Lookalike Segments Live or Die on Seed Quality
A lookalike segment finds new people similar to a seed audience, but it can only copy what the seed represents. A seed built on buyers becomes an audience of likely buyers. A seed built on contest entrants or accidental clicks becomes a larger group of people who will also never buy.
The seed decides the outcome
Every lookalike is a portrait of its seed. Include your best customers and the model hunts for people with similar value. Include every visitor and the model learns that your target customer is anyone with a browser. The most common failure is not the algorithm. It is a seed whose definition was never written down.
Be precise about the source of truth. Purchasers with a minimum order count and a recent date are a strong seed. Subscribers who never bought are a different seed with a different use. Mixing them into one bucket averages away the very signal you wanted to amplify.
Clean the seed before you expand it
Remove the noise first: refunds and chargebacks, fraud and abuse flagged accounts, gift buyers in unusual quantities, and one time anomalies like a staff testing spree. Those records do not represent customers you want more of, and leaving them in teaches the model the wrong lesson.
Size matters but quality matters more. A few hundred genuine best buyers usually outperforms several thousand loosely defined engagers. If the seed is too small for the platform minimum, narrow the question instead of diluting the seed: look alike the repeat buyers, not the whole customer base.
Test the lookalike, not the vanity metric
Run a holdout: send the same creative to a random audience and to the lookalike, and compare conversion per dollar, not reach or click rate. Lookalikes almost always reach more cheaply. The question is whether they convert better than the baseline, which is the only claim worth paying for.
Watch the first purchase quality, not just the first purchase count. New customers from a strong seed should repeat at a healthy rate over the following months. If they convert once at a discount and never return, the seed was bargain hunters and the lookalike faithfully multiplied them.
Refresh seeds on a schedule
Seeds decay as your business changes. A seed from last year customer mix represents last year customers, and after a pricing, product, or positioning change it may no longer describe who you want. Review the seed definition each quarter and rebuild from recent buyers.
Document what worked: seed definition, size, exclusions, and the resulting repeat rate. That record turns lookalikes from a hopeful experiment into a repeatable acquisition channel, and it stops teams from quietly reusing a seed that stopped working months ago.
Frequently asked questions
How many people does a lookalike seed need?
Platform minimums vary, but a few hundred genuine high value customers usually produce a better audience than thousands of loosely defined engagers. When the seed is too small, narrow the seed definition instead of padding it with low quality records.
Should we exclude one time buyers from the seed?
It depends on what you want more of. For long term value, seeds built on repeat buyers and high lifetime value customers usually outperform seeds of all purchasers. One time discount buyers, in particular, tend to produce more one time discount buyers.
How do you test a lookalike audience properly?
Send identical creative to the lookalike and to a comparable baseline audience, then compare conversion per dollar and downstream repeat rates. Clicks and cheap reach do not prove the audience works. Profitable customers do.
How often should the seed be refreshed?
Quarterly is a good default, and sooner after major changes to pricing, product mix, or positioning. A stale seed keeps optimizing for customers you no longer prioritize, so the refresh date should be a standing calendar item.