Facebook Lookalike Audiences: Source Quality, Size Thresholds, and Why Creation Fails

A Facebook lookalike audience takes a list of people you already value — buyers, subscribers, high-LTV users — and asks Meta to find more accounts that resemble them. The mechanic is simple to describe and easy to misuse. Most of the outcomes people blame on "the algorithm" trace back to three earlier decisions: what went into the source list, how big that list was, and how wide they set the percentage.

This guide covers what the model is actually matching on, the size thresholds that decide whether the audience builds at all, why creation fails, and how to read a lookalike's performance without fooling yourself.

What a lookalike is actually built from

The input is a source audience — a custom audience built from a customer list, pixel or Conversions API events, app activity, or engagement with your Page and videos. Meta profiles the people in that source, then scores everyone else in the chosen country for similarity to that profile.

Two things follow from this that decide most of your result:

The model inherits your source's biases, including the accidental ones. If your source list is "everyone who purchased in the last 180 days" and 70% of those purchases came from one discount campaign that attracted bargain hunters, the lookalike will faithfully find you more bargain hunters. The model has no concept of which customers you wanted; it only sees the list you handed it.

Similarity is measured against the source, not against your business. A lookalike of your newsletter subscribers finds people who resemble newsletter subscribers — people likely to sign up for things. That is a different population from people likely to pay. Choosing the source is choosing the outcome, and it is a decision made before any budget is spent.

The practical consequence: build the source from the behavior closest to the money. A lookalike from a purchase-event audience and a lookalike from an all-site-visitors audience are not two versions of the same thing.

The size thresholds, and what happens near them

Meta requires at least 100 people from a single country in the source audience before it will build a lookalike. That is the floor for the system to accept the request — it is not the point at which results become reliable.

The useful working range is considerably higher. A source in the low hundreds gives the model very little signal to profile, and the resulting audience tends to be unstable: rebuild it a month later with slightly different members and you get a materially different population. Sources in the thousands produce lookalikes that are more stable across rebuilds, which matters because you will rebuild them as your customer base changes.

Bigger is not monotonically better, though, and this is where the "just use your biggest list" instinct goes wrong. A source of 50,000 mixed-quality contacts profiles as a blurry average. A source of 3,000 genuinely high-value customers profiles as something specific. When you have to choose between size and coherence, coherence usually wins — as long as you clear the floor with room to spare.

The percentage is a separate decision from the source. The 1%–10% slider sets how much of the country's population to include, ranked by similarity. A 1% lookalike in a large country is already millions of people. Going wider does not "improve" the audience; it appends progressively less similar people to the bottom of the ranking. Start at 1%, and widen only when delivery is genuinely constrained — not as a default.

Why lookalike creation fails

Creation errors are common and usually have mundane causes:

The source hasn't finished populating. Custom audiences from a customer list need to finish matching, and event-based audiences need to accumulate qualifying events. A source that shows as "populating" or displays a size below the threshold cannot be used yet. Waiting is the fix.

The source is below 100 in the selected country. A list with 5,000 people spread across forty countries may have fewer than 100 in the one country you selected. The threshold is per-country, and the audience size shown is the total — those two numbers disagree far more often than people expect.

The source is itself a lookalike. You cannot build a lookalike of a lookalike. Go back to the original custom audience.

The source has been made ineligible. Audiences built from certain event types or from data sources that later became unavailable can stop qualifying. If a source that used to work stops working, check whether the underlying pixel, Conversions API integration, or app event stream is still delivering — the failure is often upstream of the audience.

The account has a restriction in place. If the ad account is limited or under review, audience creation is one of the things that can be blocked. That is a different problem wearing an audience-shaped costume; if your account is in that state, resolve the account issue first.

Automation has changed where lookalikes sit, not what they do

Meta's automated audience products increasingly treat your targeting inputs as suggestions rather than boundaries — the system explores beyond the audience you specified when it predicts better results elsewhere. This has led to a recurring question about whether lookalikes are going away.

The useful way to read the shift: lookalikes are becoming an input signal rather than a fence. In older buying flows, selecting a 1% lookalike meant your ads were shown to that audience and no one else. In newer automated flows, that same audience is a strong hint about where to start, with the system free to move outward.

What this changes in practice:

• Testing lookalike variants against each other gets less clean. If the system is exploring beyond both audiences, differences in delivered population are smaller than differences in what you selected.
• Source quality matters more, not less. A hint is still a hint. A source built from real purchasers still points the system somewhere better than a source built from all traffic.
• Broad-versus-lookalike is no longer a clean either/or. The honest test is whether adding the lookalike signal changes your cost per result at comparable spend — run it as a test, don't assume the answer in either direction.

Reading performance without fooling yourself

Three habits prevent most of the wrong conclusions:

Judge on cost per result, not on audience-level rates. A lookalike will usually show a better click-through rate than broad targeting, because it is selecting for people who resemble people who already engaged with you. That is close to circular. What matters is whether it produced results at a lower cost.

Give the source time to refresh. Lookalikes built from event audiences update as the source updates. A lookalike built from "purchasers in the last 30 days" is a moving target — comparing its performance in week one against week six is comparing two different audiences.

Check overlap before concluding one audience beats another. Two lookalikes from related sources can overlap heavily. When they do, running both is not a test; it is one audience with two names and a budget split between them.

Attribution deserves the same skepticism: if your lookalike results look dramatically better or worse than expected, verify that the conversion data feeding the comparison is complete before you redesign the audience. Deduplicated server-side events and consistent link tracking decide whether the numbers you are comparing mean anything.

A build order that avoids the common mistakes

1. Pick the behavior closest to revenue and build a custom audience from it. Purchases beat add-to-carts; add-to-carts beat page views.
2. Check the per-country count, not the total. Confirm you clear 100 in the target country with meaningful headroom.
3. Build at 1% in a single country to start. One country per lookalike gives you cleaner reads than a multi-country audience.
4. Wait for it to finish building before attaching budget. An audience still populating will underdeliver and you will misread that as a targeting failure.
5. Change one thing per test. New source or new percentage — not both.
6. Rebuild on a schedule as your customer base shifts. A lookalike built from last year's buyers profiles last year's business.

Where this fits

Lookalikes are a distribution decision that depends on a measurement decision. The source list is only as good as the events that populate it, which means pixel and server-side event quality sits upstream of every audience you build. If your events are incomplete or duplicated, your source is wrong, and a well-configured lookalike of a wrong source is just a faster way to reach the wrong people.

DeepClick works on the layer underneath this — link and campaign infrastructure for Meta and TikTok advertisers, where tracking consistency and landing page delivery determine whether the events your audiences are built from reflect what actually happened.