A measurement plan is a short written agreement on which business decisions you need evidence for, which method will answer each one, who owns it, how often it runs and what result would change the decision. For a marketing mix modeling project, the plan starts with a brief: one page that names the decision the model supports, the outcome it should explain and why, the time horizon, the channels and levers, the known events and constraints, and the evidence you already have. If you only do one thing before commissioning a model, write that page, because the brief is what turns “build us an MMM” into a project that can actually finish.
I’ve seen more MMM projects struggle because of a vague question than because of bad data or the wrong library. When I build marketing mix models for clients, the first meeting is almost always spent on the brief, and the most useful question in that meeting is “who will act on this, and what could they actually change?”. This page covers what goes in the brief and why, gives you a builder that exports one, and then shows how the brief fits into a wider measurement plan alongside experiments and platform data.
Give the model a decision to support
The Vexpower course opens with a familiar situation. In the invented Northstar case, the marketing lead asks “What is our ROI?” and the finance lead asks “What should we do with next quarter’s £300,000 budget?”. Those questions overlap, but they’re not interchangeable. A historical return describes the past under past spend levels, while a planning decision also needs the effect of changing spend, the uncertainty around it, costs and practical limits. A model built to answer the first question can be a poor fit for the second, and you don’t want to find that out in the readout.
So the brief starts with the decision, and the course boils it down to four parts:
| Part | Northstar’s example |
|---|---|
| Decision | Allocate next quarter’s £300,000 across the current channels |
| Outcome | Incremental revenue, with a separate contribution-margin check |
| Horizon | The quarter, with enough follow-up to include carryover |
| Constraints | Keep a minimum brand presence; do not exceed the team’s delivery capacity |
The model has a role in this decision, but it doesn’t get to choose the business’s objectives or invent its operating constraints. Those come from the people running the business, and they should be written down before anyone looks at a coefficient, otherwise there’s a strong temptation to choose the objective that makes the model’s answer look good. A useful test is whether you can complete this sentence: “We will use the model to decide ___, over ___, measured by ___, while respecting ___.” If you can’t, the project isn’t ready for a model yet.
It’s also worth noticing the missing information early. “Maximize revenue” isn’t enough if the business can’t fulfill extra orders. “Maximize ROI” can prefer a tiny budget with a high average return over a larger, more valuable opportunity, because the first pounds in any channel are always the cheapest (that’s what a saturation curve means). The objective has to match the decision, which usually means “the most contribution from the budget we’re prepared to spend, within these limits”.
Choose the outcome KPI, because revenue, units and profit are different questions
The outcome is the quantity the model explains, and its definition determines the units of every contribution and return estimate that comes out. In the Northstar case, the brand sells 1,000 extra units in a promotion week while prices also fell, so a model of units and a model of revenue tell different stories without either calculation being broken. The units follow the outcome:
| Model outcome | Estimated contribution | Divide by spend to get |
|---|---|---|
| Revenue | Incremental revenue | Revenue per £1 of advertising |
| Orders | Incremental orders | Orders per £1 of advertising |
| Units sold | Incremental units | Units per £1 of advertising |
An order is not revenue, and revenue is not profit. Suppose a campaign is estimated to generate £25,000 of incremental revenue on £10,000 of advertising, a revenue return of £2.50 per £1. If only 30 percent of revenue remains after variable costs, that’s £7,500 of contribution before a £10,000 advertising bill, so the revenue result is positive while contribution after advertising is negative. This is why I like briefs that name a revenue or order outcome for the model and a margin for the decision, so the conversion to contribution is agreed in advance rather than argued about later (the breakeven ROAS guide has the arithmetic).
The definition should be specific enough to travel with every chart: gross or net of returns, which geography, new or all customers, which currency, which window. “Incremental net revenue per £1 of media spend, UK, 13-week window” is a KPI. “ROI” is not. If you’re inheriting an existing report, a quick exercise is to write down the numerator and denominator of its return metric in full. If the definitions are missing, that’s your first measurement question.
What a good MMM brief contains
Beyond the four core parts, a brief that’s ready to hand to an analyst or vendor answers ten questions. I’ve listed them in the order I’d fill them in:
- Decision. What someone will change because of the model: the budget, the period and the levers that can move.
- Decision owner and date. Who acts, who signs off, and when the decision is made, which sets the deadline for the whole project.
- Outcome KPI and why. The quantity the model explains, defined in full, and why it matches the decision.
- Time horizon and granularity. History available (two or more years of weekly data is the usual starting point; see MMM data requirements), geography, and the planning horizon including carryover.
- Channels and levers. Every paid channel with the metric you’ll use, plus non-media levers like price, promotions, distribution and email.
- Known events and other drivers. Holidays, promotions, price changes, outages, stock-outs, launches and competitor moves, with dates. These become the control variables that stop the model crediting media for things it didn’t do.
- Constraints. Total budget, channel minimums and maximums, contracts, fulfillment capacity and brand presence: management limits, not model outputs.
- Existing evidence. Experiments, lift studies, holdouts, platform reporting and previous models, with dates and scope, and whether any have already been used to calibrate something.
- How results will be used. Where the output goes (quarterly allocation, annual planning) and what it won’t be used for (weekly bidding, creative decisions).
- What would change the decision. The threshold, agreed before the model runs, that would make you choose differently.
Items 6 and 8 are the ones most often left thin, and they’re the ones that cause the most trouble later. The known-events list is mostly institutional memory, so get it from the people who were there; Google’s Meridian documentation, for instance, suggests asking marketing managers how non-media treatments such as price changes or promotions were timed and decided, because how a lever was set tells you whether it’s confounded with demand. The existing-evidence list matters because the same experiment can easily get used twice, once as a prior and again as “validation”, which creates the appearance of two pieces of evidence where there is one.
The builder writes the one-sentence brief at the top from your decision, horizon, KPI and constraints fields, and marks anything you haven’t filled in as “to be agreed” rather than dropping it. That’s deliberate: a gap in the brief is useful information to take into the kickoff meeting. Click “Fill with the example” to see a complete Northstar brief, then overwrite it with your own.
From brief to measurement plan
An MMM brief covers one model, but most marketing teams have several sources of evidence that each answer a different question. In the Northstar case the team has a platform dashboard, customer research, web analytics and an MMM proposal, and they want one number that settles every disagreement. The course is blunt that this expectation is the first thing to change. Different methods answer different questions, at different levels of detail and under different assumptions:
| Evidence | A useful role | A limitation to investigate |
|---|---|---|
| Platform reporting | Delivery, optimization and attributed outcomes | Attribution rules do not establish all incrementality |
| Analytics | Recorded customer journeys and site behavior | Missing devices, consent and cross-channel pathways |
| Experiments | A defined treatment contrast in a designed comparison | Scope, precision, interference and transferability |
| MMM | Aggregate causal modeling and planning scenarios | Data, identification and model assumptions |
A measurement plan puts that table into practice. It’s a list of business questions, each with the method that answers it, the outcome metric, the population, an owner, a cadence, the next read-out date and the threshold that would change the decision. The MMM shows up as one row (or a few), next to the incrementality tests that check its riskiest assumptions and the platform reporting that handles weekly pacing. If you want a broader view of how these methods fit together, the marketing measurement framework page has a selector for which method answers which question, and the MMM vs multi-touch attribution guide covers the most common disagreement.
When two methods disagree, investigate before averaging. A geo experiment may estimate the effect of extra spend in selected regions, an MMM may report a channel’s average historical contribution against a different baseline, and a platform may report conversions credited within a specific window. Those numbers need definitions before they need a reconciliation formula, and the disagreement often reveals an estimand mismatch, missing context or a weakness worth fixing. The plan should make each method’s question explicit enough that this check is easy.
Build a learning loop, not a one-off report
The plan earns its keep over time. Start with a business decision, record what’s already known, identify which uncertainty could change the action, and pick the evidence that would reduce it. Then use each new result to improve the next question: a geo test on social at higher spend can calibrate the MMM, which narrows the next quarter’s scenarios, which tells you which channel to test next. Keeping an evidence register (the question, method, population, dates, comparison, uncertainty, and whether the result already informed another model) helps avoid both false agreement and double-counting.
The workbook below has the plan, a methods reference, the evidence register and a summary sheet that counts questions per method and flags any read-out that’s overdue or missing an owner or threshold.
Mistakes that are cheaper to fix in the brief
Most of the problems I see in MMM readouts could have been caught in the brief. The objective doesn’t match the decision (the model maximizes ROI when the business wanted contribution). The outcome isn’t defined, so the revenue model and the finance spreadsheet disagree on returns and nobody knows which to trust. There’s no decision owner, so the results get presented, admired and never used. Known events were never listed, so a website outage or a price rise gets absorbed into a channel coefficient. And nobody agreed in advance what result would change the decision, which makes it very easy to accept the model when it agrees with what people already believed and to question it when it doesn’t.
None of these need sophisticated statistics to fix. They need an hour with the decision owner and a page of writing, which is cheap compared with a model refit. When the results do come in, the brief is also what you return to when you present the recommendation: has the model answered that question, a narrower one, or a different one?
Templates to take away
The brief template has the one-sentence brief, the ten sections above with a prompt and an example for each, and a sign-off table, in Word and Markdown. It’s the starting point the Vexpower marketing mix modeling course uses before any data work, and you’ll find it alongside every other download on the MMM templates page.
Keep the brief somewhere you’ll revisit it. Later in a project the questions get more precise, the evidence register fills up and the constraints change, and it’s much easier to see whether the model is still doing the job you gave it when that job is written down.
