The Bayesian optimization process treats snippet edits as hypotheses, applying Thompson sampling to select content variations and iteratively update posterior win-rates based on LLM ranker feedback.