TL;DR
Lookalike audiences are not dead; bad seeds are dead. A lookalike is only as good as the source audience, and most advertisers build lookalikes from bloated, low-intent seeds (all website visitors, every email subscriber since 2019) and then blame the algorithm. The 2026 rules: seed from your highest-value converters (ideally 1,000 to 5,000 of your best customers), refresh the seed as your customer base evolves, and test 1% against broader percentages instead of worshipping 1%. And know when to skip lookalikes entirely: with Advantage+ audience expansion doing the finding for you, a strong broad setup often beats a mediocre lookalike.
Why the obituaries keep coming
The "lookalikes are dead" take resurfaces every year, and it always has the same parents: iOS 14 shrank the signal pool, Advantage+ targeting finds people without lookalikes, and broad campaigns keep winning head-to-head tests. All true. None of it means lookalikes stopped working. It means the bar for a lookalike worth running got higher, and most advertisers' seeds never cleared it.
Think about what a lookalike actually is. Meta takes your seed audience, finds the patterns in it, and finds more people with similar patterns. If your seed is "1,000 people who bought twice and never refunded," the patterns are buyer patterns. If your seed is "40,000 people who visited the site, including the ones who bounced in four seconds," the patterns are noise with a purchase event somewhere in the distance. The algorithm is a mirror. Complain about the reflection all you want; it's showing you your seed.
The seed rules
Quality over size, but size has a floor. The sweet spot most practitioners land on is 1,000 to 5,000 source users: enough for the system to find real patterns, small enough to stay selective. Below a few hundred, the patterns are guesswork. Above tens of thousands, you've usually diluted the seed with mediocrity to get there.
Seed your best, not your most. Rank your customers and seed from the top: highest lifetime value, repeat purchasers, the segment you'd clone if you could. A lookalike of your best 2,000 customers will outperform a lookalike of all 50,000 customers nearly every time, because "people like your average customer" is a weaker brief than "people like your best customers."
Value-based seeds when you can. If you have the data plumbing, value-based lookalikes (seeding with purchase values attached) tell the system not just who buys but who buys big. It's the difference between cloning your customer base and cloning your profit.
Refresh the seed. Your customer base drifts: new products, new markets, new seasons. A seed built in 2023 is modeling 2023's buyer. Rebuild or refresh source audiences periodically, especially after meaningful business changes.
The percentage question
The old religion said 1% lookalikes or nothing: the tightest similarity, the most qualified audience. In 2026, test it instead of assuming it. The trade-off is real: 1% gives you the closest matches but a small audience that can saturate fast; 3%, 5%, or 10% gives you scale at the cost of similarity. For big budgets, the broader percentages are often where the volume lives. For niche offers, 1% might still be the whole game.
The honest test: run 1% against 3-5% with identical creative and budget, and let the CPA decide. Percentages are a dial, not a doctrine.
When to skip lookalikes entirely
Here's the part lookalike loyalists don't like. Advantage+ audience and broad targeting have gotten good enough that a well-fed broad campaign, strong creative, solid conversion signal, often matches or beats a lookalike built from a mediocre seed. The algorithm is doing similarity-matching under the hood anyway; your lookalike is just a suggestion it may or may not need.
Our rule of thumb: if you have a genuinely excellent seed (large, high-value, fresh), test the lookalike. It can still win, especially for prospecting with tight CPA targets. If your seed is "everyone who's ever touched the brand," skip the lookalike theater and run broad with great creative. You'll get the same discovery with less maintenance and fewer tiny audiences starving in learning limited.
FAQ
Are lookalike audiences dead in 2026? No. They work when built from high-quality seeds. What's declined is the effectiveness of lazy lookalikes built from bloated, low-intent source audiences.
How big should a lookalike seed audience be? Aim for 1,000 to 5,000 of your best customers. Enough for pattern-finding, selective enough to stay meaningful.
What makes a good lookalike seed? Your highest-value converters: repeat purchasers, high-LTV customers, ideally with value data attached. Not all website visitors or your entire email list.
Should I use 1% or broader lookalikes? Test both. 1% gives the closest matches but limited scale; broader percentages (3-10%) give volume with looser similarity. Let CPA decide per account.
Do lookalikes beat broad targeting? Sometimes, with excellent seeds. Often, a well-fed broad campaign matches a mediocre lookalike, since Advantage+ does similarity matching under the hood anyway.
How often should I refresh my seed audience? Periodically, and always after meaningful business changes (new products, new markets, new seasons). A stale seed models a stale buyer.
MetaMaxd is a Meta ads agency. We run the accounts; you run the business. Want a second pair of eyes before Q4? That's what we're for.
