Our hypothesis: Perplexity's heavy use of analytics domains (KenPom, Barttorvik) makes it more likely to surface statistical mismatches between seed and actual team quality. A 12-seed with elite defensive efficiency that's seeded low due to a weak conference often looks identical to a 5-seed by the numbers.

GPT-4o, which leans on national media coverage, may be more consensus-following - and therefore worse at identifying under-covered upsets. Gemini, with its broader data blend, likely sits between the two.

We track upset accuracy separately on the Upsets page. Round of 64 will be the first real test - First Four games don't produce classic upsets.