A marketing measurement framework is an agreed plan for which evidence answers which marketing decision: marketing mix modeling for budget allocation, experiments for causal checks, attribution for in-flight optimization, and brand tracking for long-term demand, each with an owner and a refresh cadence. Its main job is to stop a team from asking one method to answer every question, and to decide in advance what happens when two methods disagree.
In the Vexpower marketing mix modeling course, the invented brand Northstar has a platform dashboard, customer research, web analytics and an MMM proposal, and the team wants one number that settles every disagreement. That expectation is the first thing the course asks you to change, and it’s the first thing I’d change in most real marketing teams too. I’ve built marketing mix models for companies including Monzo, Booking.com and The AA, and in my experience a model changes decisions when it has a clear job alongside the attribution dashboard and the experiments, and a clear rule for what to do when the numbers don’t line up, rather than when it tries to replace every other source of evidence.
I’ll show you how to give each method a job, how to pick the right evidence for the decision in front of you, and how to run the whole thing as a learning loop rather than a reporting exercise.
What a marketing measurement framework is (and isn’t)
You’ll see the same idea sold under several names: “unified marketing measurement”, “triangulation”, a “measurement stack” or a “measurement plan”. The vendor versions differ in detail, but they share one core idea, which is that no single method is good enough on its own, so you combine methods whose weaknesses don’t overlap. An MMM can see TV and can’t see creative; attribution can see creative and can’t see TV; an experiment can see causation, but only for one channel, in one period, in one place.
What a framework is not is a formula that blends three numbers into one. Averaging an attributed ROAS of 4.0 with a modeled ROI of 1.5 to get 2.75 produces a number that no method actually estimated and that nobody can defend. The framework’s value is in the definitions: knowing that the 4.0 is attributed revenue over a seven-day click window and the 1.5 is incremental revenue against a zero-spend baseline over two years tells you they answer different questions, and that the real work is deciding which question your decision depends on.
It also isn’t a one-off project. A framework that gets written up in a slide deck and never revisited goes stale within a couple of quarters, because channels change, tracking changes and the questions the business cares about change. The version that works is closer to an operating rhythm, with a calendar, named owners and a register of what each piece of evidence has already told you.
Give each source of evidence a job
The simplest way to start is to write down what each source you already have is good for, and what it can’t tell you. This is the table from the course lesson, and it makes a good first page for any measurement plan:
| Evidence | A useful role | A limitation to investigate |
|---|---|---|
| Platform reporting | Delivery, optimization and attributed outcomes | Attribution rules don’t establish incrementality |
| Analytics | Recorded customer journeys and site behavior | Missing devices, consent gaps 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 |
| Brand tracking | Awareness and consideration over time | Survey noise; not revenue until linked to sales |
Read across each row and the jobs separate naturally. Marketing mix modeling is the only method that covers every channel, including offline ones, and models diminishing returns, so it gets budget allocation and scenario planning. Experiments are the only method that establishes causation by design, so they get the job of checking the model and settling disputes (incrementality testing covers the designs). Attribution is the only method fast and granular enough to choose between creatives on a Tuesday afternoon, so it gets in-flight optimization, with the caveat that its totals are not incremental (the attribution guide explains why). Brand tracking gets the long-term questions that sales data takes too long to answer.
Which method answers my question?
The table above is general. In practice the right evidence depends on the decision, the budget, the kind of channel and how fast you need an answer. A team spending $30,000 a month on search and social shouldn’t build the same measurement program as one spending $3 million a month across TV, audio and digital, and a decision you need to make this week can’t wait for a geo test. The selector below encodes the rules of thumb I use; change the inputs and see how the recommended stack shifts.
A few patterns are worth calling out. Allocation and scaling questions almost always lead with an MMM, because they are questions about marginal returns across channels, and only a response curve gives you those (the marginal ROAS guide explains the difference from average ROAS). “Is this channel working at all?” questions lead with experiments, because the honest answer to “would these sales have happened anyway?” needs a control group. Budget size matters more than most frameworks admit: Below roughly $50,000 a month, the variation in spend is often too small relative to the noise in sales for a full MMM to say much, and a few well-designed go-dark tests will teach you more for less money. And when the channels are mostly offline, attribution drops out of the stack almost entirely.
Start from the decision, not from the data
The most common mistake I see is building measurement around the data a team happens to have, rather than around the decisions it actually makes. A company with beautiful user-level tracking will lean on attribution for everything, including questions attribution can’t answer, and a company that has just bought an MMM will try to use it to choose between ad creatives. Both are using the evidence they have rather than the evidence the decision needs.
The course’s approach is to start with a business decision and work backwards: Record what is already known, which uncertainty could change the action, and what evidence could reduce that uncertainty. If the decision is whether to move 20 percent of the search budget into video, the uncertainty that matters is whether video’s incremental return at the higher spend level beats search’s at the lower one. That points to the MMM’s response curves for both channels, and if the model’s interval for video is wide, a geo test at the higher spend level. It doesn’t point to attribution at all, because attribution can’t tell you what the next dollar is worth. The measurement plan template walks through writing this down for your own decisions, from the business question to the outcome KPI.
It’s also worth being honest about which decisions are big enough to measure properly. A decision that moves $5,000 doesn’t justify a $50,000 geo test, and a framework that treats every question with the same rigour will either be ignored or bankrupt the analytics team. I’d rather a team runs two or three excellent experiments a year on the questions with the most money riding on them than a dozen underpowered ones (the statistical power guide shows how to check before you start).
Cadence and owners for a marketing measurement program
Every method in the framework needs a rhythm and a name next to it, or it quietly stops happening. The cadence should match how often the decision it serves is actually made, not how often the data could in theory be refreshed. Budgets are usually reallocated quarterly, so a quarterly MMM refresh is plenty for most teams; refreshing monthly mostly produces small wobbles in the estimates that tempt people into moving money for no good reason.
A typical rhythm looks like this:
| Evidence | Cadence | Owner | Feeds |
|---|---|---|---|
| Attribution and platform reporting | Daily to weekly | Channel owners | Bids, creative, campaign pausing |
| MMM refresh | Quarterly | Analytics (with finance) | Channel budgets, scenarios, multipliers for attribution |
| Geo experiments | Two to four a year | Analytics, design agreed with the channel owner | MMM calibration, keep-or-cut decisions |
| Platform lift studies | Per major campaign | Channel owner, reviewed by analytics | Campaign-type decisions, MMM priors |
| Brand tracking | Continuous, read quarterly | Brand or insight team | Long-term investment case |
| Evidence register review | Quarterly | Analytics lead | Next quarter’s test plan |
The owner column matters more than it looks. When the person who runs paid social also designs and reads the paid social lift test, the results tend to be generous, not through dishonesty but because every judgment call (which weeks to exclude, which regions to match) gets made by someone who wants the channel to look good. Having analytics own the design and the read, with the channel owner agreeing the design before the test starts, removes most of that.
When methods disagree, investigate before you average
Disagreement between methods is normal, and it’s useful, because it points straight at an assumption worth checking. However, most disagreements turn out to be definitional rather than real. A campaign experiment may estimate the effect of extra spend in selected regions, an MMM may report a channel’s average historical contribution against a zero-spend baseline, and a platform may report conversions credited within a seven-day window. Those numbers need definitions before they need a reconciliation formula. Google’s Meridian documentation makes the same point about calibration: An experiment that measures lift from a reduced-spend baseline isn’t estimating the same quantity as an MMM whose counterfactual is zero spend (Meridian docs).
When two methods disagree on a channel, I work through the same questions in order:
The fifth question catches more teams than you’d expect. If a geo test was used to set the MMM’s prior for a channel, and the model then agrees with the test, that agreement isn’t two independent pieces of evidence; it’s one piece of evidence counted twice. This is exactly why the course recommends an evidence register that records, for every result, whether it has already informed another model.
Once the definitions are straight, the gaps that remain can be turned into something usable day to day. Divide each channel’s incremental conversions (from the MMM or a test) by its attributed conversions over the same period, and you get a multiplier that translates the attribution dashboard into incremental terms. The disagreement log in the download flags any channel where that ratio falls outside a band you choose, which is a reasonable shortlist for next quarter’s experiments.
Run it as a learning loop
The difference between a framework that improves decisions and one that just produces reports is whether each new piece of evidence changes the next question. After a geo test, the register should record what it measured, against what comparison, with what uncertainty, and what it changed: Did it shift the budget? Did it go into the MMM as a calibration prior (calibrating an MMM with experiments covers how)? What is now the biggest remaining uncertainty? That last question becomes the brief for the next test.
Over a year or two, this loop does something no single method can. The MMM gets better because its weakest channel estimates have been checked by experiments, the attribution dashboard becomes more trustworthy because its multipliers are grounded in incremental evidence, and the team gets better at asking questions that measurement can actually answer. Gordon and colleagues’ 2019 study of Facebook lift experiments, which found that observational methods often failed to reproduce randomized results, is a useful reminder of why experiments sit at the top of that loop: They are the check on everything else.
Common ways measurement frameworks fail
The first failure is crowning a winner. A team that decides the MMM is “the truth” stops running experiments, the model’s priors drift away from reality, and nobody notices until a big reallocation fails. The same happens in reverse with teams that trust only experiments and end up making cross-channel decisions from a handful of tests that were never designed to compare channels.
The second is reporting without deciding. Plenty of teams have an MMM, a lift study program and an attribution tool, and still allocate budget by last year’s numbers plus 10 percent, because nobody connected the evidence to the planning calendar. If the MMM refresh lands two weeks after the budget is locked, it doesn’t matter how good it is. Put the decision dates on the cadence calendar first and schedule the evidence to arrive before them, and when it arrives, present it as a decision with a range rather than a wall of charts.
The third is ignoring the long term. Attribution windows are measured in days, experiments in weeks and most MMMs in a couple of years of weekly data, so brand effects that build over years are easy to miss in all three. That’s the case for keeping a brand tracker in the framework even when it can’t produce an ROI, and for reading the work of Les Binet and Peter Field on brand and activation before deciding that everything worth doing pays back inside a quarter. A good marketing measurement framework can be simple, as long as every method has a job it’s actually good at and the team treats each disagreement as the next thing to test.
