that this dated subset is not random. News articles and structured blog posts commonly publish machine-readable dates; homepages, product pages, documentation, forums, and many other page types may not. The figure describes dated content in the sample, not every new page on the web.
Pew also compared its open model with Pangram's commercial detector on 62,370 pages from seven crawls. The two agreed on 96% of page classifications, but their Cohen's kappa score was 0.61, a measure that adjusts for agreement occurring by chance. That supports using the detector to observe a broad movement across a large corpus. It does not support treating its verdict on an individual article as proof.
The detector is an instrument, not an authorship registry
AI-text detection has a stubborn underlying problem: language models learn from human language, while humans adopt conventions they encounter in machine-written text. The two distributions are not fixed and separate. Editing, translation, templates, corporate style guides, and non-native English writing can all complicate classification.
Pew openly identifies evidence of false positives. On pages from early 2021 and 2022, before ChatGPT's release, the open model estimated an AI share of roughly 1%, while the commercial detector produced much lower estimates. The researchers say those early open-model readings probably include human writing classified incorrectly.
The study also reports aggregate changes in familiar stylistic features between 2023 and 2026. Em dashes appeared about twice as often, Oxford commas increased by 63%, and a comparison pattern built around “not just X” nearly tripled, although it remained uncommon overall. These are population-level clues, not a checklist for accusing a writer. Every one of those constructions predates modern language models.
This distinction is central. A detector can be useful as a measuring instrument even when it is unsuitable as an authorship tribunal. Weather radar can show a storm front without identifying where every raindrop landed. In the same way, a stable classifier applied to repeated samples can reveal a directional shift while still being wrong about particular documents.
TechCrunch's report on the study foregrounded the one-third result. The methodological caveat is what determines how that number should be used. It is reasonable evidence that AI-assisted writing has become common among newly dated pages. It is not evidence that exactly 35% of everything published since late 2022 was generated by AI.
A provenance problem for search and model training
For web users, the immediate concern is quality rather than authorship. A carefully checked page does not become useless because an AI system helped edit it, and a fully human page is not automatically accurate. The operational problem is that automation can lower the cost of publishing plausible, redundant, or fabricated material at enormous volume. Origin and accountability become harder to trace precisely when production becomes easier to scale.
Search systems must decide which pages deserve to be indexed, ranked, summarized, or cited. Conventional signals such as backlinks, domain history, structured metadata, and textual originality can all be manipulated. If large clusters of sites paraphrase the same generated answer, simple measures of apparent consensus may count copies as independent support. Retrieval-augmented applications inherit the same weakness when they fetch several sources that ultimately repeat one unverified claim.
Model builders face a related feedback loop. Common Crawl has long served as raw material for machine-learning datasets. If a growing share of newly crawled prose is written or heavily edited by previous models, future training sets can contain more synthetic output and less direct human expression. That does not automatically ruin a dataset, but it makes source selection, deduplication, temporal filtering, and provenance records more important. A corpus labeled only by URL and language is increasingly an incomplete description of what it contains.
The study itself demonstrates the metadata gap. Publication date is a basic field, yet it was detectable on a small minority of sampled pages. Authorship disclosures are rarer and inconsistent. There is no universal web field that reliably says which model touched a paragraph, whether a human reviewed it, or which sources supported it. Voluntary labels can help responsible publishers but are unlikely to describe the entire public web.
Developers building search or retrieval products should therefore resist collapsing the issue into a binary AI=true flag. More useful records would preserve crawl time, claimed publication time, canonical URL, content hashes, detected duplication, source links, revision history where available, and detector scores with model versions. Those fields do not solve truth assessment. They make later audits possible and keep an uncertain estimate from masquerading as ground truth.
Commercial pages show the sharpest concentration
The domain split adds another useful signal. In Pew's 2026 samples, around 10% of pages on .com domains showed AI markers, compared with 4.6% on .org and roughly 1% on both .edu and .gov. At ChatGPT's release, the linguistic patterns used by the detector appeared at more similar rates across these domain groups.
Top-level domains are coarse categories. A .com address can host a personal blog, a global retailer, documentation, or a publication, while institutional domains can contain many kinds of writing. Still, the gap fits a practical incentive structure: commercial publishers often have strong reasons to produce marketing, support, and search-oriented copy quickly. The result points toward where automated text production may be concentrating, without establishing why any particular page was created.
It also warns against treating the web as one uniform dataset. A crawl's composition can shift as domains change access rules, paywalls block collection, sites disappear, and platforms generate new pages. Pew notes that Common Crawl contains only publicly accessible material, so paywalled and login-gated sites are likely underrepresented. The measured web is the crawlable web, not the whole of online publishing.
What to watch next
The most useful follow-up will not be another isolated percentage. Researchers need repeated measurements across languages, page types, detectors, and independently labeled samples, with uncertainty reported alongside the headline figure. Search companies and model labs, meanwhile, should show how they identify duplicated synthetic material and preserve source lineage without penalizing legitimate AI-assisted editing. Pew's study establishes that the shape of new web text has moved substantially. The next question is whether the systems consuming that text can distinguish volume from evidence.