AI is not just changing how people find information. It is weakening the business model that creates the information AI needs.
The internet has always had a bargain at its center. Publishers, bloggers, reviewers, analysts and small businesses made useful pages. Search engines sent them readers. Ads, subscriptions, leads or sales paid for the next round of work. AI search is now breaking that bargain by answering the question before the reader reaches the source.
That is why Bloomberg's recent warning about an AI death spiral matters. This is not only a media story. It is a startup story, an SEO story and, increasingly, an investment story. If fewer people click through to original work, fewer people can afford to create it. If fewer humans create useful material, future AI systems have less fresh human knowledge to learn from. The machine starts eating its own output.
The traffic data is already ugly. Axios reported in March that Chartbeat data showed traditional search referrals fell 60% over two years for small publishers, compared with 47% for medium publishers and 22% for large publishers. Page views from Google Search fell 34% from December 2024 to December 2025, while Google Discover fell 15%. ChatGPT referrals rose by more than 200%, but still accounted for less than 1% of publisher referral page views.
That last number is the one founders should sit with. AI platforms may create a new discovery channel, but today they are not replacing what they are taking away. A publisher cannot pay reporters with the idea that a chatbot mentioned its work somewhere inside an answer. A software startup cannot build a content engine around impressions it never sees. A local service business cannot optimize for a click that never happens.
Google's AI Overviews put the pressure in plain sight. A recent arXiv study of 55,393 trending queries across 19 categories found that AI Overviews appeared on 13.7% of all queries studied and 64.7% of question-form queries. That is exactly the territory where informational publishers used to win: how to, why does, what is, best way to, should I.
The same study found that 11% of the atomic claims in AI Overview responses were unsupported by the cited pages. It also found that well over half of cited pages carried display advertising, which means the publisher may supply the raw material while the economic value stays on the search results page. That is not a small product tweak. It changes who gets paid for knowledge.
Regulators have noticed. The Associated Press reported last week that the UK's Competition and Markets Authority ordered Google to give publishers tools to opt out of having their content used for AI Overviews, AI Mode and model fine-tuning, while still allowing them to remain in ordinary search. That distinction matters because publishers have long faced a brutal choice: accept scraping, or disappear from search.
For startups, the lesson is uncomfortable but useful. The old content playbook assumed that good SEO could compound for years. Write authoritative pages, collect backlinks, rank, convert. That still works in pockets, but it is no longer a dependable foundation by itself. Distribution now has to include direct audience channels, email, communities, partnerships, product-led loops and content that creates enough trust for readers to seek out the source.
The data problem comes next
The second part of the death spiral is quieter but more dangerous. AI systems are trained on the web. If the web fills with AI-written summaries, rewrites and low-cost content farms, future models will train on a weaker version of reality. The problem is not that synthetic data is always useless. It is that indiscriminate synthetic data can smooth away the strange, rare and specific details that make human knowledge valuable.
A 2024 Nature paper by Ilia Shumailov and co-authors called this model collapse. The researchers found that training generative models on recursively generated data can cause the tails of the original data distribution to disappear, producing irreversible defects over time. In ordinary language, the model forgets the unusual cases first. That is a serious problem because the unusual cases are often where expertise, creativity and real-world judgment live.
Newer research on retrieval collapse points to a similar risk for search and retrieval systems. One 2026 paper found that when synthetic content contaminated 67% of the source pool, exposure contamination rose above 80%, creating a web-grounded system that still appeared healthy while relying heavily on synthetic evidence. That is the trap. The answer can look clean while the underlying source diversity quietly shrinks.
This creates an odd market signal. Companies selling cheap AI content generation tools may be helping customers win short-term output races while making the broader information environment less valuable. The same goes for publishers that flood their sites with machine-written posts to chase search volume. It may work briefly. But if everyone does it, the reward pool gets smaller and the training pool gets worse.
The practical answer is not to stop using AI. That will not happen. The answer is to put a higher value on provenance, original reporting, first-party data, expert communities and human judgment that cannot be scraped from yesterday's summaries. Startups that can prove where information came from will have an advantage. So will publishers that build reader loyalty outside the search box.
What comes next is a fight over who owns the economic value of knowledge. If platforms keep the answer, publishers lose the click. If publishers stop producing original work, AI loses the raw material. The companies that understand both sides of that loop will be better positioned than those still treating AI search as just another SEO update.
Also read: Apple is trying to make Siri matter again by opening its AI stack • AI stock deals are testing how much Wall Street can absorb • Trump’s AI memo redraws the defense market
