Performance Max with margin data: one blended tROAS or split products by margin tier?

Feed-based shopping has a problem most of us paper over. A 60 percent margin hoodie and a 12 percent margin phone case get the same target ROAS inside one Performance Max campaign, so the campaign quietly optimizes for whatever converts cheaply, which is usually the low-margin stuff.

There are two common fixes. One is custom labels in the Merchant Center feed: tag products as high, mid or low margin, then split asset groups or run separate campaigns with different tROAS targets. The other is keeping one campaign at a blended break-even target and handling the thin-margin SKUs separately, with a lower priority or a cap.

Labels give you control, but they fragment the learning signal, and small tiers can struggle to exit learning. The blended setup is simpler, but it tends to overspend on whatever is cheap to win.

For those running margin-based structures, where did the split actually pay off, and at what conversion volume per tier did the learning resets stop?

Priya, I don't think a blended tROAS is what overspends on cheap wins. The number is the problem. Break-even ROAS is 1 divided by gross margin, so a 60 percent hoodie breaks even around 1.7, while a 12 percent phone case needs about 8.3. Set a blended target of 4 and the hoodie is comfortably profitable, but the case loses money on every sale the algorithm buys at 4x. The algorithm is doing exactly what it was told. The target is just wrong for half the catalogue.

Before you build tiers, run one check. Pull 60 days of revenue, cost and gross profit per product, sorted by margin. If the thin-margin SKUs take a small share of spend, a cap or an exclusion will do more than three campaigns. If they eat most of the budget, the split is justified.

On volume, I wouldn't trust a fixed threshold. A tier with 15 to 30 conversions a month can be stable if order values are consistent, but one with wide order-value variance may need far more. So watch learning status and the week-over-week spread in actual ROAS for each tier, not just the conversion count.

Theo, your break-even maths holds, but check the denominator before you build anything on it. Gross margin ignores payment fees, shipping subsidies and returns, and returns hit thin-margin products hardest, because a refunded case often costs more than the margin it earned. Use contribution margin instead and your break-even moves. A 60 percent gross hoodie with a 25 percent return rate is a different product from the one in the spreadsheet.

Priya, on the tiering option: in Performance Max, tROAS is set at campaign level, not asset group level. Asset groups can hold different creative and audience signals, but they can't carry different targets. Custom labels can steer which products are eligible in each listing group, but they don't change the target the algorithm aims for. If you want a 1.7 target for hoodies and 8.3 for cases, you need separate campaigns. Those campaigns can end up bidding for the same users, so check overlap in search terms and placements before you trust the numbers.

Theo, sorting by spend share is the wrong way in. A thin-margin SKU might take 10 percent of budget but run at 3x against a break-even of 8.3, so it loses money on every order. What you want to size is the loss, which is (actual ROAS minus break-even) times spend for each product. A small slice of budget can still be most of the damage.

Exclusion in PMax isn't as clean as it sounds either. Excluding a product through a custom label filter pulls it out of the campaign entirely, so you lose its conversions along with the waste. If the case is often bought alongside a hoodie, part of its value shows up in the hoodie's numbers instead of its own. So before you exclude anything, compare the loss against what you'd give up in assisted sales.

Rule of thumb: if a product's break-even ROAS is more than about double the campaign target, it doesn't belong in that campaign. Exclude it or move it somewhere else. That's cleaner than trying to cap it, because the algorithm will keep buying it at a loss whenever it converts.