# What is personalization?

> Personalization is a closed decision loop. A system chooses among alternatives (content, ranking, offer, layout, timing) for a specific person or context, using data about them, and its value can only be known through controlled measurement.

_Last updated 2026-10-30. From The Personalization Guide, https://personalization.guide_

Personalization is a closed decision loop. A system chooses among alternatives, such as content,
ranking, an offer, a message, the timing, or the layout, for a specific person or context, using
data about them, and its value can only be known through controlled measurement. That is the whole
idea. Most of the confusion in the field comes from treating one part of the loop, whether
segmentation, targeting, recommendation, or testing, as if it were the whole thing.

The word promises something intimate: a system that knows you. But every personalization system is a
model of a person, and every model has an owner and a purpose. Being precise about the loop is what
separates a real capability from a relabeled one, and it is what the rest of this guide builds on.

## The loop in five questions

Every personalization system answers five questions. Naming them is the fastest way to see what a
given system actually does, and to tell two systems apart when both are sold under the same word.

- **Unit: who is the decision for?** Everyone, a segment, an account, an individual, or a single
  session. This is what separates segmentation from individualization, and B2B from B2C.
- **Inputs: what data does it read?** A declared profile, behavioral events, the current context,
  content metadata, and sometimes second-party or third-party data. The inputs set the identity and
  consent requirements.
- **Decider: what makes the choice?** A human-written rule, a statistical model, an LLM or agent, or
  the user. As the decider changes, so do the explainability, the authoring effort, and how fast the
  system can adapt.
- **Action and surface: what changes, and where?** The content, the ranking, an offer, a price, a
  message, the timing, the channel, or the layout, delivered on the web, in an app, in search, in
  email, in ads, in pricing, in support, or inside an AI assistant.
- **Feedback: how is the effect measured?** Through attribution, before-and-after, an A/B test, a
  persistent holdout, uplift modeling, or a geo experiment. Without this step, a change is a belief,
  not a result.

## The axes that actually differ

Two systems can both be called personalization and share almost nothing. What makes them different is
where they sit on a handful of axes.

| Axis | Range | Why it matters |
|---|---|---|
| Decision logic | Rules, then statistical scoring, then online learning, then LLM or agent planning | Explainability and how fast it adapts |
| Freshness | Batch (daily or hourly), near-real-time (minutes), in-session (sub-second) | Whether it can respond to intent expressed now |
| Decision unit | Everyone, segment, account, individual, session | Separates segmentation from individualization |
| Data read | Profile, events, context, catalog metadata, outside data | Identity needs and consent requirements |
| Surface | Web and app slots, search, recommendations, email, ads, pricing, support, assistants | The integration pattern |
| Measurement | Attribution, before and after, A/B, persistent holdout, uplift, geo | Whether any claimed effect is causal |

These axes are independent. Simple rules can run in-session, and sophisticated machine learning can
run in overnight batch. One further axis, the owner, explains much of the market: marketing, product,
engineering, data science, and customer service each grew their own personalization tools, which is
why a single company often runs several unrelated personalization systems at once.

## What personalization is not

A reference should be precise about the neighbors, because vendor content uses the words
interchangeably. In short:

- **Segmentation** groups people by shared attributes or behavior. By itself it changes nothing.
- **Targeting** fixes the message and varies the audience; personalization fixes the person and
  varies the content.
- **Recommendation** is one mechanism of personalization, a ranked set of items; a "most popular"
  list is recommendation with no personalization.
- **Customization** is a change the user makes through explicit settings; personalization is a change
  the system makes from data about them.

The full set of distinctions, each with a test that tells it apart, is on
[personalization vs segmentation, targeting, recommendation, and customization](/library/comparisons).

## Why the definition matters

Taking the loop seriously has three consequences that shape everything else in this guide.

**The hard part is usually the data, not the algorithm.** A model fed day-old segments cannot react
to what a visitor is doing now, and when identity is fragmented the individual being personalized to
is really several partial profiles. This is where personalization most often fails, and it gets its
own page: [the data and identity layer](/data).

**The value is only as real as the measurement.** Most published lift figures are vendor case studies
or consultancy estimates; credible evidence comes from randomized experiments with a persistent
holdout. How to read a lift claim is on [how personalization is measured](/measurement).

**The decider keeps changing, and so does the risk.** As the choice moves from rules to models to
generative agents, the system can produce the artifact itself rather than select from a fixed set,
which sharpens old questions about opacity and manipulation. The methods are in
[how personalization works](/library/techniques), and the ethics are in
[personalization and its critics](/ethics).

The same loop has been rebuilt many times as the technology changed, from the first collaborative
filters of the 1990s to today's generative systems. That story is on
[where personalization comes from](/history).