Five things brands get wrong with MMM

If you are a founder spending serious money on marketing and your measurement is supposed to be helping you make better decisions, this piece is for you. Marketing Mix Modelling (MMM) is having a moment. Half of marketers worldwide use some form of it, and 67% of marketing leaders plan to increase their investment over the next two years (Gartner, 2024). But adoption is running ahead of execution. 75% of marketers say their current measurement is not delivering the speed, accuracy or trust they need (IAB/BWG, State of Data 2026).

So what is going wrong? After fifteen years inside this industry, sitting on the brand side commissioning measurement at Heineken and Diageo, then on the platform side at Meta building data systems, and now running Hanya, I see the same five mistakes over and over. Here is what they look like, and what to do about each one.

Mistake 01

The results arrive too late to matter

A typical agency-led MMM project takes between three and six months to deliver. In larger enterprise rollouts, it can stretch to a full year. By the time the deck lands, the budget has been signed off, the agency briefed and the planning window closed. The insight is now historical. The decision has already been made.

The reason it takes so long is rarely the modelling. The VP of Measurement at Funnel put it bluntly: the model is only 20% of the work. The other 80% is consolidating and transforming data. Circana's research backs this up. Data preparation eats 30% of project time for organisations with strong data infrastructure, and 70% or more for everyone else.

That 70% figure is where most brands sit. Spend data scattered across platforms. Agencies reporting on different cycles. Naming conventions that change quarterly. Someone always needing one more week to reconcile their numbers. The modelling itself is fast. Meta's Robyn can test thousands of model combinations per minute. Google's Meridian launched a no-code Scenario Planner in early 2026. The bottleneck sits upstream of the maths.

There is a real cost to slow refreshes. OptiMine's research found that with an annual cadence, models can be off by 50 to 100%. Their conclusion: "in some cases those models can do more harm than good." A tool that misleads is worse than no tool at all.

Mistake 02

The output is built for the people who built the model, not the people who need to act on it

Only 20% of analytical insights deliver business outcomes (Gartner). Only 3% of practitioners at the ANA Media Conference said their measurement solution does everything they need (AdExchanger, 2025). And only 21% of marketing leaders say they receive useful insights in real time, down from 26% two years ago (NielsenIQ). The trend is moving the wrong way.

The reason most MMM results get filed rather than used is structural. Traditional engagements run on annual or semi-annual refresh cycles. One study, once or twice a year, producing a single set of recommendations. But marketing teams do not make decisions once a year. They make them monthly, sometimes weekly.

The output format compounds the problem. Most MMM deliverables are built to be defensible rather than usable. A 40-page methodology deck looks impressive in a steering committee, but the person trying to decide whether to push more spend into Q3 paid social needs something else. They need to know what happens if they shift 15% of TV budget into social. They need a number, a confidence range, and a recommendation they can defend on Monday morning.

When the question lands of "what would happen if we cut marketing by 10% next quarter?", the answer from a legacy MMM is usually "let me get back to you in a month." That is not measurement. That is homework you set yourself.

Mistake 03

Nobody can explain the model to the people writing the cheques

Only 49% of senior marketing and finance leaders said they could clearly explain their measurement approach to the board (Haus Decision Confidence Index, March 2026). 74% have abandoned or scaled back a marketing initiative because they lacked confidence in how to measure its impact. McKinsey's 2025 CMO research found that 70% of CEOs measure marketing on year-over-year revenue growth and margin, while only 35% of CMOs track those as their top metrics. The two groups are not even speaking the same language.

For founders, this gap is your gap. You are signing off the spend, you are answering to investors or the board, and you are the one whose head is on the block if marketing money disappears without a measurable return on growth. If your measurement output is full of Variance Inflation Factors, posterior predictive checks and Adstock decay rates, you cannot use it. Worse, you cannot defend it to anyone else.

Pressure on marketing leaders has been climbing fast. The CMO Survey (Spring 2025) puts pressure from finance up 21%, from CEOs up 20%, and from boards up 52% since 2023. Yet the measurement frameworks they rely on, MMM included, produce outputs designed for data scientists. R-squared values do not pay the rent.

This is one of the reasons measurement bodies like the Marketing Accountability Standards Board (MASB) have spent the last decade pushing for standardised, auditable definitions of marketing return that finance can trust. Most MMM engagements still ignore that work entirely.

Mistake 04

You have no way to tell whether the model is any good

This is the most technically important problem on this list, and the one most brands are least equipped to evaluate. Most receive a model with a high R-squared value, the classic measure of how well a model fits historical data. R-squared of 0.92 sounds reassuring. But as Recast's research puts it: you can have a terrible model that has a really high R-squared, and a great model that has a really low one.

The published industry benchmarks are worth knowing. Meta's Robyn flags R-squared below 0.8 as not ideal and above 0.9 as ideal. Stella's validation guide flags anything above 0.95 as suspicious for overfitting. For Mean Absolute Percentage Error (MAPE), 5 to 15% is considered reasonable for well-specified models (Funnel.io). Out-of-sample accuracy, where the model predicts a period it was not trained on, should sit within 10 to 20%.

Here is the statistic that matters most. Uncalibrated MMM models show a 25% average difference from ground truth ROAS, according to Analytic Edge research commissioned around Robyn (2024). The typical model, without experimental validation, is out by a quarter on its core output. You are making budget decisions based on numbers that are systematically 25% wrong, with no idea in which direction.

The strongest validation is calibration against real-world experiments. A geo holdout test, where you pause advertising in a subset of regions and measure the difference, lets you compare the model's prediction against what happened in reality. Meta's HBR-published research describes this as the emerging gold standard for MMM validation. The IAB's 2025 best practices guide now recommends requiring at least two supporting signals for any material budget decision, an approach Nielsen has long called "triangulated measurement".

Mistake 05

The findings are politically inconvenient, so nothing changes

This is the most common reason MMM findings get ignored, and the least often discussed.

The numbers are stark. 91% of organisations cite cultural challenges and change management, not technology, as the principal barrier to becoming data-driven (NewVantage Partners, 2025). That figure has been stable for five years running. Only 24% of organisations describe themselves as data-driven (Forbes/NewVantage, 2023).

A Gartner survey of 377 marketing analytics professionals found that one third of decision-makers cherry-pick data that supports a decision they have already made. 26% do not review the data at all. 24% go with their gut.

The sunk cost research explains why. Arkes and Blumer's foundational study found 85% of decision-makers chose to continue a failing project when prior investment was mentioned, against 10% when it was not. Loss aversion is approximately twice as powerful as the equivalent gain (Kahneman and Tversky). Confirmation bias has the strongest negative impact on marketing decision quality of any cognitive bias studied. The person who championed the underperforming channel is psychologically the least likely to accept evidence that it is underperforming.

The Uber example is a useful counter-point. Their analytics team suspected Meta rider-acquisition ads were not driving incremental users. An MMM flagged it. A three-month incrementality test (Meta ads turned off, no drop in riders) confirmed it. They reallocated $35m a year into higher-ROI channels. The reason it worked was that the test was designed before the result was known, the rules for what counted as success were agreed upfront, and the experiment was run by a team without a stake in the outcome.

So when does MMM work?

The brands that get real value from MMM tend to share a few things. Their data pipeline is automated, so results land while decisions are still live. Their output is built for the people deciding budgets rather than the people building models. And the organisation has agreed in advance what it would change based on the findings.

MMM is most useful when you spend meaningfully across multiple channels and you do not have a clean read on what is driving growth. As a rough guide, brands spending under £1m a year on media tend to find the cost of measurement outweighs the insight. Above that, the maths stacks up fast. A typical engagement adds 20%+ to marketing return on investment in the first 12 weeks, just by reallocating budget into the channels that are pulling their weight. Over a three-year window the compound impact reaches seven figures of incremental revenue for most brands in the £3-10m spend bracket. The same logic underpins our long-form essay Beyond Attribution, which explains why attribution alone caps your growth and what MMM adds on top.

If you want to know what MMM could unlock for your business specifically, the calculator below gives you a benchmarked estimate in 60 seconds, no email required.

Written by Matt Rix · Last updated 2 May 2026