Half the new internet is written by machines.

In 2021 it was a fringe conspiracy. By 2026 the head of OpenAI is publicly reckoning with it, security firms measure bots as the majority of web traffic, and Wikipedia reports losing human readers. The dead internet theory has not been proven literally — but the direction it pointed in now has data. And that data changes the question from ‘is it true?’ to ‘how much of the internet still belongs to people — and what do we do about it?’

Where it came from

The sense that something was wrong with the internet — content that felt repetitive, discussion that felt sterile, viral waves that felt synchronised — circulated on imageboards for years. It took its finished shape in 2021, in a forum thread arguing that most of the internet is fake: that human activity online is being crowded out by bots and automated content. The theory always had two layers — an observable one (machines are displacing people) and a conspiratorial one (that it is deliberate). The Atlantic covered it that September with healthy skepticism. Reasonably so: in 2021, the idea of an internet written by machines sounded overblown. ChatGPT did not exist yet.

Then machines started writing

Late 2022 changed the economics of content: text stopped costing time. Two independent studies have since measured the effect. The SEO firm Graphite analysed 65,000 articles published between 2020 and 2025 and found the share of AI-generated articles rose from roughly a tenth before ChatGPT to the point where, in November 2024, machine output briefly overtook human for the first time — and has hovered around a 50/50 split since. Ahrefs, examining 900,000 newly published pages in April 2025, went further: 74.2% of new pages contain AI-generated text. The more revealing detail is that purely machine-made pages were only 2.5% — the overwhelming majority is a hybrid of human and machine. The web was not replaced. It was grown through.

Bots are officially the majority

Content is half the story; the other half is who moves across the internet at all. The security firm Imperva has measured automated traffic for over a decade. Its 2025 report marked a milestone: in 2024, automated traffic overtook human for the first time, at 51% of all web traffic. A year later, the follow-up — pointedly titled for the ‘agentic age’ — put the figure at 53% for 2025, with the human share still falling. Independent measurement agrees by a different metric: Cloudflare data disclosed by its CEO in June 2026 put bots at 57.5% of HTML requests on its network. The most telling part is the purpose — by Cloudflare’s account, nearly 80% of AI-bot activity in 2025 was gathering training data. A loop the conspiracy never predicted: machines read the web in order to write it.

Synthetic went mainstream — and few noticed

That this is more than a traffic statistic was shown by a 2024 study from Stanford’s Internet Observatory and Georgetown: over 120 Facebook pages churning out AI images — photorealistic children, invented scenes, the notorious ‘shrimp Jesus’ — drew hundreds of millions of engagements, with operators buying bot followers and funnelling audiences to content farms. The demographics matter: younger users learned to spot AI images; older ones, for whom a photograph was a lifelong proof of truth, did not. Meta began labelling synthetic content ‘Made with AI’ in May 2024 — the existence of the label says the problem had outgrown moderation. Then music: in the summer of 2025 a psychedelic band called The Velvet Sundown surfaced on Spotify with no concerts, no history, and over a million monthly listeners before being revealed as entirely AI-generated — earning, by one account, over 34,000 dollars in a single month. Deezer said it was receiving around 10,000 AI tracks a day, roughly a tenth of all uploads; Spotify said it removed over 75 million spam tracks in a year.

The click dies before the internet does

How are people responding? The hardest data comes from Pew Research, which tracked the real behaviour of 900 American adults across nearly 69,000 searches in March 2025. When an AI summary appeared, users clicked a link just 8% of the time, against 15% without one. They clicked a source cited inside the summary 1% of the time. And they more often ended the session entirely. The consequence was described in October 2025 by the Wikimedia Foundation itself: after correcting its bot detection, it found human pageviews of Wikipedia down about 8% year on year — because search engines increasingly answer directly, often from Wikipedia’s own data, without the click. The site often called the last good place on the internet is losing people, while its content feeds the models that divert them.

Where it is heading

Three trajectories are worth watching. The agentic web: Imperva named its 2026 report for it deliberately — AI agents that search, compare and buy on people’s behalf, blurring the line between legitimate and malicious automation. The risk of degeneration: a 2024 paper in Nature described ‘model collapse’, where models trained recursively on their own output degrade — which turns clean human data into a strategic resource. And a cultural shift: the head of OpenAI has publicly admitted he never took the dead internet theory seriously until he noticed how many social-media accounts now appear to be run by language models. When the phrase is uttered by the man whose company helped cause the flood, it has stopped being a conspiracy.

The twist: the dead internet has a living core

The most surprising number points the other way. Graphite also measured where the AI assistants themselves get their answers — and 82% of the articles ChatGPT and Perplexity cite were written by humans. AI-written articles, meanwhile, rank worse in search. Pew found the most-cited sources in AI summaries are Wikipedia, YouTube and Reddit — platforms built on human contribution. A paradox the theory never predicted: machines flooded the supply, but demand — from people and from the machines themselves — concentrates on verifiable human content. The AI writing half the internet goes to the other, human half for its facts.

What it means for brands

When the average content is free and infinite, it loses value — and the scarce asset becomes verifiable authenticity: first-hand data, named authors, dated sources, real faces. And because buying decisions are migrating from blue links into the answers of AI assistants, a new discipline decides who gets recommended: being the source a machine chooses to cite. That is the whole of GEO, and it is where we concentrate. The internet may be dying. What stays alive on it has never been worth more — and making sure a machine can find and cite your authentic signal is now part of the job.

Sources

Graphite, 2025 (65,000 articles, Common Crawl) · Ahrefs, April 2025 (900,000 pages) · Imperva/Thales Bad Bot Report 2025 & 2026 · Cloudflare Radar, via Matthew Prince, June 2026 · Pew Research Center, July 2025 · Wikimedia Foundation, October 2025 · Stanford Internet Observatory & Georgetown, 2024 · The Velvet Sundown coverage, 2025 (NBC, Marketplace, NME) · Shumailov et al., Nature, July 2024 · The Atlantic, September 2021 · Time (Sam Altman).