Techniques

Rule-based personalization is personalization by hand-written conditions: a person writes if-then rules that decide what each visitor sees. It is transparent, fast to set up, and easy to measure, which is why it is still the most common form, but it never learns and it does not scale to every case.

Collaborative filtering is a personalization technique that predicts what a person will like from the behavior of many other people, rather than from any description of the items themselves. It powers much of recommendation, and it works best paired with content and context.

Content-based filtering recommends items whose attributes or embeddings resemble what a person has already engaged with. It reads the content rather than the crowd, so it handles new items and explains itself, but its results drift toward sameness.

A hybrid recommender combines more than one technique, usually collaborative filtering, content, context, and rules, arranged in stages. Retrieval narrows a large catalog to a shortlist, a heavier model ranks it, and re-ranking applies diversity, inventory, and business rules. Almost every production recommender and search ranker works this way.

A bandit splits traffic among options while it learns, balancing trying new options against using the current best. A contextual bandit also reads features of the user and the moment, which makes it a form of in-session personalization.

Propensity scoring is a supervised model that predicts the probability a person takes an action, such as a purchase, a click, or churn. It is widely used to prioritize who to target, but it answers what will happen, not what a treatment changes, so it can spend budget on people who would have acted anyway.

Uplift modeling predicts how much a treatment changes an individual's behavior, the causal effect rather than the raw likelihood of an outcome. It sorts people into persuadables, sure things, lost causes, and sleeping dogs, and it targets only the persuadables.

Next-best-action is a personalization method that arbitrates among the actions a person is eligible for at a moment, scoring each on value and picking the best one, where doing nothing is always an option.

Lookalike modeling trains a classifier to tell a seed audience apart from a wider population, then selects the closest matches as new prospects. It is an audience-expansion technique that runs in ad platforms and clean rooms, and despite the personalization label its goal is acquisition, not tailoring for people you already know.

Reinforcement learning extends bandits to sequences of decisions, choosing each action for its long-term value rather than the next click. In marketing it is more often promised than practiced, because rewards are sparse and delayed and the feedback loop is hard to close.

Generative personalization uses a model to produce the tailored artifact, such as a product description, an email, or an interface, rather than selecting it from a fixed inventory. It appears in three distinct patterns, and it moves the hard part from writing content to selecting and measuring it.

Agentic personalization is personalization run by an LLM agent. Given a goal, a set of tools, and guardrails, it plans and carries out the steps itself, rather than executing a flow a person designed. The label is used loosely, so treat it as a claim to test.