The average that's destroying your marketing decisions

A healthy ROAS on the account report can sit on top of a campaign that loses money on every sale it is credited with. The budget keeps flowing to it, because the headline number says everything is fine.

The blended ROAS looks healthy. The component parts tell a different story. Budget is allocated to the blend. The component parts determine what the business actually returns.

What the average hides

The account below is an anonymised composite of accounts I have reviewed; the numbers are illustrative and have been changed.

A Google Ads account spending $15,000 per month shows a blended ROAS of 4.2x. On the headline number alone, that reads as a performing account.

Break it down by campaign type:

Branded search: $1,500 spend, 11x ROAS
Generic search: $5,000 spend, 4.25x ROAS
Performance Max: $6,000 spend, 3.75x ROAS
Prospecting display: $2,500 spend, 1.1x ROAS

The blended 4.2x is accurate. It is also a mathematical artefact that obscures the fact that $2,500 per month, one sixth of the total budget, is operating at a ROAS that, on a 28% gross margin, is deeply unprofitable. Breakeven at 28% margin is 3.6x. The display campaign at 1.1x is losing money on every attributed conversion.

The account looks healthy. One campaign is a problem. The average conceals it.

This is an averaging problem: one number standing in for groups that behave very differently. It sometimes gets called Simpson's paradox, which is a stronger and rarer case, where a relationship that holds in every subgroup reverses once the groups are combined (Simpson, 1951; Bickel et al., 1975). Nothing reverses here. One weak campaign is simply diluted by three stronger ones.

Why averages are structurally misleading for budget decisions

Budget allocation follows performance signals. If the account shows 4.2x, the instinct is to increase budget to the account, or at minimum, to maintain current allocation.

But budget does not go to the account. In Google Ads it is set per campaign, or per shared budget. If budgets are raised across the board, say 20% on every campaign, the 1.1x campaign grows alongside the 11x campaign.

Every dollar added that way is divided across all the underlying economics. The average looks stable. The actual marginal return on new spend is determined by which campaigns absorb it, and the blended number does not tell you which ones those are.

This is how accounts scale in a way that feels productive and deteriorates economically. The blend holds. The composition shifts.

The categories that averages routinely conceal

Branded vs. non-branded split. Branded campaigns capture existing demand at high efficiency. Non-branded campaigns generate new demand at higher cost. Blending them produces a number that overstates the value of non-branded spend and understates the cost of acquiring genuinely new customers. A business making budget decisions on blended ROAS is making decisions based on a number that includes conversions it would have received anyway.

Product categories with different margins. A 4.5x ROAS on a 40% margin product is profitable. A 4.5x ROAS on a 15% margin product is not. When a campaign or ad group promotes both, the ROAS averages across them. The budget decision treats them as equivalent. The margin calculation does not.

New vs. returning customers. Platform attribution often counts returning customer purchases at the same value as new customer acquisitions. A business paying $45 in customer acquisition cost (CAC) to re-acquire a customer who costs $8 to retain through email has a CAC problem hidden in its acquisition numbers.

Seasonal performance averaged into a stable baseline. A campaign with a 7x ROAS in November and a 1.8x ROAS in March has an "annual average" of somewhere around 4x. Budget set to that annual average is over-invested in March and under-invested in November.

What disaggregation actually requires

Breaking the average down into useful components requires three things.

Segment by the dimension that matters to the business. For e-commerce businesses, the splits that usually matter are: branded vs. non-branded, campaign type, product category margin, new vs. returning customer. Each of these produces a different economic picture.

Apply the right denominator. ROAS alone is the wrong measure for budget decisions because it ignores gross margin. The correct comparison is ROAS against the breakeven ROAS for each segment. A campaign at 3.8x against a product category with 24% margin (breakeven: 4.2x) is unprofitable. That number does not appear in the platform report.

Evaluate marginal return, not average return. The average ROAS across all spend tells you what the existing budget is producing. It does not tell you what additional budget will produce. Marginal return is the return on the next dollar spent. It sits below average return because of diminishing returns: once a campaign has reached the buyers most likely to convert, each extra dollar reaches people a little less likely to. In some campaigns the marginal return is negative while the average holds.

The decision that follows from this

Once the account is disaggregated, the budget decision changes.

The question is no longer "should we increase total spend?" The question becomes: which segments are operating above breakeven margin-adjusted ROAS, what is their capacity for additional spend before returns diminish, and which segments are operating below breakeven and should not receive additional budget regardless of how the blend looks?

These are different questions. They produce different answers. They require working from the component economics rather than the aggregate.

Accounts are often managed at the aggregate level because that is how platforms present data and how reporting is usually structured. The aggregate is useful for a quick read on performance. It is not useful for budget decisions, because budget does not flow to the aggregate. It flows to the components.

The average is real. It is also the least informative number in the account for the decision that matters most.

References

Bickel, P. J., Hammel, E. A., & O'Connell, J. W. (1975). Sex bias in graduate admissions: Data from Berkeley. Science, 187(4175), 398–404.
Simpson, E. H. (1951). The interpretation of interaction in contingency tables. Journal of the Royal Statistical Society: Series B, 13(2), 238–241.