Blog » Strategy

Marketing Budget Allocation: Average ROAS Is Not Marginal

Name: kpitree.io
Availability: InStock
Author: kpitree.io

October 23, 2026 · 12 min read

Marketing budget allocation from a KPI tree ranks channels by average ROAS. Moving spend to the highest ROAS channel can cut total revenue.

The Channel They Cut Got Better and Revenue Fell

A direct-to-consumer brand spent 1,000,000 dollars on paid acquisition last quarter, and its agency built a KPI tree to find the waste.

The tree was correct. Brand search took 200,000 dollars and returned 1,600,000 in attributed revenue, a ROAS of 8.00. Non-brand search took 300,000 and affiliate took 100,000, each returning a ROAS of 4.00. Paid social took 400,000 and returned 1,200,000, a ROAS of 3.00. Blended ROAS was 4.40.

The agency moved 100,000 dollars from paid social, the worst channel in the tree, into brand search, the best. The case took one line: 100,000 dollars earning 8.00 instead of 3.00 adds 500,000 of revenue.

The next quarter held total spend at 1,000,000 and produced 4,250,000 of revenue. Non-brand search and affiliate were untouched. Brand search landed at 1,720,000 on 300,000 of spend, a ROAS of 5.73. Paid social landed at 930,000 on 300,000 of spend, a ROAS of 3.10. Blended ROAS fell from 4.40 to 4.25.

The 100,000 added to brand search brought back 120,000. The 100,000 taken from paid social cost 270,000. The channel they starved improved, the channel they fed deteriorated, and the total fell 150,000 dollars. No node in the tree was wrong. It had been asked a question it cannot answer.

Total spend was unchanged, 100,000 dollars moved from a 3.00 ROAS channel to an 8.00 ROAS channel, and revenue fell 150,000. A tree built from realized spend reports all three figures and can justify none of the decision.

What Is Marginal ROAS and How Is It Different From ROAS?

Average ROAS is a channel's total attributed revenue divided by its total spend. Marginal ROAS is the revenue the next dollar produces, holding everything else fixed. The first is a division of two numbers already in the data. The second is a counterfactual, and no arithmetic performed on money already spent will produce it.

Return on ad spend is a ratio node. In a KPI tree it sits above two summable columns, revenue and spend, and divides one by the other at whatever level is being read.

That node answers a backward-looking question exactly: for every dollar this channel has consumed so far, how many dollars of attributed revenue came back. It is the average over the whole spend, including the first dollar, which was almost certainly the most productive one.

Marginal ROAS answers the forward question. Google's Meridian documentation defines marginal ROI as the incremental outcome generated by a small increase in spend at the current spend level, read off a response curve rather than off a ratio.¹

Wherever a channel shows diminishing returns, every dollar after the first returns less than the one before it, so the marginal rate sits below the average rate. The average is held up by cheap early dollars that have already been bought and cannot be bought again.

Why Does a KPI Tree Rank Channels Backwards for Budget Decisions?

A KPI tree is an accounting identity over money already committed, so every node reconciles to what happened. Ranking channels by average return assumes the next dollar earns the average. The channels with the highest averages are usually the ones nearest saturation, so their next dollar earns the least, and the ranking inverts.

The inversion is not a quirk of the example. It is the normal case, because the conditions that produce a high average ROAS are the same conditions that produce a low marginal one.

A channel earns a high average when it is small against its addressable demand and when it captures demand that was already there. Brand search is the clearest instance of both. It is capped by the number of people searching the brand name, and most of those people already intended to buy.

A channel earns a low average when it is large, broad, and buying colder audiences. That is also the description of a channel with room left on its curve.

So the tree ranks by a quantity that tracks scarcity of headroom, and the decision reads that ranking as abundance of headroom. The arithmetic is sound and the inference runs backwards.

The Elasticity That Links the Two Numbers

One clean relationship connects the two, and it is worth memorizing.

If a channel's revenue responds to spend as a power curve, revenue equal to a constant multiplied by spend raised to the power beta, then marginal ROAS equals beta multiplied by average ROAS. Beta is the spend elasticity: the percentage change in revenue produced by a one percent change in spend. It runs between 0 and 1 wherever returns diminish.

Apply it to the opening quarter, with elasticities assumed for the illustration rather than measured. Brand search at a ROAS of 8.00 and an elasticity of 0.15 has a marginal ROAS of 1.20. Paid social at a ROAS of 3.00 and an elasticity of 0.90 has a marginal ROAS of 2.70.

The averages differ by a factor of 2.67 in brand search's favor. The marginals differ by a factor of 2.25 in paid social's favor. Elasticity decides which way the money should move, and elasticity is not in the tree.

Those are the two rates the quarter actually produced.

Average, Marginal and Incremental Return Compared

Three measures carry the word return and support three different decisions. Treating them as interchangeable is what produced the 150,000 dollar loss.

The third row is the one most often skipped. A response curve fitted to historical spend variation estimates the marginal rate from correlation. A holdout measures it from a randomized difference. When the two disagree, the experiment is the arbiter, and they do disagree: Gordon, Zettelmeyer, Bhargava and Chapsky compared observational methods against large randomized experiments at Facebook and found the observational estimates often far from the experimental benchmark, in both directions.²

None of the three is a property of a channel. Each is a property of a channel at a spend level, in a period, under one attribution rule. Change any of those and the number changes.

Measure	What it divides	Where it comes from	What it decides
Average ROAS	Realized revenue by realized spend	A KPI tree over uploaded spend and revenue columns	Reconciliation, mix and trend. Not allocation
Marginal ROAS	Change in revenue by change in spend, at the current level	A response curve fitted to historical spend variation	Where the next dollar goes, inside the observed range
Incremental ROAS	Treated minus untreated revenue, by the spend difference	A randomized holdout, usually by geographic region	Whether the channel pays at all, at the tested level

Brand Search Is Where the Gap Is Widest

The canonical evidence was published in 2015 and still stands.

Blake, Nosko and Tadelis ran large-scale paid search experiments at eBay and reported them in Econometrica. Switching off branded keyword advertising produced close to no measurable drop in traffic or sales, because the users who had been clicking those ads arrived through organic links instead.³

That is a marginal return near zero underneath an average return any attribution model reports as the highest in the account. The ad was credited with revenue it did not cause, in proportion to how ready the customer already was.

The same mechanism operates at smaller magnitude wherever retargeting, branded display and loyalty email appear in the tree. Each sits close to the conversion, collects a large share of last-click credit, and has the least headroom in the account.

A tree built on attributed revenue inherits every one of those credits and reproduces the attribution model's opinion with perfect arithmetic.

The channels an attribution model ranks highest are, by construction, the ones closest to the conversion. Proximity to the conversion and causal contribution are different things, and only one of them is in the tree.

How Do You Measure Marginal Return Without a Media Mix Model?

Run a holdout. Cut or stop the channel in a randomly chosen set of geographic regions, hold it steady everywhere else, and compare the two groups over a fixed window. Vaver and Koehler published the standard geo experiment method at Google Research in 2011. The result is one point on the curve, at one spend level, and it ages.

A geo holdout is the cheapest honest answer available to a mid-sized advertiser, and it needs no model, no data science team and no vendor.⁴

The design is simple. Split regions into a treatment group and a control group at random, change spend in one group only, and measure the difference in outcome against the difference in spend. What comes back is an incremental return at the tested spend level.

The cost is the point of it. A holdout deliberately forgoes revenue in the control regions to learn what that revenue was worth. Treat the forgone amount as the price of the measurement rather than as waste.

The limits are real. A holdout returns one number per channel per test, it takes weeks, and it cannot run on every channel at once without contaminating the comparison. Media mix modeling fits a curve across all channels from historical variation instead, which is cheaper and weaker. Google's Meridian optimization documentation is explicit that its reallocation recommendations are bounded by the spend range the model actually observed.⁵

The Rule Is Equal Marginal Return, Not Equal ROAS

A fixed budget is allocated correctly when the last dollar in every channel earns the same marginal return. If one channel's marginal return exceeds another's, moving a dollar across the gap raises the total.

Run that rule on the opening quarter. Brand search was returning 1.20 at the margin and paid social 2.70, a gap of 1.50 in paid social's favor. The correct move was the one the agency inverted.

Moving 100,000 dollars from brand search into paid social gives up 120,000 and gains 270,000. Revenue rises 150,000 to 4,550,000 and blended ROAS rises to 4.55, on identical total spend. The same 100,000 is worth 300,000 of swing depending on which way it travels, and the tree points at the losing direction with complete internal consistency.

One caution. Marginal rates are local. Treating 1.20 and 2.70 as constant across a 100,000 dollar step is a first-order estimate, and the rates converge as money moves. Beyond a third of a channel's budget, re-measure rather than extrapolate.

Where the Tree Is Still the Right Instrument

None of this makes the decomposition useless. It is the wrong tool for allocation and the right one for four other jobs.

It reconciles. Channel spend sums to total spend, so a number quoted in a meeting traces back to a row. A media mix model gives a coefficient with an interval and no audit trail to the invoice.

It fixes denominators. Blended customer acquisition cost computed as the average of channel CAC values is not blended CAC. A tree that divides summed cost by summed customers returns the right figure whatever the grouping, the same failure as averaging a ratio column in a spreadsheet.

It separates rate from mix. When blended ROAS falls and no channel's ROAS falls, the move is entirely mix, and the price, volume and mix decomposition sizes each part.

It exposes saturation over time. Plot a channel's average ROAS against its spend across eight quarters. A channel whose average falls as spend rises is on the diminishing part of its curve, and the slope between two points is a crude marginal estimate. Crude beats absent.

Five Tests Before Moving Budget

Run these in order. The first three take an afternoon.

1. The attribution test. State which model produced the revenue column and how it treats assisted conversions. If the answer is last click, every channel near the conversion is overstated and the ranking is unusable for allocation.

2. The headroom test. Divide this quarter's channel spend by the largest quarterly spend that channel has absorbed without its average ROAS falling. Above 0.8, there is little room left, whatever the average says.

3. The direction test. Plot average ROAS against spend for each channel across the last eight periods. A negative slope means diminishing returns and a marginal rate below the average. A flat slope means the channel has never been pushed hard enough to find its ceiling.

4. The step size test. Cap any single reallocation at a size the receiving channel has previously absorbed. A 50 percent increase lands outside the range any model or history has seen, so no measured rate applies.

5. The holdout test. Above a defined threshold, run a geo holdout on the receiving channel first. A reallocation worth arguing about is worth measuring, and a tree that reconciles perfectly can still be wrong.

Building the Allocation Tree in kpitree.io

kpitree.io builds this tree from one uploaded CSV. The file needs one row per channel per period and the summable columns under every rate: spend, attributed revenue, customers acquired.

Inside the tree, parent to child relationships are addition and subtraction, so channel spend genuinely sums to total spend. Every rate is a derived node dividing two summed columns at the level being read, which is why blended ROAS recomputes to 4.25 when the channel weights move.

Period comparison is where the saturation signal shows up. Put eight quarters in the file, group by channel, and read each channel's average ROAS against its own spend. That reading is the direction test.

What it does not do is produce a marginal rate. There is no response curve inside the tree and no experiment behind it. It computes the identities, shows which channel absorbed which dollars, and reports the direction an average is moving. The elasticity comes from a holdout or a model built outside it.

CSV upload is the evidenced ingest path today. The smallest useful version is sixteen rows: four channels across four quarters, with spend and revenue.

Frequently Asked Questions

Is average ROAS ever the right basis for a budget decision? Only when every channel sits on a flat part of its response curve, which is rare and testable with the direction test.

What elasticity should I assume if I have no model? None. Run the headroom and direction tests instead, which use data already in the tree. An invented elasticity is an invented decision.

How large does a budget have to be to justify a geo holdout? Large enough that the decision at stake exceeds the revenue forgone in the control regions. Marketing budgets stood at 7.7 percent of company revenue in Gartner's 2025 CMO Spend Survey, read October 2026, so a single channel reallocation usually clears that bar.⁶

Does this apply outside marketing? Yes. Any node where a ratio ranks where the next unit of a shared resource goes has the same defect, including headcount ranked by revenue per head and capital by divisional return.

Should the KPI tree and the planning model be the same artifact? No. The plan projects forward from assumed drivers and the tree explains backward from realized ones, a split the Association for Financial Professionals tracks in its integrated planning benchmarking⁷ and one a driver-based plan and a driver tree answer differently.

Closing: Rank the Next Dollar, Not the Last One

A ratio over money already spent describes history. An allocation decision is about money not yet spent. Those are different objects, and the tree holds only the first.

Use the decomposition for what it is good at. Reconcile the spend, fix the denominators, separate rate from mix, and watch each channel's average ROAS against its own spend across periods. That last reading is the only saturation evidence a tree can give.

Then stop. Before money moves, name the elasticity or measure it. A geo holdout costs real revenue in the control regions and returns the one number the tree cannot compute.

Read that way, the opening quarter stops being a story about a bad agency. It becomes what it was: a correct tree, read as a ranking, pointing 100,000 dollars at the channel with the least room left in the account.

Upload one CSV with your channels across four quarters, and check whether your best ROAS channel's average has fallen every time you have fed it.