How to Allocate Retail Media Budget Across Platforms

Retail media budget allocation is one of the most consequential decisions a media planner makes each week. The client has a fixed budget. The planner’s job is to distribute it across Amazon, Walmart, Criteo, and potentially several other networks in a way that maximizes return on the total investment.

In practice, this decision is almost always made by comparing platform-reported ROAS figures and shifting budget toward whichever platform is reporting the highest number. It is a logical approach. It is also systematically wrong, for reasons that compound over time.

Why Raw ROAS Is the Wrong Input for Allocation

Platform-reported ROAS figures are internally consistent within each platform’s measurement methodology and structurally incomparable across platforms. Each network defines its own attribution window, its own model type (first-click, last-click, linear), and its own rules about which interactions count. The result is a set of numbers that appear comparable—all expressed as revenue per dollar of spend—but are measuring different things.

Using these numbers for allocation decisions is equivalent to comparing the prices of two apartments where one is quoted in square feet and the other in square meters, without converting. The ratio looks meaningful until you realize the denominator is different.

The second problem with using reported ROAS for allocation is that ROAS is an average, not a marginal metric. A platform reporting a 5x ROAS on $50,000 of spend does not mean the next $10,000 of spend on that platform will also generate 5x. Almost universally, the next dollar spent on any platform generates less return than the current average—because the most efficient inventory is always purchased first.

The Two Inputs That Actually Matter

Normalized ROAS: the fair comparison

Before any allocation decision, all platform ROAS figures must be brought to a common measurement standard. This means correcting for attribution window length, model type, and view-through inclusion simultaneously. The output is a normalized ROAS figure for each platform that reflects performance under identical measurement conditions.

Without normalization, allocation decisions systematically favor platforms with generous measurement methodologies. With normalization, the comparison reflects actual performance differences rather than measurement differences.

Marginal ROAS: the allocation signal

The correct question for budget allocation is not “which platform has the highest average ROAS?” It is “which platform generates the most additional revenue from the next dollar of spend?” This is marginal ROAS (mROAS)—the derivative of the revenue curve at the current spend level.

Marginal ROAS is lower than average ROAS for any platform that is not severely underfunded. The relationship follows a diminishing returns curve: early spend on a platform is highly efficient because it captures the best auction opportunities; later spend captures progressively less efficient inventory. The optimal allocation across platforms is the one where marginal ROAS is equalized—because if platform A is returning more at the margin than platform B, moving a dollar from B to A increases total revenue.

Return Curves: Modeling Marginal Returns

A return curve models the relationship between spend and revenue on a specific platform. At low spend levels, revenue grows steeply—there is plenty of efficient inventory available. As spend increases, revenue growth decelerates as the most efficient placements are exhausted and the campaign moves into lower-quality inventory at higher prices.

The mathematical form that best describes this relationship is a Hill curve (also called a saturation curve):

Revenue(spend) = R_max × spendⁿ / (Kⁿ + spendⁿ)

Where R_max is the maximum achievable revenue at infinite spend, K is the spend level at which revenue reaches 50% of R_max, and n is the steepness of the curve. These parameters are fitted to historical spend-revenue data for each platform using nonlinear regression.

Once the curve is fitted, marginal ROAS at any spend level can be calculated as the first derivative of the curve at that point. This gives the planner a precise estimate of how much additional revenue each incremental dollar would generate on each platform.

A Worked Example

An agency manages $120,000 monthly across three platforms for a household goods brand. Current allocation and normalized performance:

The normalized ROAS figures show Amazon leading, followed by Walmart and Criteo. But the marginal ROAS signal is more instructive: Amazon is returning 3.8x at the margin while Walmart is returning only 2.4x. Every dollar shifted from Walmart to Amazon generates an additional 1.4x in revenue on that dollar.

The optimized allocation shifts approximately $12,000 from Walmart to Amazon, equalizing marginal returns near 3.1x across platforms and projecting an 8.3% increase in total revenue on the same $120,000 budget.

Constraints That Bound the Allocation

Pure marginal return optimization ignores real constraints that planners must account for. Allocation recommendations that ignore these constraints will not survive client review.

Minimum presence requirements

A client may have contractual minimum spend commitments with specific platforms, or strategic reasons to maintain presence on a network regardless of short-term return. The allocation model must respect these floors while optimizing within the remaining budget.

Auction saturation limits

On some platforms, particularly Amazon Sponsored Products in competitive categories, increasing spend above a certain level produces rapidly diminishing marginal returns because the most efficient auction slots are already captured. The return curve steepens sharply at these saturation points. The allocation model must identify these thresholds and treat them as effective spend ceilings.

Attribution confidence

Not all platform data has equal reliability. A platform with only two weeks of data has a return curve fitted to a small sample. The confidence interval around the marginal ROAS estimate is wide, meaning the allocation recommendation carries higher uncertainty. Planners should weight allocation changes toward platforms where the data is more mature.

Presenting Allocation Changes to Clients

The most technically correct allocation recommendation is worthless if the client won’t act on it. Presenting a budget reallocation requires three things: a defensible methodology, an expected outcome, and a validation plan.

The defensible methodology is normalization. The recommendation is based on comparable numbers, not platform-reported figures where every platform looks good within its own rules. This is the answer to the client who says “but Walmart’s ROAS is higher.”

The expected outcome is a projected revenue increase on the same total budget, expressed as a percentage and a dollar figure. Clients respond to concrete projections, not abstract optimization theory.

The validation plan is a commitment to review results over the next two to four reporting cycles. Reallocation effects take time to materialize as campaigns optimize to new budget levels. Setting this expectation upfront prevents the client from reversing the change before it has time to demonstrate results.