Brand Bias and Manipulation in LLM Recommendation Systems

A study investigating brand dynamics within Large Language Model (LLM) recommendation systems has uncovered a phenomenon termed "Conditional Monopoly." Well-known brands are recommended almost exclusively when product specifications are identical, but this dominance can be broken by even a small rating advantage (e.g., +0.1 stars) for a competitor. The research, conducted across commercial LLMs like GPT-4o-mini, Claude Sonnet, and Gemini 3 Flash, also demonstrates that "authority-style" marketing language, including potentially fabricated clinical evidence claims, can significantly influence recommendations. This manipulation can overcome the incumbent brand advantage, with each LLM responding differently to such tactics. Furthermore, the study highlights a "social dilemma" in multi-brand Generative Engine Optimization (GEO) competition. If all brands adopt the same optimization strategies, individual payoffs diminish drastically, and non-participating brands risk receiving no recommendations. This suggests GEO is not just a security risk but an emerging marketing practice shaping market competition.