DEJAN's Content Optimizer operates a multi-role experimental loop where initial hypotheses—termed claims—are generated by a seeder, categorized against a Rank Factor Class taxonomy, and modelled as Beta distributions representing win probabilities. During each optimization round, a two-stage acquisition strategy samples a diverse rank-factor class and selects a high-potential claim via policies like Thompson Sampling or UCB. An ideator model then generates a targeted candidate edit (operating in either snippet mode or page mode), which a dedicated judge model scores against a competitive set across repeated samples to determine median rank movement and reward. The system records the posterior probability shift of each hypothesis, tests for positive or negative rank factor patterns, and can evaluate the blast radius measurement to ensure improvements on one target query do not harm performance across related entities.