Meta lookalike seed: all purchasers or only your high-value buyers?

Seeding a lookalike with every purchaser versus a smaller group of high-value buyers keeps coming up. The usual advice is to use the biggest pool you can get, since Meta needs volume to find patterns. But if your price range is wide, a seed that includes everyone who bought a $12 keychain can pull in bargain hunters who never come back.

Take a skincare brand with 3,000 buyers a year, 400 of whom spent over $150. Seeding on those 400 gives you a narrower audience, and the first-order cost per purchase often looks worse because the lookalike is going after pricier buyers. So I'd run both for a few weeks and compare second-purchase rate and margin instead of the cost per purchase in Ads Manager. Just keep in mind that a seed under a few hundred people is usually too thin to build a useful lookalike anyway.

Agreed on seed size, but the comparison is the harder part. Meta won't report second purchases as a result of the lookalike, so the only clean version of your test is cohort-based. Tag each new customer by the ad set that acquired them, using separate UTMs or landing pages, then pull repeat rate from your order data over the same window. Otherwise a second order gets credited to whichever retargeting campaign happened to catch it, and both seeds end up looking better or worse than they really are. Are you pulling those repeat numbers from Shopify or a CRM, or from the ad platform?

Cohort tagging is the right call, but when you read it depends on how often people rebuy the product. A serum on a 45-day cycle will show a usable repeat rate within a few weeks, while a gift set or a bath product might need six months before the two seeds even separate. Read at three weeks on a low-frequency product and you're mostly measuring who reordered fast, which isn't the same as who came back. I'd pull your existing reorder curve from order history first and set the window from that, not from the test calendar.

Agreed on the reorder curve. The seed list itself can quietly skew this test, though. If existing purchasers aren't excluded from both prospecting ad sets, the lookalike keeps serving people who already bought, and they land in your "new customer" cohort as fast reorders. A high-value seed could then look like the winner just because its audience keeps catching past buyers. So add an exclusion for all customers to both ad sets before launch, then check that each cohort's count grows with new names.

The exclusion is the right move, but the age of that customer file matters. Upload it in January and leave it alone, and every December buyer stays eligible for both prospecting sets, which is exactly the reorder contamination you're trying to avoid. With a 45-day cycle, a list can go stale within a few weeks. Refresh it weekly or sync it straight from Shopify, and make sure both ad sets get the same version on the same day. If one list gets rebuilt more often than the other, the gap can look like a performance difference between the seeds.

Depends on whether your customer file carries order value. If it's just emails, getting a high-value seed means filtering on order totals first, and that's one more step where the two lists can drift apart.

Freeze the filter in one export. Pull both seed files from the same query at the same timestamp and save that CSV, so neither arm gets re-pulled later with different orders. Keep in mind a plain email upload doesn't tell Meta who spent more, so the value split only exists in the file you built. If you'd rather let Meta weight by value, the value-based lookalike option uses purchase value from your pixel or conversion events. That's a different seed with its own sample size, so don't mix the two methods across your test arms.

Freezing both files is where I'd disagree. Freezing the seed makes sense, since a customer-list lookalike is built from whatever you uploaded and won't quietly change. The exclusion list is a different story. If your December buyers are still on it in week six, they're still eligible for prospecting, and a frozen export just makes that contamination identical in both arms. The test looks fair but isn't measuring new customers. I'd freeze the seed, sync the exclusion from Shopify on a schedule, and log the sync date next to each cohort's count. That way you can see whether the "new" cohort is actually growing with new names or quietly refilling with past buyers.