Meta Advantage+ shopping: exclude past buyers or let them into prospecting?

A lot of stores running Advantage+ shopping hit this exact question. Excluding past purchasers makes new-customer CPA look better, and that's the number most people report. But it can also starve the algorithm of the purchase signal it uses to find lookalikes.

Take a supplement brand where customers reorder every 30 days. If you exclude anyone who bought in the last 180 days, you either pay to re-find people who would have bought anyway, or you push them toward a competitor's ad.

My view is to leave them in prospecting and split new versus returning revenue in reporting, so the efficiency numbers don't hide what's actually happening. Has anyone found a split that holds up when the repeat rate is high?

The new-versus-returning split is only as good as the match behind it. Meta can only flag someone as a past buyer if your purchase events arrive with usable identifiers, so if CAPI matching is patchy, the returning bucket gets undercounted and new-customer CPA looks better than it really is. And if you want to know whether repeat buyers would have purchased anyway, run a lift test or hold out a customer list. Attribution tells you where a sale got credited, not whether the ad caused it.

Jun's right about matching, but I'm not sure a lift test is the fix for a supplement brand. With a 30-day reorder cycle, a holdout big enough to read costs real money, and the people you hold out are exactly the ones who'd reorder anyway. You'd be paying to learn something the reorder curve already tells you.

Cohort repeat rates by first-purchase month, plus a plain check of how many new buyers came in during a stretch with no spend, is usually cheaper and gets you most of the way there.

The cohort curve is the cheaper place to start, but a no-spend check only works if Meta is your main source of new buyers. For a supplement brand that's often not the case. Brand search, email flows and organic listings keep sending first orders during a pause, so a drop in new buyers mostly tells you how much paid you lost, and the rest is noise.

So it comes down to what share of new customers comes from non-paid channels. If that share is small, Iris's shortcut is probably enough. If it's large, filter the pause data to first orders with paid UTMs or Meta-specific landing pages instead of looking at total new orders. It's messier, but it's the only way the pause tells you anything about the ads.

The pause data is noisy, but you can sidestep most of that by not using one exclusion window for everyone. Tie it to the reorder cycle instead. Exclude past buyers while they're still inside the time they'd normally reorder, roughly 1 to 1.5 times the median gap between orders, then let them back into prospecting. For a 30-day supplement that works out to about 30 to 45 days, not 180.

Inside that window, a Meta impression mostly takes credit for a sale that was going to happen anyway, so you're paying for nothing. Past it, they've lapsed, and a lapsed customer is really a prospect with a purchase history, which gives the algorithm something useful to work with. Your cohort curve already shows the median gap, so you don't need a holdout to set the window.