Every ad platform grades its own homework. That is the first thing worth understanding about modern marketing measurement.
Meta says your campaign worked. Google says your campaign worked. The CFO asks which half of the budget actually made money, and nobody in the room can answer with confidence. Not because the data is missing but because the data everyone is looking at was produced by the very systems with the most to gain from looking good.
High-growth companies have figured this out. The gap between them and everyone else rarely comes down to campaigns or creativity. It comes down to whether the marketing team can walk into a budget meeting with numbers that nobody in the room can argue with.
Why No Single Measurement Method Works Anymore
Last-click attribution hands all the credit for a conversion to whichever channel the customer touched last. In practice, that usually means paid search collects the reward for purchases that awareness campaigns, email sequences, and organic content spent weeks building toward. Budgets follow the credit. Channels that do the upstream work get defunded because there is no mechanism to see what they actually contributed.
It was always a flawed model. What changed is that the data layer underneath it has thinned considerably. Usercentrics and Sapio Research’s State of Digital Trust 2025 report, drawing on 10,000 consumers across Europe and the US, found that 38% of Americans are accepting cookies less often than three years ago.
The signal that last-click relied on is deteriorating, and the measurement approaches built on top of it are deteriorating with it.
The problem with any single measurement method is that each one sees only part of the picture. Attribution tells you what happened. Platform dashboards tell you what platforms want you to believe happened. What high-growth companies have figured out is that you need several methods in parallel, each catching what the others miss.
Marketing Mix Modelling as Strategic Guidance
CFOs do not care whether YouTube has a better click-through rate than Meta. They care whether shifting budget from one channel to another will grow revenue, and they want to see the model that says so. That is the question MMM exists to answer.
The method has been around since the 1960s, which tells you something: it predates digital advertising entirely, and it is still the most reliable cross-channel view available.
The modern version uses Bayesian statistics and incrementality calibration rather than manual regression, but the underlying logic is unchanged. Analyse historical data across all your marketing activity, account for seasonality and external factors, and surface which inputs are actually driving outcomes.
Here is what that looks like when it works. One Google-backed MMM study showed that YouTube’s true contribution to business outcomes extended far beyond what last-click attribution gave it credit for. Meta was collecting the credit at the point of sign-up because it served the retargeting ad close to conversion. YouTube was doing the trust-building work earlier in the journey that made someone a real prospect in the first place.
The company reallocated the budget. The total spend stayed the same. The mix changed. This is not a case study from a vendor deck; it is a representative example of what MMM consistently surfaces when teams look honestly at where credit has been flowing versus where value is actually being created.
If you are still allocating next quarter’s budget using platform ROAS, you are effectively letting Meta and Google negotiate against each other using your own money.
Nearly half of US marketers now plan to invest more in MMM over the next year, and the majority named it the single most reliable measurement methodology available, per a July 2025 survey by EMARKETER and TransUnion.
That shift is not happening because MMM is new. It is happening because the alternatives have stopped working.
MMM does have a constraint worth naming. It works at an aggregate level. It can tell you that paid social drove a certain percentage of revenue over a period. It cannot tell you which specific campaign or creative did the work within that channel. For that, you need the next layer.
Incrementality Testing as the Causal Engine
Every platform takes credit for every conversion it can see. That sentence explains why platform ROAS is one of the most systematically abused metrics in digital marketing.
Incrementality testing is the antidote. The methodology is deliberately simple: withhold ads from a defined control group, expose your target group as normal, then measure the difference in outcomes. The gap between the two groups is the lift your marketing actually caused. Not the lift your dashboards reported. The lift that would disappear if you stopped spending.
Stella’s 2025 DTC benchmarks, drawn from 225 geo-based tests, put the median incremental ROAS across paid channels at 2.31x. Branded Google Search came in at 0.70x. Below the breakeven line. Brands were paying to intercept customers who were already searching their name and would have found them anyway. The platform reported strong ROAS because it was counting the sale. The brand was essentially buying its own organic demand back from Google.
The brands spending most on branded search tend to be the ones most surprised by what incrementality testing finds. The platform confidence is highest exactly where the true lift is lowest, because the audience is already decided and the ad is just the last thing they clicked.
Over half of US brand and agency marketers are now using incrementality testing, with 36.2% planning to increase investment in it over the next 12 months, according to the same EMARKETER and TransUnion survey. What is driving that adoption is not enthusiasm for methodology. It is the gradual realisation that privacy restrictions have made the alternative less viable, and that the numbers platforms report are increasingly hard to reconcile with business outcomes.
Used well, incrementality testing does not replace MMM. It calibrates it. An MMM running on historical correlations alone drifts over time. Periodic incrementality tests anchor it to actual causal evidence and keep the model honest.
Platform Data Still Matters, But Only for Optimisation
Platform dashboards are useful. Just not for the decisions most people use them for.
Inside a single channel, real-time data tells you things that matter: which creative is outperforming this week, which ad sets are approaching budget exhaustion and where to pull spend before it wastes. These decisions happen at a cadence that neither MMM nor incrementality testing can match, and platform data is the right tool for them.
The mistake is using in-platform data to make cross-channel budget decisions. Asking Google whether Google Ads should get a bigger share of next quarter’s budget is not a measurement exercise. It is outsourcing the allocation of your money to the company that benefits from spending more of it.
The cleaner way to think about it: platform data tells you how to optimise what you are already doing. MMM and incrementality testing tell you whether what you are already doing is worth optimising at all. Those are different questions and they need different answers.
The Pioneer, Settler and Town Planner Measurement Model
Most organisations that try to build better measurement capability eventually hit the same wall: the people being asked to run incrementality tests are the same people maintaining the weekly reporting dashboard, which is the same team trying to evaluate new MMM vendors. Everything competes with everything else, and the experimental work always loses to the urgent operational work.
We see this regularly when working with brands that have invested in measurement tools but are not getting value from them. The infrastructure exists. The capability to actually use it does not exist, because everyone with the relevant skills is buried in reporting.
Researcher Simon Wardley mapped this through a framework describing three fundamentally different kinds of organisational work. Pioneers operate in uncertainty, running experiments whose outcome they cannot predict. Settlers take successful experiments and turn them into reliable, repeatable systems. Town Planners take those systems and industrialise them: standardised, efficient, scalable enough that the whole organisation depends on them.
In a marketing measurement context, pioneers are the analysts running geo-holdout tests on new channels or exploring whether causal inference models can improve MMM accuracy. Settlers are the people who identify what worked and build the process around it. Town Planners own the dashboards, the reporting cadences, the budget tracking systems that teams actually use week to week.
The trouble most organisations run into is that they only have town planners. Execution is efficient. Experimentation never happens. Or they have pioneers but no settlers, so successful experiments never become standard practice and the knowledge walks out the door when someone leaves.
Getting this right does not require a restructure. It requires honestly naming which kind of work each person is being asked to do, and whether the incentives around them are set up to support it.
How High-Growth Companies Allocate Measurement Resources
Seven in eight US marketers plan to invest more in at least one measurement methodology over the next year, per EMARKETER’s measurement trends analysis. More investment without better structure tends to produce more tools, more data, and the same number of clear answers.
The allocation pattern that actually works keeps the majority of measurement budget on ongoing infrastructure: data collection, the MMM platform, standard reporting. A smaller portion goes toward periodic incrementality tests, timed to sync with major budget decisions rather than run on an arbitrary calendar. A small slice funds exploratory work.
One practical development worth knowing: Google reduced the minimum budget for incrementality experiments from approximately $100,000 to $5,000 by adopting Bayesian statistical methods that require less data to reach meaningful conclusions. Mid-market brands that previously could not afford rigorous testing now can.
The pattern that consistently separates high-growth companies from everyone else is not budget size. It is the directness of the line between measurement output and budget decision. Teams where the MMM report sits unread in a shared folder until after the quarterly budget has been set are spending on measurement theatre.
Seven Steps to Evolve Your Measurement System
The path from platform dependence to a proper multi-method system is not complicated, but it is sequential. Skipping steps is how teams end up with sophisticated tools sitting on top of unreliable data.
Start with your data, not your tools. MMM and incrementality testing both require at least 18 months of clean, consistent historical data: media spend by channel, revenue at a matching level of granularity, and enough documentation of your seasonality patterns that a model can separate genuine marketing effects from the annual spike that happens every December regardless of what you ran.
Run one incrementality test before you do anything else. Pick your highest-spend channel. Run a geo-holdout or conversion lift study. The goal is not a complete picture of your marketing mix. It is one credible data point that either validates your current allocation or forces a harder conversation.
Build the MMM on top of that foundation. An MMM calibrated against real incrementality data is considerably more reliable than one built purely on historical correlations. The calibration is what separates a model you can defend to a CFO from one that is mathematically impressive and strategically useless.
Connect the output to a budget decision, in advance. Before the MMM is built, decide who will review the findings, when, and what authority those findings carry over budget allocation. If you cannot answer those questions before the model runs, the model will not change anything.
Treat the system as infrastructure, not a project. Measurement capability compounds. A team that reviews MMM outputs quarterly and feeds incrementality test results back into the model improves its accuracy over time in ways that a team running annual analyses simply cannot match.
The two steps most teams skip are the first and the last. Data foundation work is unglamorous and generates no immediate output. And treating measurement as permanent infrastructure means accepting that it requires ongoing resources rather than a one-time investment. Both are worth getting right before the rest.
Key Takeaways
- Every ad platform grades its own homework. Building a measurement system that can challenge platform-reported numbers is the foundational capability high-growth companies have that most others do not.
- Attribution tells you what happened. Incrementality testing tells you what caused you to happen. MMM tells you which channels drove business outcomes across the full mix. All three answer different questions. The mistake is using any one of them for a job it was not built to do.
- Platform ROAS is one of the most abused metrics in digital marketing. Branded search campaigns frequently show strong platform-reported ROAS while delivering incremental returns below breakeven, because they are capturing demand that would have arrived anyway.
- Nearly half of US marketers now plan to invest more in MMM. That is not a trend. It is the industry admitting that attribution has stopped answering the questions finance teams actually ask.
- The Pioneer, Settler, Town Planner model is useful because it names a real organisational failure: most measurement teams are either all execution and no experimentation, or all experimentation and no operational follow-through.
- Measurement investment matters less than measurement integration. Data that does not change decisions is not measurement. It is record-keeping.
FAQs
Our platform numbers look healthy. Do we actually need incrementality testing?
Probably more than teams whose numbers look bad. Healthy platform numbers are often the sign of a measurement problem, not a marketing one. Branded search consistently reports strong ROAS because it captures customers who were already going to buy. Retargeting looks efficient because it reaches people near conversion who were already in the funnel. The channels that look best on platform dashboards are often the ones with the lowest true incrementality. That is not a coincidence. It is how attribution works.
We already have an MMM. Is incrementality testing still necessary?
Yes, and this is the question worth getting right. An MMM built purely on historical correlations will drift over time as the marketing mix changes. Incrementality tests give you causal anchors to keep the model honest. A 2025 Sellforte study found that an MMM-estimated ROI of 3.91x was validated by an incrementality test returning 4.00x, precisely because the model had been calibrated against real experimental data. Without that calibration, the gap can be considerably wider.
How much historical data does MMM actually need?
Two years is the comfortable threshold. Eighteen months works if your seasonality is relatively predictable. Below that, the model lacks enough variation in its inputs to separate genuine marketing effects from background noise. Data quality matters more than volume: weekly spend by channel matched to weekly revenue, consistently structured, is what the model actually needs. Teams that skip this step and go straight to modelling end up with outputs that are mathematically sophisticated and strategically unreliable.
What happens if the MMM says cut a channel the team is emotionally attached to?
This is the real test of whether measurement is actually integrated into decision-making or just used for post-rationalisation. The honest answer is that this happens regularly, and it is where many measurement programmes stall. The answer is not to run a better model. It is to agree before the model runs what authority the findings will carry. If that conversation has not happened, the output becomes a negotiating document rather than a decision input.
Is this level of measurement only realistic for large brands?
The cost barrier has dropped significantly. Google reduced the minimum budget for incrementality experiments from approximately $100,000 to $5,000 using Bayesian statistical methods. Accessible MMM platforms now exist for mid-market budgets. The constraint for most brands is not cost; it is having someone who can own the process end-to-end, from test design through to connecting the findings to an actual budget decision.
Conclusion
The companies that win on measurement are not the ones with the most data. They are the ones willing to let better measurement prove that their favourite channel is not actually working.
Most marketing teams respond to measurement uncertainty by buying more dashboards. The best ones use it as the reason to ask harder questions. There is a version of every marketing budget conversation where the MMM says one thing and the platform dashboard says another, and someone in the room has to decide which one they trust. High-growth companies have already done the work to know the answer before that moment arrives.
At Ellipsis Digital, that is where the measurement conversation usually starts. Not with which tools to use, but with what decisions those tools need to serve. If you are building that capability or rebuilding one that has stopped working, that is a conversation worth having before the next budget cycle.
