What Are Lookalike Audiences?

What a lookalike audience actually is

A lookalike audience is not a list of people who look like your customers on a spreadsheet. It's a statistical similarity score. You give the platform a seed — a group of real people, usually your customers, email list, or people who converted on your site — and the platform's model scores everyone else in its user base on how closely they match that seed's behavioral and demographic pattern. The people who score highest become the new audience.

This is different from two things it often gets confused with. A custom audience is people who already know you — past visitors, past buyers, your email list, retargeted directly. An interest audience is people the platform guesses share a trait, based on declared interests or inferred category affinity. A lookalike sits between them: it's new people, but the match is derived from real behavior of your real customers rather than a guessed category.

Facebook built the first version in 2013. Google Ads, Outbrain, Taboola and LinkedIn Ads have since shipped comparable tools under different names — Google calls it 'similar audiences,' LinkedIn 'lookalike audiences.' Amazon's DSP has its own version of the same idea inside the audience builder, matching shoppers to a seed drawn from purchase or browse behavior on Amazon's own retail signal, which is a different, arguably richer, data set than a pixel or an email list.

How the matching actually works, with real numbers

Facebook's own methodology, and most platforms that copied it, run three steps. First, pick the seed — page fans, site visitors, or a customer list, with a recommended minimum of 500 to 1,000 people; smaller seeds work but get noisier. Second, pick the location — the pool of people the model is allowed to search within. Third, pick the size, expressed as a percentile from 1% to 10% of that location's population. A 1% lookalike is the tightest match and the smallest reach. A 10% lookalike reaches far more people but the resemblance to your seed gets looser with every point you add.

Here's what that looks like on a real book of business. Across 30 advertisers running Amazon DSP in July 2026, one group's campaigns produced 57,137 attributed purchases, and 20.1% of those buyers were new to the brand. That full purchase list — not just the new-to-brand slice — is exactly the kind of seed a lookalike model wants: it's homogeneous (everyone in it did the same thing, bought), it's recent, and it's large enough to be statistically stable. Feed that list in, and the model isn't guessing who might like the brand — it's finding people who behave like 57,137 people who already bought.

The underlying technology is standard machine learning: distance-based clustering, keyword-based models, and classification algorithms scoring individuals against the seed's behavioral and demographic signal. The size and quality of the seed matters more than any tuning the marketer does afterward — a small, noisy seed produces a small, noisy lookalike no matter what percentile you pick.

How to create a lookalike audience, step by step

Every platform's process boils down to the same three inputs. The table below covers the three common seed types used in ecommerce, which one to reach for, and what it's actually good for.

• Homogeneity beats size. A tight seed of 1,000 people who all bought the same product line will outperform a messy seed of 50,000 people who did unrelated things.
• Minimum size matters less than people think. Facebook's technical floor is 100 users from one country; its own recommendation is 1,000 to 5,000. Below that, the model has too little signal to find a real pattern and starts matching on noise.
• Percentile is a trade-off, not a setting to maximize. Smaller percentages (1-3%) give tighter resemblance and lower reach. Larger percentages (7-10%) give more scale but the new users drift further from the seed's actual behavior.

Do lookalike audiences actually work — and where they don't

Academic work on the technique generally shows a real lift over blind interest targeting, and it's been listed as a standard PPC tactic for years. That's not the same as saying it always works. Two things break it reliably: a seed that's too small (startups and small brands often don't have enough converters yet to give the model anything to learn from), and a seed that's mixed — bundling browsers, cart-abandoners, and buyers into one list dilutes the pattern the model is trying to find.

There's also a compliance history worth knowing. In 2019, Facebook removed age, gender, and ZIP code as available inputs for housing, employment, and credit ad categories after concerns the tool could be used to exclude protected groups. In 2022, the DOJ's Civil Rights Division sued Meta over exactly this — alleging the lookalike tool discriminated in housing ad delivery — and Meta settled the same day. If you're advertising in those regulated categories, the targeting options are narrower than they used to be, on purpose.

The other limit is one platforms won't tell you: the reported ROAS on a lookalike campaign is almost always last-click. Last-click can't tell you whether the person matched by the model would have bought anyway. That's not a lookalike-specific problem — it's an attribution problem that sits underneath every audience tactic on every platform. The only way to know if a lookalike audience is adding incremental sales is a holdout: hold a matched group of otherwise-eligible users out of the audience, run the campaign against the rest, and compare. Reconciling DSP delivery against organic and sponsored-ad activity in a clean room like Amazon Marketing Cloud is the only honest way to see whether display added anything the other channels weren't already going to get credit for. Dr. DSP, the managed Amazon DSP product from Full Circle, builds every lookalike-style test this way: evidence behind the change, a measurement plan, and a rollback trigger, before it runs.

The common mistake, and what to do when it isn't working

The single most common mistake is treating the reported campaign number as proof the audience worked. It isn't proof. It's a last-click number sitting on top of an audience that was, by definition, already similar to people who convert — some of that number would have shown up anyway. We've made this same read ourselves on early campaigns before insisting on a holdout: a lookalike segment looked like the best-performing audience in the account, and a matched-control test later showed a meaningful chunk of it wasn't incremental at all.

If a lookalike audience isn't performing, work through this order before killing it: check the seed size first — anything under a few hundred people is too thin to trust. Check seed homogeneity second — mixed intent signals (browsers and buyers together) produce a mixed, unfocused match. Check seed freshness third — a seed built from data six months old is matching against who your customer was, not who they are now. Only after those three checks should you touch the percentile size, and even then, move it one point at a time and measure against a holdout, not against the platform's own reported number.