Definition
A lookalike audience is a platform feature that creates a new audience of people who share characteristics with a source audience. Meta's Business Help Centre, where the term originates, describes it as built from a source such as a custom audience of customers, website visitors or people who engaged with content; the advertiser selects the location the lookalike should cover and a size expressed as a share of the population in that location, within the documented range, and the platform identifies the people in that location most similar to the source. The registry's note captures the essentials: Meta builds it from a source audience such as customers or high-value visitors. Other platforms offer the same idea under their own names, and Google Ads retired its standalone similar-segments feature in favour of optimised targeting and audience expansion, which its help centre describes as automated extensions of a campaign's own signals rather than a separate audience object.
How it works
The platform takes the identifiers in the source, matches them to its own user accounts, and uses the attributes and behaviours it holds for those accounts to train a similarity model. It then scores the users in the chosen region and selects the top share requested. A smaller share is closer to the seed and smaller in ReachUnique people who saw an ad.; a larger share trades similarity for scale. The model is opaque to the advertiser, the help centre does not document which signals carry weight, and the audience refreshes as the source changes and as the platform's data updates. Meta documents a minimum source size and recommends a source made of the advertiser's best customers rather than everyone, because the model learns whatever the seed has in common.
| Seed type | What similarity is based on | What to watch |
|---|---|---|
| Customer list (hashed) | Accounts matched to the uploaded identifiers | Match rate; recency and value of the customers included |
| Website or app audience | People whose tracked events met the rule | Consent coverage; the event chosen defines the seed's quality |
| Engagement audience | People who interacted on the platform's surfaces | Large but shallow; interest is not intent |
| Value-based source | Customer list or events carrying a value per person | Value data must be accurate for the model to weight it |
The lookalike audiences guide covers seed choice, the size trade-off, value-based lookalikes and how they interact with the platform's automated audience options; the custom audiences guide explains how the sources themselves are built.
In practice
Lookalikes are a prospecting tool, used when the warm audiences from RemarketingAdvertising to previous visitors or engagers. are too small to spend against and interest targeting is too blunt. The working sequence is to build a seed from the outcome that matters, usually purchasers or qualified leads over a recent period, exclude existing customers and current visitors from the campaign so the lookalike reaches only new people, start at a small share and widen as results hold. Seeds can also come from analytics: audiences built in Google Analytics and shared to Google Ads, as the GA4 audiences guide describes, play the same role on Google's side even though the expansion mechanism there is now automated rather than a named lookalike.
Two seeds, two outcomes
A subscription brand builds two lookalikes for the United Kingdom at a 1% size, each around 440,000 people. Seed A is all 95,000 website visitors from the last 30 days; seed B is the 6,200 customers who completed a paid signup in the last 90 days. Both campaigns run for four weeks at £8,000 each with purchasers and recent visitors excluded. Seed A delivers 230 signups at a CPA of £34.80; seed B delivers 410 at £19.50, and its customers show a 23% higher three-month retention rate. The brand widens seed B to a 3% size, where CPA rises to £24.10 but volume more than doubles, and retires seed A.
What it is not
A lookalike audience is not RemarketingAdvertising to previous visitors or engagers.. The source list is made of people the brand knows; the lookalike is made of people it does not, chosen for their resemblance, and the two must not be confused when reading results or setting exclusions. It is not interest or demographic targeting, which the advertiser defines by rule; a lookalike is defined by the platform's model, and the advertiser cannot inspect the rule. It is not a guarantee of similarity in the dimension that matters: the model finds whatever the seed shares, which may be a demographic coincidence rather than purchase intent. It is not the same as the platforms' automated audience expansion features, which widen a campaign's targeting using its own conversion signals rather than a supplied seed. And it is not portable: a lookalike built on one platform cannot be exported to another.
Related terms
RemarketingAdvertising to previous visitors or engagers. supplies the seed and defines the exclusions, so the two features are configured together even though they reach different people. CPACost of each conversion. is the usual yardstick for a lookalike test, read alongside downstream quality such as retention or lead acceptance, because a cheap acquisition from a shallow seed can be a poor customer. ReachUnique people who saw an ad. is what the size setting trades against similarity, and the right size is found by widening until marginal results fall below the target. The lookalike audiences guide covers the decisions in order, the custom audiences guide the construction of good seeds, and the GA4 audiences guide the analytics-side equivalent of a seed shared to Google Ads. Incrementality testing belongs in the same conversation, since a lookalike that reaches people already drifting toward the brand can look efficient without adding anything a holdout would confirm.
Common mistakes
- Seeding with everyone. A seed of all visitors teaches the model what a visitor looks like, not what a customer looks like.
- Skipping exclusions. Without excluding customers and warm visitors, the lookalike's results include people remarketing would have reached anyway.
- Judging on CPA alone. Check what the acquired customers do afterwards; a cheap seed can buy cheap customers.
- Stale seeds. A customer list from years ago describes a past customer base; refresh the source regularly.
- Starting wide. Begin at a small size to learn whether the seed works before trading similarity for reach.
- Running several lookalikes against each other. Lookalikes from similar seeds overlap heavily, so parallel ad sets bid against themselves and inflate frequency.
- Trusting the platform's audience estimate as reach. The estimate describes the audience object, not the people the campaign will actually serve at its budget.
