Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

Meta Ads › Targeting & Audiences

Lookalike Audiences

How lookalike audiences are built from a seed, choosing seeds by value, the size trade-off, country and stacking rules, and lookalikes with Advantage+ audience.
Guideintermediate10 min readUpdated

How a lookalike is built

Meta's documentation describes the process in three parts. First, you choose a source, which can be a custom audience, the people who like or follow a page, or data from a pixel or app. Second, you choose the location: a country, a group of countries, or a region, which defines the population the model searches. Third, you choose a size, expressed as a share of that population, where the smallest setting contains the people most similar to the seed and larger settings reach progressively less similar people.

Meta then identifies the common qualities of the people in the seed (the documentation refers to demographics, interests and behaviours) and finds people in the chosen location who share them. The result is an audience that excludes the seed members themselves, so a lookalike of purchasers does not contain those purchasers. The audience is refreshed on a schedule as long as the seed is still updating, so a lookalike built from a live website audience evolves with the site's traffic, while one built from a static customer list stays fixed until the list is replaced.

Meta also documents a minimum number of people the seed must contain and recommends a seed in the low thousands rather than the hundreds; the documented figures are in the lookalike Help Centre page. Larger seeds are not automatically better. Past a few thousand people, additional seed members add little to the model and can dilute it if they are less representative. What the model cannot do is see your intent: it finds people like the seed, so a seed made of the wrong people produces a confident audience of the wrong people.

Choosing the seed: quality over quantity

Everything downstream depends on the seed. The table ranks the common choices by how well they tend to represent future customers.

SeedWhat it representsStrengthCaveat
Customer list with valuesKnown buyers, weighted by what they spentStrongest: real buyers, real valueNeeds a clean list and a value column; static until re-uploaded
Purchase event audienceBuyers tracked by the pixel or Conversions API within a windowStrong and self-refreshingDepends on event coverage; a leaky pixel makes a partial seed
Lead submitters or qualified leadsPeople who completed a form, ideally filtered to qualified onesStrong for lead generationUnqualified leads teach the model to find more unqualified leads
Add-to-basket or checkout startersHigh intent without purchaseUseful where purchases are scarceIncludes people who chose not to buy
Product viewers, long-window visitorsInterest without commitmentLarger seed, weaker signalModels browsers, not buyers
Video viewers at deep watch depthAttention to a messageFast to build, good for considerationWeak link to purchase
Page or profile engagers, followersAffinity with the brandEasy and largeOften the least like actual customers

Two practical rules follow. Choose the seed from as far down the funnel as volume allows: purchasers over basket adders, basket adders over viewers. And filter the seed to the customers you want more of, not every customer. A seed that includes one-off discount buyers alongside repeat full-price buyers teaches the model a blend; a seed of the repeat buyers teaches it the ones that matter. The custom audiences guide covers how to build each of these as an audience; this page is about which one to pick.

Size: the reach vs similarity trade-off

The size setting is a slider from the smallest share of the chosen population to a larger one. Meta documents that smaller sizes match the seed more closely and larger sizes increase reach at the cost of similarity. Because the size is a share of the location's population, the same setting produces an audience many times larger in a populous country than in a small one, which is why size should be chosen per market rather than copied between them.

The trade-off is not linear. The narrow end concentrates the people most like the seed, and performance there is typically the best the lookalike will give, but the audience exhausts quickly under meaningful spend and cost per acquisition (CPA) rises as frequency builds. Widening adds people who resemble the seed less, so the average quality falls, but it also gives the delivery system more room to find converters it would otherwise have missed. In practice the right size depends on the budget, the market and the strength of the seed: a strong seed tolerates a wider setting because even its looser matches are decent, while a weak seed is only worth using at the narrow end, if at all.

Meta lets you create several lookalikes from one seed in one step, each covering a distinct size tier that does not overlap with the others. That feature exists for exactly this question: run adjacent tiers as separate ad sets, compare them, and widen or narrow with evidence. The testing section below sets out how to do that fairly.

Value-based lookalikes

A value-based lookalike is built from a seed that carries a value for each person: a customer list with a value column, or a pixel or catalog audience where purchase events include a value parameter. Meta documents that the model uses those values to weight the seed so that people who resemble high-value customers are favoured over people who resemble low-value ones. The lookalike is still a single audience with a size setting; what changes is which seed members the model treats as most representative.

This matters because an unweighted purchaser seed is dominated by whoever buys most often, which is often the lowest-value, most promotion-driven segment. Weighting by value shifts the model towards the customers who make the economics work, and that shows up as better return on ad spend (ROAS) from the resulting audience rather than merely a lower cost per order. If you hold margin or lifetime value per customer, use it as the value rather than first-order revenue; the model will then favour people who look like your best long-term customers rather than your biggest single orders. Where value data does not exist, filtering the seed to repeat buyers is a rough substitute. Preparing that seed file, from the segment to the value column and the upload, is covered in the customer list upload tutorial.

Country, stacking and multiple lookalikes

Country. Every lookalike is built against a chosen location, and Meta documents that the seed does not need to contain people from that location: a seed of customers in one country can be used to build a lookalike in another. That is useful for market entry, with the caveat that the model is then finding people in the new market who resemble customers in the old one, which assumes the two markets buy for similar reasons. Selecting several countries in one lookalike produces one audience modelled across the combined population; separate lookalikes per country keep the sizes and the reporting separate.

Stacking. A single ad set can include several lookalikes, which widens the audience to their union. Stacking is a reasonable way to give the delivery system a larger pool when each individual lookalike is small, but it removes the ability to see which seed is working, and under the original audience options the stacked audiences may overlap heavily. The exclusions and overlap guide explains how to check that and when it matters.

Multiple lookalikes. The documented create-multiple feature produces non-overlapping tiers from one seed. Separate seeds (purchasers, leads, value-weighted customers) produce audiences that do overlap, and running them in separate ad sets invites them to compete. The sensible pattern is to test seeds against each other for a period, keep the winner, and either retire the others or stack them into one ad set.

Illustrative example — not a real result

A subscription meal kit tests three seeds and two size tiers

A meal-kit subscription business in the UK had been running one lookalike built from all website visitors in the last 90 days at the 3% setting. Cost per first-box order was about £48 against a target of £40. The test rebuilt the seeds: a customer list of roughly 4,200 subscribers who had stayed past their third box, uploaded with lifetime revenue as the value column; a pixel audience of about 1,900 purchasers in the last 60 days; and the original visitor audience as the control. Each seed produced two lookalikes, 1% and 2% to 4%, giving six ad sets with identical creative and daily budgets of £150 for three weeks. The value-based subscriber seed at 1% delivered first-box orders at about £31 and, more importantly, its subscribers were still active at a higher rate after two months. The purchaser-event seed came second at around £37. The visitor seed at either tier did no better than the account's broad ad set. The account kept the value-based seed at both tiers, retired the visitor lookalike, and refreshed the subscriber list monthly.

Lookalikes alongside Advantage+ audience

Under Meta's original audience options a lookalike is a boundary: the ad set reaches people in the lookalike and nobody else, subject to the other targeting. Meta also documents an Advantage lookalike option that, when enabled, allows delivery beyond the chosen lookalike size if the system predicts better results; for some objectives it is applied automatically. Under Advantage+ audience, which is the default on new ad sets for most objectives, a lookalike added to the ad set is an audience suggestion. Meta documents that suggestions inform where delivery starts and that the system may reach people outside them when it expects better performance. The lookalike is a hint, not a wall.

That changes what a lookalike is for. As a suggestion, it is a way to point a broad ad set at a good starting population, which is useful early in an account's life when the system has little conversion history of its own. As the account accumulates purchases, the suggestion matters less and the system's own learning takes over. If you need the lookalike to be a hard limit, for a controlled test or because a licence restricts who may be reached, switch the ad set to the original audience options. The Advantage+ audience guide covers the documented controls and the decision between automated and defined audiences in full.

Testing lookalikes fairly

Lookalike comparisons go wrong in predictable ways: tiers that overlap, ad sets that compete in the auction, creative that differs between arms, and conclusions drawn before the learning phase ends.

A fair test holds the creative, the optimisation event, the placements, the budget and the schedule constant across arms and varies only the audience. Use the create-multiple feature or explicit tier ranges so size tiers do not overlap. Where seeds are being compared rather than sizes, accept that the audiences overlap and use Meta's A/B test tool, which splits people between arms so the same person does not see both. Run each arm long enough to leave the learning phase and to collect enough conversions that the difference is not noise; for low-volume accounts that may mean weeks, and it may mean testing on a higher-funnel event. Judge on the business metric, not the proxy: cost per qualified lead or return on spend, not clicks. The CPA calculator is a quick way to turn spend and conversions per arm into a comparable figure.

Finally, be clear what the lookalike is being tested against. The relevant comparison today is usually the lookalike versus a broad or Advantage+ audience ad set with the same creative. If the lookalike wins, keep it as a boundary; if it merely matches, it is probably better used as a suggestion inside the broad ad set. The author's Meta Ads expertise page describes how that decision tends to go on accounts of different sizes, and the targeting hub places lookalikes in the wider audience map. For a prospecting decision read on purchases and new-customer share, the audience targeting breadth test lays out the full design, with a lookalike as a possible control arm.

Whether a lookalike should stay a boundary or become a suggestion inside a broad ad set is weighed in the comparison of lookalike audiences and Advantage+ audience, including how to test one against the other fairly.

Common mistakes

  • Seeding from the top of the funnel. Visitors and engagers are easy seeds and poor ones; model buyers.
  • Chasing seed size. A few thousand good customers beat tens of thousands of mixed ones; filter rather than inflate.
  • Ignoring the value column. Without it the model weights by frequency of purchase, which favours the wrong customers.
  • Copying a size setting between countries. The setting is a share of population, so the same number means very different audiences in different markets.
  • Overlapping tiers in separate ad sets. Two ad sets reaching the same people compete; use non-overlapping tiers.
  • Testing a lookalike against nothing. The question is whether it beats broad or Advantage+ delivery with the same creative, not whether it converts at all.
  • Letting a static seed age. A customer list from a year ago models last year's customers; schedule the refresh.

Frequently asked questions

What is the best source audience for a lookalike?

The seed that best represents the customers you want more of, which is usually purchasers or high-value customers rather than visitors or engagers. Meta documents that a seed of a few thousand people works well and that quality matters more than volume, so a smaller list of real buyers beats a larger list of page fans. Where you hold value data, a value-based seed lets the model weight the best customers more heavily.

Should I use the smallest lookalike size or a larger one?

The smallest setting is the most similar to the seed and the least reach; larger settings add people who resemble the seed less closely. Start narrow when the seed is strong and the budget is modest, and widen when the narrow tier is exhausted or when the account has enough conversion data for broader delivery to work. Testing adjacent, non-overlapping tiers in separate ad sets is the fair way to find the point where similarity stops paying.

Do lookalikes still matter now Advantage+ audience exists?

Yes, but their role has changed. Under Advantage+ audience a lookalike is a suggestion that steers initial delivery rather than a wall around it, so the system can and will reach beyond it. Lookalikes remain a hard boundary under the original audience options, which is where they are still useful for controlled prospecting, licence-restricted markets and fair tests.

Terms used on this page

Tools to try

Where this page is used

Part of these learning paths

Also related

Sources

  1. About Lookalike Audiences — Meta Business Help Centre
  2. Meta Business Help Centre — Meta Business Help Centre

Related