Anatomy of the decision loop
Every engine, whatever its marketing, runs the same four-stage loop. Ingest: signals arrive — behavioral events, context, profile data if any exists. Decide: the engine maps those signals to an experience, by rule evaluation, model inference, or behavioral scoring. Act: the chosen variant renders on the page, ideally before first paint so there is no flicker. Measure: outcomes flow back, ideally against a randomized holdout so the engine's contribution is proven rather than assumed. Weakness at any stage caps the whole loop — most engines are strong at deciding and weak at acting fast or measuring honestly.
The label hides the architecture
"Personalization engine" gets applied to three very different builds. A suite feature: the rule evaluator inside a marketing cloud, powerful if you've adopted the surrounding ecosystem and inert if you haven't. A generative optimizer: AI producing and testing its own variants, strong on volume, usually thin on knowing who the visitor is. And a decision layer: a standalone runtime that scores in-session behavior and chooses per visitor, independent of any suite.
The questions that separate them: Where does the decision run, and how fast? Does it work on the anonymous majority of traffic or only resolved identities? Is the live path deterministic, or is a model improvising per pageview? Is lift holdout-proven? And can an agent operate it — is there a real tool surface, or only a UI? The comparison pages walk these questions vendor by vendor.