What Is a Lookalike Audience, and How Does It Actually Work?

What a lookalike audience actually is

You don't build a lookalike audience by hand. You give the platform a seed audience — an existing customer list, people who bought after clicking a pixel event, people who engaged with a page — and the platform's model looks at everything it knows about those people that you can't see directly: browsing patterns, purchase timing, device and interest signals. It then finds new people who match that pattern closely enough to be worth showing an ad to.

On Facebook, this sits under Audiences: you pick a source (customer list upload, Pixel purchasers, page engagers, app installs), pick a country, and pick a size. The size is expressed as a percentage — roughly, how large a slice of that country's population the platform is willing to match against your seed. Smaller percentage, tighter match. Larger percentage, broader reach, looser match.

How it's built, with a worked example

Say a brand uploads its 3,000 highest-lifetime-value repeat purchasers as the seed. A 1% lookalike is the tightest match the platform can build from that list — it returns the smallest audience, but every person in it scored closest to the seed's pattern. Move up to a 5% or 10% lookalike and the audience gets several times larger, but the match loosens: the platform is now including people who share fewer of the traits that made the seed audience valuable in the first place.

That tradeoff — precision versus reach — is the entire mechanic. There's no setting that gives you both a huge audience and a tight match. Most advertisers start at 1%, prove the seed works, then widen once performance justifies more spend.

The common mistakes — including ones we've made

The seed list is where almost every bad lookalike goes wrong, not the algorithm:

• Mixed-quality seeds. Lumping one-time discount shoppers in with repeat, full-price buyers teaches the model the wrong pattern. It'll go find you more coupon hunters.
• Forgetting to exclude existing customers. A lookalike built without an exclusion list will happily spend budget re-showing ads to people who already bought.
• Letting the seed go stale. A customer list from eighteen months ago describes who your customers were, not who they are now.
• Treating a conversion as proof it worked. A sale attributed to the lookalike campaign doesn't mean that person wouldn't have bought anyway through search or direct.

One of ours, on the Amazon DSP side rather than Facebook: we once built an audience straight off all-time purchasers, including deal-only buyers, for a household brand. It matched more people who'd only ever transact on a discount. Rebuilding it off repeat, full-price purchasers held CPA in a much narrower band. Same mechanic Facebook uses, same failure mode, same fix — check the seed before you blame the algorithm. Brands as different as HexClad and Ridge have hit this exact issue with badly-mixed seed lists.

When the answer is bad news: it stopped working, or you can't tell if it ever did

If a lookalike audience's numbers went south, check in this order: seed size and quality first, then whether you're excluding current customers, then whether the campaign is still in its learning phase, then attribution. That last one is the part nobody checks and it's usually the real problem. Last-click reporting will tell you a lookalike audience drove a conversion. It cannot tell you whether that person would have converted anyway through another channel. Only a holdout — a matched group that didn't see the ad — or a clean reconciliation against your other channels answers that.

This is the exact question we build our own reporting around. Across 30 of our advertisers in July 2026, the book delivered 6.04x return on ad spend, 78.4 million impressions at a $4.00 CPM and a blended $1.42 cost-per-click. Blended cost per acquisition was $5.49 across 57,137 attributed purchases, and 20.1% of them were shoppers new to the brand. That new-to-brand share is the number that actually tells you whether an audience — lookalike or otherwise — expanded who's buying, versus just claiming credit for people who were going to buy regardless.

Lookalike audiences on Amazon: same idea, different substrate

Amazon DSP doesn't use Facebook's exact 1%-to-10% slider, but it offers a comparable idea: audiences modeled on shopper similarity, built from Amazon's own retail signals — browse behavior, search terms, purchase history — rather than social engagement. The honest tradeoff is different too. Facebook's version is fast and self-serve; you can spin one up in minutes from a pixel event. Amazon's version is built and managed through the DSP console or a managed service, and it tends to reward brands with real purchase history over brands with only page views, because purchase signal is a stronger predictor on a retail platform than it is on a social one.

Neither version is better in the abstract. Facebook's lookalikes suit anyone who's already collecting pixel data and wants to scale prospecting fast. Amazon's version suits brands whose strongest signal is what people actually bought, not what they clicked.

Where Dr. DSP fits

Dr. DSP is Full Circle's managed version of Amazon DSP — not the Delivery Service Partner courier franchise, the ad platform — run by a team that has managed more than $500M in revenue across 100+ brands. Every audience change, lookalike-style or otherwise, ships with the evidence behind it, a measurement plan, and a rollback trigger, and results get reconciled in Amazon Marketing Cloud so DSP and sponsored ads stop double-counting each other's conversions. There's no published price — a demo, the first 30 days free, and terms set on a call against real budget and scope, with the full Orbit software suite included either way. If you're only running Facebook lookalikes, none of that applies to you yet — but the moment you start asking whether an audience actually added sales rather than just claimed them, this is the question to bring.