The competition between content providers who can craft webpage text to manipulate an LLM-based search engine's rankings can be modeled as an infinitely repeated prisoners' dilemma, where each provider chooses each round to refrain from or launch a ranking manipulation attack. Incorporating stochastic attack success rates, attack costs, discounting of future profit, and market degradation reveals when cooperation is sustainable and why some defenses that only cap attack success rates fail or even encourage attacks.
LLM-based search engines answer queries with the Retrieval-Augmented Generation (RAG) pattern: they retrieve documents and feed them, with the query, into a language model that ranks and summarizes t…