The optimizer selects claims through a two-stage process that first samples a Rank Factor Class based on freshness to ensure diversity, and then applies a Bayesian acquisition method like Thompson sampling across the claims in that class.