Method comparison

Ways to split ad budget — even, historical, marginal efficiency

How should I split ad budget across channels?

Split on marginal efficiency, not average efficiency.

The most common budget mistake is adding spend to the channel with the lowest average CPA. Channels usually face diminishing returns, so the efficiency of the next dollar differs from the average. Moving the allocation rule from average to marginal is the point of this comparison.

Side by side

Budget allocation methods
MethodRuleStrengthWhere it breaks
Even splitDivide by number of channelsEasy to explain and unbiasedIgnores efficiency differences entirely
Historical shareKeep last period's spend sharesLow variance and operationally stableKeeps growing already-saturated channels
Average CPA weightingGive more to the lowest average CPAIntuitive and simple to computeIgnores diminishing returns and overinvests
Marginal efficiencyEqualize the efficiency of the next unit of spendMaximizes total outcomeRequires estimating a response curve per channel

Which one to use

When spend varied enough per channel
Allocate on marginal efficiency. The curve is estimable.
When spend was nearly flat
No curve emerges. Vary spend deliberately first to widen the observed range.
When channels differ by less than 1.2x
There is little to reallocate. Look at creative or targeting instead.
Before shifting a large amount
Check saturation to see whether the channel is already in the flat region.

Frequently asked questions

Why not put everything into the lowest average CPA channel?

It is not recommended. Most channels lose efficiency as spend rises, so a good average can coexist with poor efficiency on the next dollar. The decision rule is marginal, not average, efficiency.

How is marginal efficiency estimated?

Fit a curve to the relationship between spend and outcome per channel, then take the rate of change at the current spend level. If past spend barely varied, that curve estimate is unstable.

Can I execute the recommended split as-is?

Moving everything at once is not recommended. A fitted curve is only trustworthy inside the observed spend range, so shift gradually, observe the result, and re-estimate.

Tools for this

← All method comparisons