# What is lookalike modeling?

> 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.

_Last updated 2026-11-01. From The Personalization Guide, https://personalization.guide_

Lookalike modeling is a way to find new people who resemble an existing audience. You give the system
a seed audience, usually your best customers or recent converters, and it learns what separates that
group from the wider population, then returns the people who look most similar as prospects to reach.
It is one of the most common techniques in paid acquisition, and it is often sold under the
personalization banner, but it belongs to a different job.

## How it works

Under the hood a lookalike model is a classifier. The seed audience is the positive class, a broad
sample of everyone else is the negative class, and the model learns the patterns of attributes and
behavior that tell the two apart. Scoring the wider population with that model produces a ranked list,
and the highest-scoring people become the lookalike audience. Most platforms let you trade reach
against similarity: a tight lookalike is smaller and closer to the seed, a broad one is larger and
looser.

The work usually runs where the wider population lives, which means inside an ad platform or a data
clean room rather than on your own site. The seed is matched in, the modeling happens against the
platform's or partner's population, and the resulting audience is activated for targeting. Because the
raw populations stay on their own side, clean rooms have become a common place to build lookalikes
without sharing the underlying records. The mechanics of that data layer are covered in
[the data and identity layer](/data).

## Why it is not personalization

The distinction is worth drawing clearly, because the label hides it. Personalization adapts an
experience for a person you already have some data about: it changes what a known or in-session
visitor sees. Lookalike modeling does the opposite. It reaches outward to strangers who resemble your
customers, and it says nothing about what to show them once they arrive. Its output is an audience,
not a decision about content, ranking, or an offer for an individual.

In the terms this guide uses, lookalike modeling is acquisition and expansion, close to targeting,
while personalization fixes the person and varies the experience. The two are often run by the same
team with the same data, which is why they get conflated, but they answer different questions: who
should we try to reach, versus what should this person see.

## Strengths and limits

Its strength is leverage. A good seed and a large population can surface prospects that rules or
hand-built targeting would miss, and the method needs no understanding of why the seed behaves as it
does.

Its limits follow from the same mechanics. The audience is only as good as the seed: a biased or stale
seed produces a biased or stale lookalike, and a seed that is too small gives the classifier little to
learn from. Lookalikes also tend to reproduce whatever is already common in the seed, so they can
narrow rather than broaden who a brand reaches, and they can quietly encode sensitive attributes even
when those attributes are never named, which is the fairness concern that applies to any optimization
over population data. Because the modeling happens inside a platform, the advertiser usually cannot see
exactly how the audience was formed.

Treat a lookalike audience as a starting hypothesis about who to reach, measured like any other
acquisition tactic, and keep it separate from the personalization that decides what those people
experience once they are through the door.