Contents6
The image of a web that “eats itself” brings together two different phenomena. The first: sites produce AI-generated content in bulk, and the next models learn from that content. The second: answer engines read pages in place of readers, who no longer always need to click through to the sites the information comes from.
Both are real. Neither condemns the web. They change what a site has to offer to deserve being read, cited or visited.
What the data shows
The impoverishment of models is demonstrated, under controlled conditions. In July 2024, Ilia Shumailov (University of Oxford) and his co-authors published in Nature the study “AI models collapse when trained on recursively generated data”. It shows that when a model is trained in a loop, indiscriminately, on content generated by models, it suffers irreversible defects: the “tails” of the original distribution disappear. Put simply, rare cases fade out first; and those rare cases, the authors remind us, often concern marginalised groups. A correction by the authors, published on 21 March 2025, fixes a notation in the theoretical part; it does not change this conclusion.
One important limit: this is controlled recursive training, carried out notably on a small language model (OPT-125m, published by Meta), fine-tuned at each generation on the texts produced by the previous one. Production models mix human and synthetic data, and the scale of the effect on them has not been established.
AI assistants are becoming a gateway to information. The Reuters Institute’s Digital News Report 2025, which covers 48 markets, asked the question for the first time:
of respondents use an AI chatbot every week to get their news
of under-25s use an AI chatbot every week to get their news
The survey was conducted online by YouGov, from mid-January to the end of February 2025 (methodology). It is still a minority, twice as common among the under-25s, and too recent to speak of a trend.
What remains to be seen is whether these readers still click through to the sites. It is also the question raised by the summaries Google displays at the top of its results, which we cover in our analysis of AI Overviews.
Position 1: the web is exhausting itself
The first camp describes a spiral.
The web exhausting itself, according to the first camp
- Abundance
Generated content costs almost nothing to produce, so it floods search results.
- Traffic
Answer engines summarise pages without sending visitors, so the sites that produce the information lose the traffic that financed them.
- Impoverishment
Less revenue, less original content, more generated content; the web grows poorer, and so do the models that feed on it.
This reasoning has in its favour the Nature study and a concern the Reuters Institute notes among news publishers: as the big platforms integrate AI-generated summaries, they fear losing yet more traffic to their sites and apps. What counts against it is that it assumes nothing will come along to correct the trend.
Position 2: mediocre content has always existed
The second camp points out that content farms, text written for robots and pages copied from other pages existed long before AI, and that search engines have always had to sort. The authors of the Nature study remind us of this themselves: content farms led search engines to change their algorithms, and Google demoted mass-produced articles. AI labs, for their part, have every interest in identifying quality human data: the same study concludes that data collected from genuine human interactions with these systems will become more valuable.
This camp is right about history. It may underestimate the scale, and again it is the Nature study that says so:
What is different with the arrival of large language models is the scale at which such poisoning can happen once it is automated.
What we think
We operate a multilingual content engine in production, in six languages, and we apply the fixes we recommend to this site. Our reading agrees with both camps at once: it has become very easy to publish a page, and much harder to publish one that serves a purpose.
Our view fits in one sentence: a page that contains only what a model would have written on its own has no reason to be cited, whether by Google, by an answer engine or by a reader. The model already knows how to produce the average. What it does not know is what happened at your company: a measurement you took, a mistake you corrected, a price you charge, an opinion you stand behind.
From the field
The robots.txt served was not the one in the repository
The audit of Palmora Property, a property listings portal in Thailand that we built and operate, showed that the robots.txt served by the CDN blocked the whole site to GPTBot, ClaudeBot, CCBot and Google-Extended, although nothing of the kind appeared in its repository.
A finding like this, specific to one site and verifiable, is precisely what a model cannot invent. We tell the story in detail in our article on getting cited by ChatGPT.
What would change our mind: if answer engines stopped citing their sources altogether, or if it were shown that production models really do degrade because of synthetic content. Neither is established as of October 2026.
What this changes in practice for a website
- Publish less, but publish well. Ten pages with figures you have measured are worth more than a hundred reworded pages. AI can help with the writing; it must not supply the substance, and human review remains non-negotiable, as we explain in what an AI SEO agency should really deliver.
- Source and date. A claim with its primary source and its date is easier to cite than a vague one, for a reader as for an assistant.
- Answer straight away. A page that answers its question from the first lines is the one an answer engine can pick up, with a link.
- Decide who reads the site. Opening or closing your content to AI crawlers is a legitimate editorial choice. Having it imposed on you without knowing it is not, and that is what was happening on Palmora Property. The llms.txt file helps agents that read a site on demand, but no major engine says it uses it to choose whom to cite.
- Measure what comes from assistants. The traffic sent by assistants is tracked separately from classic search traffic, along with the other readings in our protocol for measuring visibility in AI.
And if your team knows things that no model can write in its place, our SEO and GEO content engine starts from there: we prepare the drafts, you review and approve them before anything is published.
Frequently asked questions
Is AI-generated content penalised by Google?
Not as such. In its guidance published on 8 February 2023, Google writes that it focuses on the quality of content rather than on how it is produced. Its spam policies, on the other hand, target the production of many pages to manipulate ranking, for example with AI, with no value for users: a site that does this can be ranked lower, or even removed from the results.
What is AI model collapse?
Model collapse is the name Ilia Shumailov and his co-authors give, in their study published in Nature in 2024, to the degradation of a model trained, generation after generation, on data produced by models. They distinguish two stages: first, the model loses information about rare cases; then, it converges on a distribution that looks little like the original, often far less varied.
Is model collapse already under way?
Nothing establishes it as of October 2026. The phenomenon is demonstrated in the laboratory, when a model is retrained in a loop, indiscriminately, on generated data. In the same study, keeping 10% of the original data at each generation led to only a minor degradation in performance. Production models mix human and synthetic data, and the scale of the effect on them has not been established.
Sources
- “AI models collapse when trained on recursively generated data” (new tab)doi.org
- Correction by the authors (new tab)doi.org
- Digital News Report 2025 (new tab)reutersinstitute.politics.ox.ac.uk
- Methodology (new tab)reutersinstitute.politics.ox.ac.uk
- Guidance published on 8 February 2023 (new tab)developers.google.com
- Spam policies (new tab)developers.google.com
Jim · founder of Brain Plus AI
Also in seo and ai visibility
Practical guide7 min
How to get cited by ChatGPT and AI search engines
What actually gets a site cited by ChatGPT, Claude or Perplexity: let the bots read it, answer straight away, source every fact, mark up only what is shown. And how to measure it without fooling yourself.
Understand6 min
llms.txt: what it is really for
The llms.txt file promises to help AI read your site. What the proposal says, what can be verified about real usage, what Google, OpenAI and Anthropic say about it, and why we publish one anyway.
Stepping back7 min
AI and disinformation: lying machine or remedy?
One study shows that AI writes more convincing false information than humans do, another that it reduces conspiracy beliefs, but its journal has expressed a reservation. What we can take from them, and what the law requires of those who publish.