Lookalike Audience
A lookalike audience is a new group of prospects that an ad platform builds by finding people who resemble a source list of your existing customers or leads.
Key takeaways
- The platform ranks a population by similarity to matched seed profiles.
- A one percent tier is closest and smallest; ten percent is looser.
- Seed size, recency and homogeneity matter more than any targeting setting.
- Mixed seeds average two customer types and describe neither of them.
- Lookalikes inherit the acquisition bias of the list they were built from.
In depth
Building a lookalike starts with a seed list of identifiers, which the platform matches to its own accounts. It then compares those matched profiles against everyone else in a chosen country and ranks the population by similarity across thousands of behavioural and demographic signals you never see. You pick a percentage cut-off: the top one percent is the closest match and the smallest reach, ten percent is looser and much larger. The audience is scoped to one country and refreshed on a schedule.
Three properties of the seed decide the outcome: size, recency and homogeneity. Below roughly a thousand matched members, the model has too little to generalise from; far above that, extra members add little unless they are genuinely similar. A seed mixing enterprise buyers and free-tier signups produces an average of two different people, which describes neither. Recency matters because behaviour drifts, so a seed built two years ago quietly encodes a market that has moved on.
In practice teams run several percentage tiers as separate ad sets and let spend settle where the economics work, excluding the tighter tier from the looser one so they do not compete. Seeds get rebuilt monthly from the most recent qualifying customers. A scorecard funnel makes the rebuild almost automatic, since each new cohort of top-tier respondents is already labelled and can be exported as the next seed without anyone deciding by hand who counts as a good lead.
A lookalike can only reproduce the customers you already found, so it inherits every bias in how you acquired them. If early growth came from one channel, the model will keep pointing back at that channel's audience and away from segments you never reached. It also cannot help a brand-new category with no customer list. And as delivery algorithms improve at broad targeting, a well-fed conversion campaign often finds the same people without a seed at all.
Example in practice
How to measure it
Compare each lookalike tier against a control audience running the same creative and offer, because a lookalike that beats nothing proves nothing. The numbers to line up are cost per qualified lead and the share of leads that reach a sales conversation. A tier that produces cheap leads with a low qualification rate is finding people who look similar on the surface only.
Check the seed itself as well as the campaign. The match rate, meaning how many uploaded records the platform recognised, sets a ceiling on quality; a low rate usually points at stale or badly formatted contact data. Then track whether the audience is still growing between refreshes. Over several months, compare the value of customers acquired through each tier, since a slightly dearer lead from a tighter tier can still be the cheaper customer.
Common mistakes
Building the seed once and never touching it again is the quietest failure. The audience keeps running, results drift down over months, and nobody connects the decline to a list assembled before the product changed. Put a rebuild on the calendar, use only customers acquired in a recent window, and keep the old audience running alongside the new one for a fortnight so the comparison is real rather than assumed.
Another common error is stacking a lookalike with heavy interest and demographic filters. The model already selected for similarity, so adding age brackets and job titles removes people it identified as promising and shrinks the audience below a workable size. If you distrust the audience, fix the seed rather than the filters. The one restriction worth keeping is excluding existing customers, which the seed itself will otherwise pull straight back in.
Frequently asked questions
How big should my seed audience be?
Most platforms recommend a seed of at least 1,000 to a few thousand quality records, but cleanliness matters more than size. A focused list of your best customers outperforms a large list padded with low-value contacts.
What makes a good lookalike seed for a quiz funnel?
Your highest-scoring, sales-qualified quiz leads make an excellent seed because the scorecard has already filtered for fit. Seeding from them teaches the ad platform to find more people who match your ideal customer profile.
How is a lookalike different from retargeting?
Retargeting re-engages people who already know you, while a lookalike finds brand-new prospects who resemble your best existing audience. Lookalikes expand reach at the top of the funnel, whereas retargeting recovers conversions further down.
Which percentage tier should I choose?
Start at one percent to see whether the seed carries any signal, then widen only if the results hold. If the tightest tier performs no better than broad targeting, the seed is the problem and a wider tier will not fix it. Larger budgets need larger tiers simply to avoid over-serving, so the right tier is partly a function of how much you intend to spend.
How often should I rebuild a lookalike audience?
Monthly is a reasonable default when new qualifying customers arrive steadily, and quarterly when they trickle. Rebuild sooner after anything that changes who buys: a price change, a new product tier, entry into a different market. Keep the previous audience live for a couple of weeks after each rebuild so you can see whether the new seed actually improved anything.
Can I use website visitors as a seed instead of customers?
You can, and it is a reasonable fallback when the customer list is too small to match well. Expect a weaker audience, because visitors include everyone who arrived by accident. If you go this route, require depth of engagement rather than a mere visit: people who reached a late funnel step, spent real time, or completed a form. That approximates intent without needing a purchase.
What is a value-based lookalike?
It is a lookalike whose seed carries a number per record, usually revenue or a lead score, so the model weights the most valuable members more heavily. It works well when the spread of customer value is wide and the data is honest. With a narrow spread it adds nothing over a plain list, and with invented values it teaches the model to hunt for the wrong pattern.
Do lookalike audiences work across several countries?
Each lookalike is built inside one country, because the model compares a seed against that country's population. To expand, create a separate audience per market from the same seed, but expect quality to vary: a seed of customers from one market may map poorly onto another where the buying context differs. Where you already have local customers, seed each market from its own.
Are lookalikes still worth building if broad targeting works?
They are worth testing rather than assuming. Broad delivery has improved to the point where a campaign with plenty of clean conversion data often matches a lookalike, especially on large platforms. Lookalikes still earn their place when the account is new and short of conversion history, when the buyer is rare, or when you want a specific customer type rather than whoever is cheapest to convert.