The engine optimizes content through an iterative multi-model loop—operating in either snippet or full-page mode—that seeds testable hypotheses, classifies them across a seven-category Rank Factor taxonomy (such as substance, alignment, and proof), and applies targeted rewrites. Guided by acquisition policies like Thompson sampling over Beta-distribution posteriors, an Ideator crafts variations while an LLM Ranker evaluates median ranking outcomes against competitive SERP sets to isolate what drives model preference.