New Shopping campaign: how long before you judge CPA when purchases lag?

Picture a Shopping campaign for a store where people tend to compare for a few days before buying. On day 3 you've got two sales and a CPA that looks awful, and the urge is to cut budget or pause. I'd judge the campaign against its conversion lag instead of the calendar. Check the conversion lag report first. If most purchases land 4 to 6 days after the click, a 3-day read is mostly noise. Set a rule before you touch anything: no structural changes until the campaign has enough conversions inside the window that matches its lag. At low volume that can mean waiting weeks, though, and waiting costs money too. Where do you draw the line?

Agree with the rule, and I'd add something about how to read the lag. The conversion lag report is bucketed by conversion date, so on day 3 the newest days look worse just because the later purchases haven't come in yet. A cleaner way is to take the clicks from a fixed past window, say the cohort from 10 days ago, and see how much of its eventual conversion volume had shown up by day 3, day 5 and day 7. Once that share stops climbing, you have the real CPA for that click period. So wait for that point, not for a calendar date.

If the campaign is bleeding while you wait, a tighter daily budget limits the damage without touching structure, which is usually what you want to avoid changing mid-read.

Agreed on the cohort view. One thing that trips people up: Shopping conversions often get reported in two layers. The first purchase can come from a click days earlier, and the conversion action's attribution window plus the "conversions" vs "all conv." column can shift the numbers depending on which you read. So lock the attribution setting before you compare cohorts, or day 3 and day 7 won't be measuring the same thing.

Locking the attribution setting makes sense, but how much it matters depends on volume. At a few sales a week, one purchase landing on a different click day can swing your CPA more than any lag curve will, so check the raw conversion counts before you trust a cohort chart.

I don't buy that one sale matters more than the lag curve. At a few sales a week, a single purchase does swing CPA hard, but that's a reason to pool clicks over a longer window, not a reason to look at the chart less. One weekly cohort is mostly random purchase timing. If you build the cohort from four or five weeks of clicks, the single-sale swings start to average out, and the curve becomes usable even at low volume. The raw counts should tell you how much to trust the curve, not whether it's worth building.

Pooling works, but the thing to watch is cohort age. If you pool five weeks of clicks and the newest week is only four days old, that week hasn't matured yet. It drags the curve down and makes late conversions look rarer than they are. Only include click weeks that are at least as old as your longest lag, so 14 days if purchases land 4 to 6 days after the click. Then check that the share still rises on the oldest weeks before you trust the plateau.

Agreed on the maturity check, and it's the step people skip. A rule of thumb that works: don't trust a cohort until it's at least twice your 90th percentile lag old and has 20 to 30 conversions in it. The age rule keeps the curve from reading low, and the volume rule keeps you from calling a plateau when it's really just two lucky or unlucky sales. Until both are met, only tighten the daily cap. That limits losses without changing what you're measuring.

I don't agree with the cap, though. Tightening the daily budget isn't neutral while you're reading a cohort. It changes how many clicks you get, which changes the cohort you're trying to measure, and at low volume it pushes the 20 to 30 conversion threshold even further out. If you're only selling a few units a week, a capped campaign could sit there for months and never get close to it. I'd keep spend steady during the read and set a hard loss ceiling up front, something like a maximum weekly cost you're willing to absorb. You only cut if you hit that ceiling.