More than one-third of English-language web pages published after ChatGPT’s public launch in November 2022 show significant signs of AI authorship or editing, according to a new analysis by Pew Research Center. The finding offers one of the clearest indications yet of how quickly generative AI has moved from an experimental technology into a major source of online content.
Pew analyzed nearly half a million webpages collected from the Common Crawl archive between January 2021 and July 2026. In a random sample of 10,000 webpages collected in July 2026, about 10% showed significant signs of AI authorship. But after researchers excluded older pages that predated ChatGPT and looked only at pages with detectable publication dates after November 30, 2022, the share rose to 35%.
35% Of Post-ChatGPT Web Pages Show AI Signals
The headline finding from Pew’s research is that 35% of webpages with identifiable publication dates in the July 2026 crawl and published after ChatGPT’s launch showed meaningful signs of AI authorship.
The researchers used Open Pangram, an AI-detection model developed by Pangram, to analyze the body text of webpages. The model produces a score between zero and one, with higher scores indicating stronger signs of AI-generated text. Pew classified pages with a score of at least 0.2 as showing meaningful signs of AI authorship or editing.
Importantly, the study does not claim that 35% of the entire internet is AI-written. The estimate applies to pages with detectable publication dates in the July 2026 sample. Pew notes that only around 10% to 15% of webpages in a crawl typically contain a publication-date field, meaning the dated subset is not a random representation of the entire web.
Key Numbers From The Study
| Metric | Finding |
|---|---|
| Total webpages analyzed | 490,000 |
| Common Crawl samples | 49 crawls |
| Pages sampled per crawl | 10,000 |
| Study period | January 2021-July 2026 |
| July 2026 random sample | 10,000 pages |
| All July 2026 pages showing AI signals | ~10% |
| Post-ChatGPT pages showing AI signals | 35% |
| ChatGPT public launch | November 30, 2022 |
| AI detection model | Open Pangram |
| Detection threshold | 0.2+ |
The distinction between the 10% overall figure and the 35% post-ChatGPT figure is important. The broader sample contains pages published years before generative AI became widely available, while the filtered figure focuses on newer content that could plausibly have been created or substantially edited using modern AI tools.
AI Authorship Has Increased Rapidly Since 2022
Pew’s data shows that AI-related writing signals have become increasingly common since ChatGPT’s release.
Before ChatGPT became publicly available, the study found relatively low and broadly similar levels of AI-authorship signals across major web domains. The numbers began rising after late 2022 and accelerated through 2024, 2025 and 2026.
The growth is particularly visible on commercial websites. By January 2026, about 9.35% of sampled .com pages showed significant AI-authorship signals on a six-month-average basis, compared with 4.59% of .org pages and around 1% of .edu pages.
AI Signals Across Major Web Domains
| Domain | Jan. 2023 | Jan. 2024 | Jan. 2025 | Jan. 2026 |
|---|---|---|---|---|
| .com | 1.63% | 3.76% | 5.48% | 9.35% |
| .org | 0.90% | 2.00% | 2.89% | 4.59% |
| .edu | 0.56% | 1.70% | 0.95% | 1.03% |
| .gov | 0.34% | 1.42% | 1.62% | 0.76% |
The difference between commercial and institutional websites is significant. By January 2026, .com pages were showing AI-authorship signals at roughly 10 times the rate of .edu and .gov pages, according to Pew’s analysis.
Commercial Websites Are Seeing The Biggest Increase
The strong growth on .com domains suggests commercial content is one of the areas where generative AI is being adopted most aggressively.
Businesses can use AI tools to produce product descriptions, blog posts, guides, landing pages, marketing material and other text-heavy content at a much lower marginal cost. The technology can therefore make it easier for companies to increase publishing volume.
However, a rise in AI-assisted publishing also creates a differentiation problem. If more businesses can generate large quantities of reasonably fluent content, simply publishing more articles may become less valuable.
January 2026 AI-Authorship Signals
| Domain Type | Share Showing AI Signals |
|---|---|
| .com | 9.35% |
| .org | 4.59% |
| .edu | 1.03% |
| .gov | 0.76% |
Pew’s figures are based on sampled webpages rather than all pages on these domains. They should therefore be interpreted as indicators of a broader trend rather than an exact measurement of every website category.
AI-Associated Writing Patterns Are Becoming More Common
The study also examined specific language and punctuation patterns that have become more common on the web since ChatGPT’s launch.
Pew identified several characteristics that AI models tend to use more frequently than human writers, including em dashes, Oxford commas, certain AI-associated vocabulary and a writing structure known as negative parallelism.
The researchers found that these patterns have increased substantially since 2023.
How AI-Associated Writing Patterns Changed
| Writing Feature | Change Since 2023 |
|---|---|
| Em dashes | About 2x more frequent |
| Oxford commas | +63% |
| Selected AI-associated vocabulary | More than 2x |
| Negative parallelism | Nearly 3x |
Negative parallelism refers to constructions such as “it’s not just X, it’s Y.” Pew says the structure remains relatively uncommon overall, despite nearly tripling in frequency.
The researchers also identified words that AI models appear to favor more frequently, including terms such as “delve,” “interplay,” “testament,” “bolstered,” “fostering” and “pivotal.”
However, the presence of any one of these words or punctuation marks does not establish that a human did not write a particular article.
AI Detection Is Not A Perfect Test
Pew explicitly cautions against treating AI-detection results as definitive judgments about individual webpages.
AI detectors are probabilistic systems. They can incorrectly classify human-written content as AI-generated, while AI-written material can sometimes avoid detection. This is particularly important when analyzing individual articles rather than large datasets.
To test the reliability of its approach, Pew also ran 62,370 webpages through Pangram’s commercial Pangram 3.3 model and compared the results with the open-weight model used for the main analysis.
The two systems agreed in 96% of cases, although the researchers noted that page-level disagreements demonstrate why individual classifications should be treated cautiously.
What The Detection Result Actually Means
| Result | Interpretation |
|---|---|
| High AI signal | Text has statistical characteristics associated with AI |
| Low AI signal | Text resembles human-written material more closely |
| AI signal detected | Does not prove AI wrote the entire page |
| Human classification | Does not guarantee no AI assistance |
| Large-scale trend | More useful than an individual-page verdict |
The study therefore measures signs of AI authorship or editing, rather than providing a definitive record of who wrote each webpage.
The Study Covers Nearly Half A Million Web Pages
The scale of the research is another important element.
Pew randomly selected 10,000 English-language webpages from each of 49 Common Crawl snapshots between January 2021 and July 2026. That produced a total sample of 490,000 pages.
Common Crawl is a nonprofit web archive that periodically captures publicly accessible pages. Websites behind paywalls or login systems are less likely to appear in the archive, creating another limitation that researchers acknowledge.
Study Methodology
| Methodology Element | Detail |
|---|---|
| Data source | Common Crawl |
| Language | English |
| Number of crawls | 49 |
| Pages per crawl | 10,000 |
| Total pages | 490,000 |
| Earliest crawl | January 2021 |
| Latest crawl | July 2026 |
| AI detector | Open Pangram |
| Cross-check | Pangram 3.3 |
| Pages used for cross-check | 62,370 |
Pew also compared its results with another 2026 study, “The Impact of AI-Generated Text on the Internet,” which used Internet Archive data and similarly found that about 35% of newly published websites contained AI-generated or AI-assisted text.
What This Means For The Future Of The Web
The findings suggest that generative AI is changing not only how people search the internet but also how the internet itself is produced.
Before generative AI became widely available, producing large volumes of written content required human writers, editors and publishers. AI tools have dramatically reduced the time and cost involved in producing text, allowing organizations and individuals to create content at much greater scale.
That does not necessarily mean AI-generated content will replace human-created content. Instead, the distinction between human and AI authorship is becoming increasingly blurred, particularly as writers use AI for research, drafting, editing, translation and rewriting.
Implications For Search And Online Publishing
The increase in AI-assisted content could make quality and originality increasingly important for publishers.
If the supply of generic articles continues to grow, publishers may need to provide information that AI systems cannot easily reproduce from existing online material. Original reporting, proprietary data, expert interviews, first-hand experience and unique analysis can become more valuable as the amount of generic AI-generated text increases.
For search engines and AI systems, the growing volume of synthetic content also creates a challenge. Automated systems must increasingly distinguish between useful information and large quantities of repetitive or low-value material.
This could influence how search rankings, recommendations and AI-generated answers are developed in the coming years.
The Bigger Picture
Pew’s research provides a quantitative view of a transformation that has been underway since ChatGPT became publicly available in November 2022. The finding that 35% of dated post-ChatGPT webpages in its July 2026 sample show meaningful AI-authorship signals suggests that AI is now deeply embedded in online publishing. The particularly high growth on .com domains indicates that commercial content is at the center of this shift.
At the same time, the study does not support the conclusion that one-third of the entire internet is AI-generated. Its methodology is based on dated pages in Common Crawl and uses a probabilistic detector. The more defensible takeaway is that AI-assisted writing has become a substantial and rapidly growing part of newly published web content.
Looking Ahead
The share of web pages showing AI-authorship signals is likely to remain an important metric as generative AI becomes more deeply integrated into publishing, marketing and online commerce. For publishers, the growing availability of automated writing tools could make content production cheaper, but it may also make generic content easier to replicate and harder to differentiate.
The longer-term impact will depend on how search engines, publishers and readers respond to the expanding supply of AI-assisted material. As machine-generated writing becomes more common, original reporting, expertise, evidence and distinctive human experience could become increasingly important signals of value in an internet where simply producing more text is no longer difficult.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
