LLMs favor well-known brands until reviews tip the scale

Testing GPT-4o-mini, Claude Sonnet and Gemini on skincare product recommendations, researchers Xi Chu and Yupeng Hou found that when competing products were otherwise identical, the well-known brand was recommended 100% of the time — but a rival needed less than a 0.1-star ratings edge to overturn that advantage, per their paper posted to arXiv. When every brand adopted the same marketing tactics, the incumbent’s edge collapsed from a payoff of +0.802 to +0.007, and brands that opted out received no recommendations at all.