What is Lookalike audience?
A lookalike audience is a set of people a platform believes resemble a seed audience the advertiser supplied. The model is only as good as the seed, and the percentage setting controls how far from it the platform is willing to go.
A model, not a list
A lookalike audience is the output of a model. The advertiser supplies a seed — a custom audience of some kind — and the platform finds people in a chosen market who resemble it by whatever similarity its own data supports. The term and the percentage scale below are Meta's; other platforms offer the same capability under their own names, with their own controls.
What "resemble" means is not disclosed and is not simple. It is not demographic matching. The features involved are behavioural and contextual and drawn from data the advertiser cannot see, which is exactly why the audience can be useful: it finds patterns nobody could have specified.
That also means the audience cannot be inspected or explained. An advertiser can describe the seed precisely and can say almost nothing about the result.
The seed is the whole of the input
Everything about the output follows from the seed, and most lookalike disappointments are seed problems.
Value. A seed of all customers describes the average customer. A seed of the top decile by lifetime contribution describes the customer the business wants more of. These produce visibly different audiences and visibly different acquisition. Where the data exists to build the second, it is almost always worth using.
Homogeneity. A seed containing two distinct kinds of customer — a discount-driven group and a premium one — asks the model to find people resembling the average of two populations, and the average of two populations frequently resembles neither. Splitting the seed and building two lookalikes produces better audiences than one combined model.
Recency. A seed built two years ago describes a business that may no longer exist in the same form: a different product mix, a different price point, a different market. Models built on stale seeds keep finding yesterday's customers.
Event depth. A seed of purchasers beats a seed of leads, which beats a seed of visitors, in exactly the order you would expect. Where purchase volume is too thin to seed a model, using the deepest event that has volume is the compromise, and it is a better one than padding a purchase seed with weaker events.
Size matters least of the five. Platforms state a minimum, and beyond a few thousand well-chosen people, more seed rarely improves the model as much as a better-chosen seed does.
What the percentage does
The percentage setting controls how far from the seed the platform will reach within the chosen country.
The narrowest setting is the closest match and the smallest pool. Widening relaxes the similarity threshold and grows the audience, roughly proportionally.
The instinct is to choose the narrowest on the theory that similar is better. That instinct is frequently wrong now, for the same reason tight targeting generally is: the delivery model selects inside whatever boundary it is given, and a narrow lookalike removes people it might have chosen.
A useful way to decide. If the budget is small relative to the audience, precision buys something — the campaign will only reach a fraction either way, and the fraction may as well be the most similar. If the budget is large relative to the audience, a narrow lookalike saturates quickly, frequency climbs, and a wider one performs better because there are more people to spend against.
Where they sit in an account
Lookalikes are a prospecting instrument. They find new people, which is what an account short of new customers needs.
They are not a retargeting instrument, and the common error of running a lookalike without excluding existing customers means paying prospecting prices to advertise to people who already bought.
They also stack usefully with the rest of the audience work. Exclude customers, exclude recent converters, exclude the retargeting pool — then let the lookalike and the delivery model work on what remains.
Why they have become less special
Two shifts have narrowed the gap between a lookalike and a broad audience.
The first is that delivery models improved. A platform given a broad audience and a conversion objective is doing something not entirely unlike building a lookalike continuously, from the conversions the campaign itself produces, and updating it as it goes. The lookalike's advantage — a head start from a seed — is worth most early, before the campaign has conversions of its own.
The second is that several platforms now accept a seed as a signal rather than a boundary. Instead of constraining delivery to the modelled audience, the seed informs the model and the campaign remains free to deliver outside it. That is usually the better arrangement, and where it is offered it largely supersedes the older form.
Running them well
Build several, from different seeds. High-value customers, recent purchasers, subscribers who convert. Each is a different hypothesis about who to find more of, and testing them against each other is a real test.
Exclude the seed and the customers. Otherwise the campaign spends part of its budget on people already in the business.
Make the comparison clean. Lookalikes from related seeds overlap heavily, and two overlapping audiences compared side by side do not produce an interpretable result. Mutual exclusion is the only version of that test worth running.
Refresh on a schedule, ideally automatically from the source system.
Judge on acquired value, not on cost per result. A lookalike from a high-value seed should be expected to cost more per conversion and to bring in better customers, and a report that only shows the first half will retire the better audience.