Uses14 November 2026Is AI already writing most of the Internet?
Articles, pages, traffic and clicks, measured from 2023 to 2026
Photo Solen Feyissa, CC BY-SA 2.0, via Wikimedia Commons.
For new articles published in English, almost: since the start of 2025, about half have been written mainly by an AI, according to a study that classified 55,400 of them with 3 detectors, and this share has not risen for 5 quarters. For the web as a whole, no: fully generated pages remain rare, and texts that were online before 2023 have not disappeared. On the reading side, AI bots account for only about 4% of page requests at Cloudflare, but some crawl thousands of pages, up to tens of thousands in Anthropic's case, for each visitor they send back, and in the United States Google searches that display an AI summary lead almost 2 times less often to a click to a website. The frequently cited forecast of "90% of generated content in 2026" is based on no measurement.
Contents
The "dead Internet" theory has circulated since the early 2020s: most online texts and exchanges would be produced by machines, for machines. The arrival of ChatGPT, at the end of 2022, gave it a concrete basis. Yet it mixes several questions, each with its own measurements: who writes the new texts, who visits the pages, and who still reads the websites.
All these measurements of writing rely on automatic detectors, whose errors are detailed in How can you tell whether a text was written by an AI?. The studies below publish their error rate, measured on texts whose author is known.
Share of new articles written mainly by an AI
The agency Graphite randomly drew 55,400 addresses from Common Crawl, a public copy of the web, and kept the articles in English published from January 2020 to March 2026. Each article went through 3 detectors, whose results are averaged. On articles predating ChatGPT, the detectors wrongly classify fewer than 2% of texts as generated.
Share of new articles in English written mainly by an AI:
The share rose quickly, then stopped at around half. The authors put forward an explanation without measuring it: generated articles would be ranked lower in search engines, which would discourage producing more of them. Their 1st version, published in October 2025 with a single detector, gave shares that were 3.3 points higher.
The scope is narrow: articles in English, tagged as such, of at least 100 words. A mixed text may be counted as written mainly by an AI: GPTZero puts in this category the texts it judges mixed, and Pangram adds passages edited with an AI to those it wrote.
Share of AI depending on what is counted
Depending on what is counted, the share of AI ranges from a few per cent to nearly three quarters. These figures do not contradict each other: they do not measure the same thing.
| What is counted | Share attributed to AI | Period |
|---|---|---|
| New pages in English containing at least one generated passage | 74.2 % | April 2025 |
| New articles in English written mainly by an AI | 49.9 % | 1st quarter 2026 |
| New English-language Wikipedia articles flagged as generated | More than 5% | August 2024 |
| New pages in English that are fully generated | 2.5 % | April 2025 |
Sources: Ahrefs, 900,000 pages, publisher's detector; Graphite, 2026; Brooks et al., preprint; Cloudflare Radar. Each study has its own detector and threshold.
The gap between 74.2% and 2.5% comes from a single study, that of the software publisher Ahrefs: almost all recent pages contain a passage flagged as generated, and very few are entirely so. On Wikipedia, detectors flag more than 5% of new English articles from August 2024, a low estimate obtained with other settings.
None of these studies measures the stock of the web. Pages published before 2023 remain online, and a new article does not replace the old ones. The share of AI in all texts online is therefore lower than in recent publications, although no study we have found puts a figure on it.
Bot traffic and page reading
Bots have long made up a large share of visits: according to the 2025 report by the security vendor Imperva, automated traffic exceeded human traffic, at 51%, and malicious bots made up 37% of it. These bots are of all kinds, from search engine indexing bots to fraud bots; most have nothing to do with generative AI.
Among customers of Cloudflare, which carries a large share of web traffic, people made 47% of page requests on 2 December 2025 and non-AI bots 44%. Over the year, AI bots made 4.2% on average, and Googlebot, counted separately, 4.5%. Their weight is better read in another measure: the number of pages an AI company crawls for each visitor it sends back to the site.
| Company | Pages crawled per 1 visitor referred |
|---|---|
| Anthropic | 25,000 to 100,000, with a peak at 500,000 |
| OpenAI | Up to 3,700, in March 2025 |
| Microsoft | Peaks of 50 to 70 |
| About 3, up to 30 in April 2025 |
Cloudflare Radar, review of the year 2025. Crawl requests by each company's bots, relative to the visits coming from its services.
A search engine crawls a page and sends readers to it. An AI assistant crawls it to answer itself, and sends back far fewer visitors, as these reports show.
Searches and clicks to websites
Assistants have become a way of searching. OpenAI announced in October 2025 more than 800 million weekly ChatGPT users, and Google 2 billion monthly users of its AI summaries in July 2025. According to Similarweb, these summaries appeared on more than 40% of American searches in August 2026, and 95% of ChatGPT users also used Google in May 2026.
Share of visits to Google results pages followed by a click to a website, in March 2025, among 900 American adults whose browsing was tracked:
With a summary, only 1% of visits lead to a click on a link in the summary itself. The reader more often stops on the Google page: the session ends on 26% of pages with a summary, against 16% without. The blog compares the footprint of a Google search and of a question to ChatGPT in ChatGPT or Google search.
Origin of the 90% forecast for 2026
The most cited forecast announced that 90% of online content could be generated in 2026. The website Futurism attributed it in September 2022 to a report from Europol's innovation lab on deepfakes, intended for police services. According to this citation, the report relied on "experts", without naming them or giving a method, and spoke of "synthetic" content, which, in a report on deepfakes, includes images and videos. We could not read the 2022 version; the January 2024 version, which replaces it, does not contain this figure.
In 2026, the available measurements give about half of new articles in English, far less for other pages and for the web as a whole.
Generated texts and model training
The web is also used to train models. In 2024, a team of researchers showed in Nature that a model trained indiscriminately on texts produced by other models loses, generation after generation, the rare cases of its original data. They call this effect "model collapse".
The risk depends on the share of generated texts in the data and on the sorting that publishers do of them, which they do not publish.
Our reading
The answer depends on the question asked. AI already writes half of new articles in English, and this share has stopped rising. It does not write most of the Internet, which keeps its older texts and pages still written by people. The 90% forecast had no measurement behind it.
In our view, the change that weighs most on websites is on the reading side: AI bots crawl many pages for few visitors sent back, and generated summaries reduce clicks to sources. The measurements available are still short, covering a month or a year, and do not say whether this trend is growing.
Measurements of writing rely on detectors, whose error is known on older texts and on a few recent models, and which may recognise future models less well. They cover English. The Cloudflare and Imperva traffic covers only their customers, and the Pew Research Center study covers only Google, in the United States; it observes different searches with and without a summary, without measuring the effect of the summary itself. No study found measures the share of AI in all texts online.
Sources
13 documents
- Graphite, AI Now Writes as Many Online Articles as Humans2026entry 548
- Graphite, More Articles Are Now Created by AI Than Humans2025entry 549
- Ahrefs, 74% of New Webpages Include AI Content (Study of 900k Pages)2025entry 550
- Brooks, Eggert and Peskoff, The Rise of AI-Generated Content in Wikipedia, arXiv preprint2024entry 551
- Cloudflare, Radar 2025 Year in Review2025entry 552
- Imperva (Thales), 2025 Bad Bot Report2025entry 553
- TechCrunch, Sam Altman says ChatGPT has hit 800M weekly active users2025entry 554
- TechCrunch, Google's AI Overviews have 2B monthly users, AI Mode 100M in the US and India2025entry 555
- Similarweb, Is AI Replacing Search? What The 2026 Numbers Actually Reveal2026entry 556
- Pew Research Center, Google users are less likely to click on links when an AI summary appears in the results2025entry 547
- Futurism, Experts: 90% of Online Content Will Be AI-Generated by 20262022entry 557
- Europol Innovation Lab, Facing reality? Law enforcement and the challenge of deepfakes, January 2024 version2024entry 558
- Shumailov et al., AI models collapse when trained on recursively generated data, Nature2024entry 559
Every figure was read in its source; the hub's methodology is described in Methodology and sources.
AI Impact, Projet Celsius, "Is AI already writing most of the Internet?", 14 November 2026. https://projetcelsius.com/en/ai/analyses/is-ai-writing-most-of-the-internet/
Electricity
