What's this about?
Researchers asked two web-grounded Perplexity models for the best products in 380 software categories — 760 calls, categories written before any results were seen — and kept every URL retrieved. Of 7,534 citations, 59.8% point at domains ranked worse than #100,000 in the Tranco top-1M list, and 23.4% at domains not in the top million at all. Two of the sites doing the grounding have given their homepage the HTML title "Facts & Grounding Page" — grounding being the retrieval step these models perform — and those two plus a third under apparently common control have published 215,128 machine-generated "best <category>" pages. None of the three domains existed before December 2023.
What is "grounding"?
The retrieval step: before answering, the model searches the web and reads pages, so its answer is anchored to sources it can cite rather than to memory. It is the standard fix for hallucination and it works — but it moves the attack surface from the model to the index. If you can get a page into what the model retrieves, you have influenced the answer without touching the model at all. Naming your homepage "Facts & Grounding Page" is that strategy stated out loud.
What's the community saying?
- The question the thread keeps circling: what protection do LLM search engines have against retrieving and training on content generated by other LLMs? Nobody has an answer.
- The pessimistic consensus: we are probably already at the point where AI-generated sites make up the majority of what most people see, whatever the true share of the web is.
- A resigned observation about incentives: this is a problem that will not blow up within a few quarters, so it is of no commercial interest.
- One commenter's response is to go back to the library, on the theory that most technical books predate 2020 and physically publishing slop is not profitable enough to flood that market.
- The practical complaint underneath it all: searching for products has become impossible unless you already know what you are looking for.