On August 20, 2026, Pew Research Center published an analysis measuring how much text on the web is now written by AI1. In a random sample from July 2026, 10% of all pages showed significant signs of AI authorship or editing, and among pages whose publication dates could be identified as post-ChatGPT, the share rose above one-third1.
The center frames the study around what it calls a basic question about the modern web, raised by these tools’ ability to generate human-sounding text1. The authors are Samuel Bestvater, Aaron Smith, Carson TerBush, Chris Baronavski, and Janakee Chavda1.
490,000 pages through one detection model
The method is published. Pew used data from Common Crawl, a nonprofit that has maintained a large web archive since 20082. Common Crawl completes a crawl of the observable internet roughly once a month, creating a snapshot of the web at that point in time2.
Pew randomly sampled 10,000 English-language pages from each of the 49 crawls created between January 2021 and July 2026, for a total of 490,000 pages2. For each page it collected both the WARC record containing the full HTML and the WET file containing just the body text, excluding HTML code and images2.
For detection, Pew used editlens_Llama-3.2-3B (Open Pangram), an open-weight model developed by Pangram2. The model takes text and returns a score from 0 to 1, where 0 indicates fully human-written text and 1 indicates fully AI-generated text2. For this analysis, any page scoring 0.2 or higher was treated as containing “meaningful” signs of AI authorship or editing2. Pew explains that the model looks at patterns in language — words, phrases, and linguistic quirks used more commonly by AI than by human authors1.
10% and 35% count different things
In the 10,000-page sample collected in July 2026, 10% showed significant signs1. Pew describes this as part of an upward trend that began in late 2022 and continued as other chatbots such as Claude and Gemini followed1.
But the web mixes new and old material. Pew notes that many pages in these samples could not have been written by AI at all, and so filtered down to pages published after ChatGPT’s release on November 30, 20221. In the July 2026 snapshot, over one-third of those pages showed signs1. The methodology page gives the figure as 35%2.
This is the point most easily misread, and Pew flags it. Only about 10% to 15% of pages in a given crawl sample carry a publication date field in their HTML2. The center states that “the subset of pages with a detectable publication date is not a random subset of the web, so our post-ChatGPT estimate reflects the prevalence of AI authorship among dated content rather than the web as a whole”2. The 35% figure therefore describes dated pages, not one-third of the web.
Pew notes another limit: because pages must be publicly accessible to enter a Common Crawl snapshot, sites behind paywalls or logins are likely underrepresented2.
The center says the result is broadly in line with related work. According to the methodology page, Dolezal et al. (2026), using Internet Archive data, found that 35% of newly published websites contained AI-generated or AI-assisted text2.
The gap between domains widened
The signs are not evenly distributed.
Pew reports that when ChatGPT was first released, the linguistic patterns that can signal AI authorship appeared at similar rates across the main top-level domains (.com, .org, .edu, and .gov)1. In 2026 samples, around one in ten .com pages show signs, compared with 4.6% on .org and around 1% on both .edu and .gov1. That makes .com roughly double .org and ten times .edu or .gov1. Plotted points in the chart represent six-month averages1.
The signs run heaviest on commercial pages and lightest on educational and government ones.
”AI-sounding” writing spread
Pew also measured changes in the writing itself.
The center first notes that AI detection models are imperfect — sometimes misclassifying human-written documents as AI, and vice versa — but that looking at very large collections of texts together reveals certain punctuation, words, and phrases appearing at much higher rates in AI-generated content1.
Pew’s explanation: AI models train on large datasets of human writing and learn to mimic its patterns. Some kinds of writing are overrepresented in that training, so the resulting models use certain words, phrases, or quirks more often than humans typically do1. The example given is the em dash (—), a mark widely used in journalistic and academic writing that AI models tend to use far more often than humans1. Models are also more likely than human authors to list items in threes and to use Oxford commas1.
Comparing today’s web with a 2023 snapshot, measured in uses per 10,000 words1:
- Em dashes: about twice as frequent
- Oxford commas: a 63% increase
- Vocabulary AI models favor (“delve,” “interplay,” “testament,” and others): more than doubled
- “Negative parallelism” (“it’s not just X, it’s Y”): nearly tripled, though still fairly rare overall
The AI-typical vocabulary Pew lists includes “additionally,” “align with,” “boasts,” “crucial,” “delve,” “enhance,” “essential,” “highlight,” “interplay,” “key,” “landscape,” “meticulous,” “pivotal,” “showcase,” “significant,” “tapestry,” “testament,” “underscore,” “valuable,” and “vibrant,” with this comparison limited to pages published after November 30, 20221.
The original does not attribute these increases to AI as a cause. It presents them as an observation that AI-favored features have become more common on the web.
Detection models are not verdicts on individual pages
Another important element is that Pew is explicit about how detection models should be used.
To check that the results were not dependent on one model, the center ran 62,370 pages from seven crawls through Pangram’s flagship commercial model, Pangram 3.3, and compared the results2. The two models agreed in 96% of cases, with a Cohen’s kappa of 0.612.
Pew then writes that the page-level disagreements are a reminder that AI detection models are probabilistic tools, and that their classification of an individual page should not necessarily be taken as a definitive verdict about that page’s authorship2. Applying the same tool consistently across a large collection of pages, however, does allow the aggregate share to be tracked systematically over time2.
There is a caveat about early crawls too. The largest deviations between the two models appeared in pre-ChatGPT crawls, where Pangram 3.3 produced very low estimates while Open Pangram came in around 1%. Pew says this suggests a higher false-positive rate for Open Pangram on pages created before AI use was common2.
Different stakes for readers and publishers
The numbers land differently depending on where you sit.
For anyone publishing content, the widening gap between domains shows that commercial writing is converging. Rising em dash and Oxford comma counts are not themselves quality problems, but the fact that machine-typical features are spreading is worth considering when deciding where differentiation actually comes from.
For anyone drawing traffic from search, this pairs with Google’s move on the same day to give publishers an embeddable “Preferred Sources” button. As AI-generated pages multiply, the question of which sources get shown moves to the front. Browsers are moving too: Mozilla recently added an Exa partnership to Firefox’s Smart Window so it can pull current web information with citations.
And for anyone hoping to decide internally whether “AI wrote this,” Pew’s caveat is closer to the point than the headline figure. The researchers state plainly that detection models work for aggregates but do not settle authorship of individual documents2. Using detection scores to judge people or content individually means deciding who absorbs the 4% of cases where two models disagree — the flip side of that 96% agreement and kappa of 0.61.
Telling machine-written material from human-written material is not only a text problem. Cloudflare in August introduced a way to give AI agents an identity and a wallet, an answer aimed at making it provable who is making a request. Pew’s study counts, through a published procedure, what the web looks like without such proofs in place.
Sources
- How Much of the Internet Is Written With AI? - Pew Research Center (August 20, 2026)
- Methodology - Pew Research Center, methodology for the above study (August 20, 2026)
- A third of web pages published since ChatGPT’s launch show signs of AI authorship, study finds - TechCrunch (August 20, 2026)