Does synthetic content in search results hide ecosystem decay?

Model collapse is usually told as a training story: feed a model its own outputs and the distribution's tails vanish. This paper relocates the same dynamic to retrieval, where it bites sooner and more quietly. Retrieval Collapse is a two-stage failure. First, Dominance and Homogenization: high-quality SEO-style synthetic content captures the top results and erodes source diversity. Second, System Corruption: low-quality or adversarial synthetic content infiltrates the pipeline. The damning detail is that in the SEO scenario, 67% pool contamination produced over 80% exposure contamination while answer accuracy stayed stable — a "deceptively healthy" state where the metric you watch looks fine while the ecosystem you depend on decays.

This is the retrieval-side completion of an arc the vault already traces. Since Does training on AI-generated content permanently degrade model quality? establishes the mechanism in model training, and since How much of the internet is AI-generated now? documents that the web is already heavily synthetic with diversity (not accuracy) as the first casualty, this paper shows the same diversity loss propagating through RAG: the pipeline silently shifts toward synthetic evidence, and a self-reinforcing decline becomes possible because tomorrow's models train on what today's retrievers surfaced.

The defensive findings cut both ways. Under adversarial contamination, BM25 exposed ~19% of harmful content while LLM-based rankers suppressed more — but at high compute cost. So the resilient ranker is the expensive one, and the cheap scalable baseline is the vulnerable one, which is exactly the wrong cost gradient for web-scale deployment. The proposed answer is Defensive Ranking that jointly optimizes relevance, factuality, and provenance — provenance being the variable that homogenization destroys first.

The strongest counterargument: stable accuracy might mean the synthetic content is simply good enough, and diversity loss is an aesthetic worry, not a functional one. The rebuttal is fragility — high accuracy resting on a monoculture has no fallback when the monoculture is wrong or poisoned, which the adversarial scenario demonstrates.