Who in your company is allowed to say how much you will spend on ads next quarter?
At $500K a month the answer is usually the founder, and the number is a feeling informed by last month. At $3M a month that stops working. The ad budget is now one of the largest lines in the plan. Finance wants to know what it buys. The board wants to know what happens if it stops. And the ad account, which used to be the constraint, may well be the most competent part of the operation.
This is the top rung of the ecommerce growth ladder. Below it, every constraint lived inside marketing: creative supply, a second channel, a new-customer engine, and then margin, offers, and AOV. Above $3M a month the binding constraints sit around marketing. There are three, and they are connected.
- Forecasting: a spend plan the finance team trusts enough to fund without cutting it mid-quarter.
- Incrementality: evidence of what each channel’s spend actually adds, strong enough to settle arguments between channels.
- Team: one named owner for each number, so decisions get made at the speed the spend requires.
A weak forecast gets the budget cut. Weak incrementality evidence turns the forecast into guesswork. A muddled team means neither gets built. Here is how to build each, and how to tell which one is binding for you.
Forecasting: plan from the marginal dollar, not the average
Most spend plans are built from average efficiency. Last quarter ran at a blended MER of 4.0, next quarter’s revenue target is $12M, so the plan is $3M of spend for the quarter. That arithmetic assumes the next dollar performs like the average dollar. At this size it does not, and the plan misses in the same direction every time: spend goes up, efficiency comes down, and finance stops believing the next plan.
A forecast finance can trust starts from a response curve: what each additional increment of spend has produced, in new customers, in your own business. You build it from your history of spend changes, from holdout tests, and, once you have enough history, from a marketing mix model. Then you plan to the point where the marginal customer still pays for itself, and you give finance a range rather than a point.
Platform forecasts help inside one channel, but they are not a business forecast. Google’s Performance Planner, for example, bases its forecasts on the last 7 to 10 days of data, refreshed daily and adjusted for seasonality. That is useful for sizing a Google budget. It knows nothing about Meta, your margin, or your repeat rate, and it forecasts the conversions Google reports, not the incremental ones.
The worked table: a month’s plan from a response curve
Here is the arithmetic for an illustrative brand, not a client result. It has run at four spend levels over time and matched the new customers each produced against its customer list. It values a new customer at $110 of contribution: the first order plus 90 days of repeat purchases, after product, shipping, and returns.
| Monthly paid spend | New customers | Average new-customer CAC | Marginal CAC of the last $100K | Contribution after spend |
|---|---|---|---|---|
| $600,000 | 9,000 | $66.67 | n/a | $390,000 |
| $700,000 | 10,000 | $70.00 | $100.00 | $400,000 |
| $800,000 | 10,800 | $74.07 | $125.00 | $388,000 |
| $900,000 | 11,400 | $78.95 | $166.67 | $354,000 |
Read the average column and every level looks fine. A $79 average CAC against $110 of value suggests room to spend $900K and beyond. Read the marginal column and the answer changes. The step from $700K to $800K bought 800 new customers at $125 each, $15 more than each is worth. Total contribution peaks at $700K and falls after it.
So the plan is $700K a month, because that is the last level where the final $100K earns more than it costs. The full treatment of the next-dollar number is in marginal ROAS, and the ad budget calculator is a quick way to rough out the spend side before you model it properly.
Then give finance the range. Over a quarter at $700K a month, the base case is 30,000 new customers and $1.2M of contribution after spend. If the curve turns out 10% weaker than measured, that becomes 27,000 new customers and $870,000. If it runs 10% stronger, 33,000 and $1.53M. A plan with a stated range and a named downside gets funded. A plan with a single number gets cut the first month it misses.
Four things the forecast has to include
- Seasonality from your own history, not a generic retail calendar.
- Inventory. A forecast that assumes stock you will not have is a forecast of stockouts. If one product carries a large share of revenue, the plan is only as good as that product’s supply, which is the risk described in hero SKU concentration.
- Creative supply. The curve you measured assumed a production cadence. If new concepts slow down, the curve bends sooner than the model says.
- A rule for mid-quarter changes, agreed in advance: which signal moves the plan up, which moves it down, and who decides.
Incrementality: test the biggest dollars in question first
At this spend, every channel reports that it is working. Brand search, retargeting, Performance Max, and prospecting all claim a healthy return, and the sum of their claims exceeds what the store banked. Incrementality is the question of how much of each channel’s reported result would have happened anyway.
The tools are documented by the platforms themselves. Google’s Conversion Lift measures the incremental conversions driven by Google Ads campaigns and reports incremental CPA and incremental ROAS. Meta’s Conversion Lift uses holdout groups to measure the incremental effect of Meta ads. Geo experiments, described by Google researchers in a 2011 paper, compare regions where spend changes with regions where it does not, and work across any channel. The end-to-end design for a Shopify brand is in the geo holdout test.
The hard part is not running a test. It is choosing which one, because each costs weeks, a held-out region or audience, and some revenue. My rule: rank every major line of spend by the dollars genuinely in question, and test the top of the list first.
Dollars in question is monthly spend multiplied by the width of your honest uncertainty about its incrementality. If you believe brand search is somewhere between 10% and 60% incremental, half its spend is in question. The ranges below are an illustrative team’s own prior guesses, written down before any test. The arithmetic is the point, not the ranges.
| Line of spend | Monthly spend | Incrementality, low to high guess | Dollars in question |
|---|---|---|---|
| Meta prospecting | $300,000 | 60% to 90% | $90,000 |
| Performance Max | $200,000 | 30% to 80% | $100,000 |
| Meta retargeting | $120,000 | 20% to 70% | $60,000 |
| Brand search | $80,000 | 10% to 60% | $40,000 |
On these numbers Performance Max is the first test, not because it is the biggest line but because it combines size with the widest uncertainty. Brand search, the usual first suspect, comes last: its range is wide, but the line is small. Once a test narrows a range, that line’s dollars in question fall and the next one moves to the top.
Feed every result back into the curve. A response curve built on platform-reported conversions overstates what the marginal dollar buys, most of all in the channels with the most non-incremental credit. Calibrated with lift results, the curve moves, and so does the plan. Google’s open-source model, Meridian, is built around calibrating a mix model with experiments for exactly this reason.
How to read a result: inside or outside the range
A lift result is only useful if it changes something, so decide before the test what each outcome will change. My rule: if the measured incrementality lands inside the range you wrote down, narrow the range to the test’s confidence interval, adjust that line’s value on the curve, and move to the next test. If it lands outside the range, your priors were wrong in a way that probably affects more than one line. Rebuild the curve before the next quarter’s plan, and revisit the ranges on every other line, not just this one.
Read the test window honestly too. Meta notes that reading results over post-test windows without maintaining holdout groups can increase variance and reduce the measured effect. A test ended early, or read long after the holdout was released, can produce a number that looks precise and is not.
Team: one owner per number, five numbers to own
The third constraint is the least technical, and it decides whether the first two get built at all. At $3M a month, a spend decision touches finance, merchandising, creative, analytics, and one or more agencies. If nobody owns the forecast, it is everyone’s spreadsheet and nobody’s commitment.
The rule is simple to state. Every number the business steers by has exactly one named owner, accountable for its accuracy and for proposing changes. Two owners means no owner.
| Number | Owner | What they decide | What breaks without them |
|---|---|---|---|
| Spend forecast and range | Head of growth, signed off by finance | Monthly spend level and mid-quarter changes | The budget is cut on the first miss |
| Incrementality priors and test plan | Measurement or analytics lead | Which test runs next and how results move the curve | Channels argue from their own dashboards |
| New-customer CAC by channel | Channel leads, in-house or agency | Budget within each channel | Returning customers get bought as acquisitions |
| Creative cadence | Creative operations lead | Briefs, volume, and when to retire tired ads | The curve bends early |
| Contribution per order | Merchandising or ecommerce lead | Offers, bundles, pricing, and returns work | Media ends up paying for margin problems |
Which roles belong in-house and which can sit with an agency? My answer, with the disclosure that I run an agency: keep in-house the numbers that depend on knowing the business intimately and that compound over years, which are the forecast, contribution per order, and the offer. Outsource where breadth across many accounts is the advantage, which usually means channel operation and help designing tests, with the reading of results kept in-house. Creative operations and analytics can go either way, depending on who you can hire. The longer argument is in in-house vs agency.
What does not work at this size is one agency that owns the forecast, the measurement, and the channel spend all at once. That is not a partnership. It is marking your own homework. Whoever spends the money should not be the only one who grades it.
Which of the three is binding: four decision rules
- If finance has cut or frozen the ad budget mid-quarter in the last year, forecasting is binding. Build the curve and the range before anything else.
- If platform-reported revenue added up across channels sits well above what Shopify banked, and nobody can say which channel to cut first, incrementality is binding. Rank the dollars in question now and schedule the first test for January, when peak weeks can no longer distort it.
- If spend decisions wait on a meeting, or two teams report different CAC for the same month, the team is binding. Assign owners before building anything, or nothing you build will be maintained.
- If none of these is true, you may be standing on a lower rung. The margin and AOV rung is the nearest one to check.
What “test everything” gets wrong
The standard advice at this size is to test everything: lift studies on every channel, holdouts on every tactic, an experiment calendar that never stops. It sounds rigorous. I think it is wrong, for a reason that is arithmetic rather than philosophical.
Every test has a price: the revenue lost in the held-out region or audience, the weeks of waiting, and the analyst time to read it. Tests of small lines rarely have the statistical power to detect the effect you care about, because the signal is small next to the normal noise in daily sales. A test that cannot detect a plausible effect is not cautious. It is a way of spending money to produce a shrug.
Testing the top of the dollars-in-question list properly beats testing the whole list badly. A few well-powered tests a year, each fed back into the forecast, change more decisions than a calendar of underpowered ones. For the smaller lines, the lighter options in incrementality testing without a data team are usually enough.
When this advice is wrong
- You reached $3M a month on one extraordinary product or moment rather than a repeatable engine. The constraint may still be the new-customer engine, and a forecast will only formalize a spike.
- Contribution per order is not yet known. A response curve valued in revenue will plan you straight into losses. Fix the margin arithmetic first.
- Most of your revenue runs through marketplaces or retail. A curve built on store orders describes only part of the business, and the effect of your ads on other channels needs its own measurement.
- You are about to change the business materially: a new category, a new country, a price change. A curve measured on the old business will not describe the new one. Plan with wide ranges and retest early.
What to do this month
- Pull the last six to twelve months of spend changes and the new customers each produced, matched against your customer list. Plot the curve, even roughly.
- Put a value on a new customer: first-order contribution plus a defined window of repeat purchases. Write the window down.
- Find the spend level where marginal CAC crosses that value. That is the plan. Add a 10% weaker case and a 10% stronger case and take all three to finance.
- List every line of spend above 5% of the total, write a low and a high incrementality guess beside each, and rank the lines by dollars in question.
- Name one owner for each number in the ownership table, and put the names where everyone can see them.
Running the channels and helping design the tests, with the forecast and the verdicts staying on your side, is the work our enterprise PPC engagements are built around. If you want an independent read of your ad accounts first, start with the free audit.
Common questions