One AI Answer, Eight Brands: Designing a Benchmark Without Multiplying the Evidence

A recent AI benchmark for 8 luxury-jewelry brands revealed a discrepancy in visibility metrics. The study used an 'answer-once design' to collect 12 raw answers, which were then evaluated against the 8 brands. This approach helped to separate the model's answer from the brand-level judgment, providing a more accurate analysis. This distinction is crucial in AI visibility datasets, as it affects the collector, denominator, and publishable results. The study's findings highlight the importance of using multiple metrics and avoiding the conflation of model answers and brand judgments.