For years, the Dead Internet Theory was a conspiratorial meme: the idea that a large part of what circulates online isn't made by people, but by bots, content farms and generative AI. In 2026, we have enough data to stop calling it a theory.
In April 2025, the Imperva Bad Bot Report confirmed that in 2024, for the first time in a decade, automated traffic overtook human traffic: 51% of global web traffic is now bots. In parallel, a Graphite study published in October 2025 found that 52% of articles published on the web are now AI-generated. And another from Ahrefs on 900,000 new pages indicates that 74.2% of web pages created in April 2025 contain AI content to some extent.
The internet isn't dead in the literal sense. But a good portion of what circulates on it isn't made by anyone, nor is it meant to be read by humans. And that proportion is accelerating.
What the theory is, briefly
The original idea was that online content and interactions are dominated by bots, scripts and AI. Likes that aren't likes, comments no one has written, articles no one has read before publishing.
Without invoking any conspiracy, the mechanics are simple: generating automated content and interactions is so cheap that a lot of people do it. System incentives — pay-per-click advertising, algorithms that reward volume, platforms that measure engagement — reward it.
No malevolent architect is needed. It's enough to let economic incentives and AI tools do their job. The theory describes an emergent property of the system, not a conspiracy.
2024 was the turning point
Three independent figures mark the shift, all published in 2025 from 2024-2025 activity:
1. Bots overtake humans in web traffic. Imperva, in its Bad Bot Report 2025, confirms that in 2024 automated traffic reached 51% of the total. Malicious bots (excluding legitimate ones from search engines) rose to 37%, the sixth consecutive year of growth. Generative AI plays a direct role: it has lowered the cost of building bots so much that unsophisticated actors can launch large-scale attacks.
2. More than half of published content is now AI-generated. Graphite analysed 65,000 English articles published between January 2020 and May 2025: 52% are AI-generated in the latest period. Ahrefs, on 900,000 new pages from April 2025, finds 74.2% with some degree of AI content. The proportion has stabilised in recent months, probably because Google has started penalising purely synthetic content.
3. Entire websites built by algorithms. Joint research from Stanford, Imperial College London and the Internet Archive estimates that 35% of new websites launched in mid-2025 are AI-generated or assisted, often without real human supervision.
Combine these three figures and the picture is clear: a meaningful portion of the noise you see when you browse isn't made by anyone. And those who do make it often don't read it before publishing.
Google no longer sends as many clicks
The most visible effect of this shift isn't so much in the content as in how people access it.
Google launched AI Overviews — AI summaries at the top of results — in May 2024. In January 2025 they appeared on 6.49% of searches. By March 2025 they were already at 13.14%. And they keep growing.
The traffic impact is well documented:
- Ahrefs (December 2025): when an AI Overview appears, top organic results lose 58% of clicks compared with December 2023.
- Pew Research: with an AI Overview present, users click on any result only 8% of the time. Without one, 15%.
- Seer Interactive: organic CTR drops between 49% and 65% when there's an AI Overview.
According to a Semrush study of 10 million keywords, 88.1% of searches that trigger an AI Overview are informational. That is, exactly the kind of search that historically funnelled people to blogs and Wikipedia.
Translation: a large share of informational traffic no longer leaves the search engine. It stays inside. And the human content that does fall outside the Overview competes for the remaining clicks against synthetic content optimised exactly for those same algorithms.
What you see every day
If you're not convinced this is accelerating, check whether you recognise any of these symptoms:
- The top 10 Google results all look like the same article. Same structure, same bullet points, interchangeable conclusions.
- Comments that are too polished, too fast, too generic. Especially on LinkedIn. "Great post!" in flawless English under an English conversation, three seconds after posting.
- Traffic spikes in any analytics that don't add up: 3 AM visits, 100% bounce, 0-second session duration, from countries that don't match the project.
- Long articles that say nothing. 500 words to reach a point that fits in a sentence. All SEO filler in between.
- Sites with hundreds of articles and no identifiable author. AI-generated profile photos, vague bios, non-existent brand. Zombie sites or MFA (Made For Ads).
None of these signs is proof on its own. All together, yes.
The layer no one was watching: model collapse
There's a deeper dimension to this story, and it directly affects the future of AI itself.
When AI models are trained massively on internet data — which is what they do — and a growing proportion of that data is itself AI output, the system enters a kind of feedback loop. A study published in Nature in 2024, signed by Oxford and Cambridge researchers, gave the phenomenon a technical name: model collapse.
The conclusion: when models feed mostly on content generated by other models, quality degrades quickly. After a few training cycles, outputs become less accurate, more homogeneous and progressively absurd. Minority voices, under-covered topics and divergent viewpoints are the first to disappear, because they have less statistical representation in training.
Put another way: if AI ends up dominating online content creation while training mostly on online content, it is doomed to get worse.
That's why big tech has started moving pieces: paying Reddit, Stack Overflow, X or specific publications for licensed access to original "human-made" data. Authentic human content is the scarce resource sustaining the system, and it's already being treated as such. The paradox is obvious: the same industry that contributed to flooding the internet with synthetic content now has to protect itself from what it created.
Who wins and who loses
In this new scenario the winners aren't those who publish the most. They're someone else.
Winners:
- AI models and their owners, who keep most informational traffic without having to send it anywhere.
- Platforms with proprietary, walled data — Reddit, X, Discord, private communities — that now sell access to their data to train models.
- Brands with authority established before the boom, which are still referenced by AI as reliable sources.
- Direct channels — newsletters, podcasts, private communities — where no algorithm gets in the way.
Losers:
- Media outlets dependent on Google's open traffic. Several reports show drops of 30% to 70% in specialised publications.
- Blogs and informational sites that lived off answering general questions (how to do X, what is Y).
- New voices trying to start from zero with classic SEO strategy.
- In the long run, AI systems themselves, if they don't solve the problem of training without cannibalising themselves.
The GEO strategy and what's starting to move
Part of the industry has started talking about GEO (Generative Engine Optimization) as the natural evolution of SEO. The idea: if most users no longer click, the goal stops being to appear in the top 10 and becomes to become a source cited by AIs.
This requires different things from classic SEO: very clear semantic structure, verifiable data, explicit authorship, original content that can't be derived from other sources. The paradox is beautiful: what makes an AI cite you is exactly what makes content human and useful. The cure for being cited by the machine is to stop thinking like a machine.
Some platforms are making more structural moves. Cloudflare has launched Pay per crawl, a system for publishers to charge the AI bots that read their content. The idea — that human content gets paid to be training data — is the antithesis of the free model that sustained the internet for thirty years. But it's coherent with a world where human content is a scarce resource.
It's also possible that within a few years we'll see human-authorship certifications, licensing contracts between creators and AI companies, marketplaces for verified human data. The internet may end up splitting into two layers: a public, synthetic, free and progressively useless one, and a private, human, licensed and expensive one.
What to take away
The Dead Internet isn't a catastrophe. It's a change of scenery.
What used to work — volume, generic optimisation, "correct" content to hit the top 10 — no longer works, because the system is saturated with synthetic noise and Google has changed the rules of the game. What now works is the opposite: less content, more authority, more verifiable humanity, more owned channels.
In the medium term, the question isn't whether the internet dies or not. It's who has the right to the living part. Platforms with proprietary data, creators with a recognisable voice, and communities with direct human relationships. The rest — blogspam, generic SEO, the correct articles no one reads before publishing them — is already replaceable. It probably already has been replaced.
And as users, writers or builders of the internet, what we choose to publish (or not publish) over the coming years will be part of the collective answer to that question.