Blog
How targeted advertising cuts wasted ad spend across display, online video and CTV
Learn more
Media mix models have helped advertisers make sense of marketing effectiveness since the 1960s. Long before dashboards and real-time optimisation, they offered a strategic promise: step back from channel-level noise and understand what actually drives growth.
Ultimately MMMs aim to answer the questions senior marketers care about most: What was my return on ad spend on TV last year? What would happen if I shifted budget next quarter? How should I allocate spend to maximise sales?
According to Boston Consulting Group research in EMEA, 76% of companies already rely on MMM for cross-channel budget allocation. In an environment defined by fragmentation, rising media costs and pressure to prove impact, a single, big-picture view is attractive.
The risk is not that MMM is wrong. It is that it feels hands-free.
As platforms lean further into automated, probabilistic optimisation, it becomes tempting to hand over complexity and accept model outputs as truth. But when MMMs are treated as decision engines rather than decision support, model error scales quickly. If assumptions are wrong, entire budgets can drift off course before anyone notices.
Unlike attribution approaches that attempt to reconstruct individual customer journeys, MMMs work at an aggregated level. They rely on historical time series data to understand how changes in marketing inputs relate to changes in outcomes over time.
At their core, MMMs are regression models. They estimate statistical relationships between outcomes, such as sales, and inputs, such as media spend, promotions, or pricing. From those relationships, they infer contribution and forecast future scenarios.
In principle this makes MMMs great at capturing interaction effects between channels, such as how connected TV supports in-store conversion, or how brand investment lifts downstream performance. They can also incorporate non-marketing factors like seasonality, economic conditions, competitive activity, weather, and major external events.
That breadth is exactly what makes MMM valuable, providing context that narrower methods miss. It is also what makes MMM fragile when conditions change.
MMMs and probabilistic AI systems are statistical approximations of reality, not direct measurements of incrementality. They infer impact rather than observe it.
Two challenges sit at the heart of this:
As a result, MMMs infer relationships from correlations that may be shaped by forces outside the model. When teams outsource judgement entirely to these outputs, they can end up accepting neat, persuasive stories about channel performance that are difficult to validate.
As AI-driven tools become more confident in their explanations, this automation bias grows stronger. Outputs feel authoritative and questioning them feels unnecessary. This is when riding hands-free becomes genuinely risky.
There is no way to guarantee that an MMM is perfectly specified. What advertisers can do is reduce model risk and increase confidence in the decisions that flow from it.
The most effective way to do that is through controlled experimentation. Experiments answer a different question to MMM. Rather than asking what likely happened in the past, they test what happens when an advertiser deliberately does something different.
By randomly splitting a population into a test group and a control group, advertisers can isolate the impact of a specific action, such as increasing spend in a channel or suppressing exposure to a particular audience. Done well, experiments provide clean causal signals.
In practice, however, experimentation is hard to scale. Clean test and control splits across fragmented media ecosystems are technically complex. Holding out users creates opportunity cost that performance-focused teams are often uncomfortable with. Many media effects are small or noisy, meaning tests must be large or long to reach statistical confidence. And individual experiments tend to answer narrow questions at specific spend levels, while MMM requires broad response curves across channels and time.
This is where deterministic data changes the equation.
Deterministic data, such as identity graphs, logged exposures, and transaction-level outcomes at the person or household level, creates a far clearer map from eligibility to exposure to outcome.
It allows advertisers to define test and control groups with greater precision, reducing contamination and leakage between cohorts. It supports more granular experiment designs by audience, publisher, or geography, without losing the ability to link outcomes reliably.
Just as importantly, deterministic outcome data from CRM, loyalty programmes, and offline sales can be tied back to exposure with far less noise than panel-based or modelled approaches. That increased fidelity boosts statistical power, making smaller and shorter experiments viable.
The result is not experimentation as a special project, but experimentation as a repeatable operating mode.
Traditional MMMs often smoothed performance over long time horizons, blurring the impact of recent changes. What modern planning requires is continuous, multi-channel calibration.
Deterministic experimentation allows advertisers to run targeted tests where uncertainty or risk is highest, such as contested budget decisions or emerging channels. The causal results from those tests can then be fed back into the MMM as anchors, strengthening response curves where real-world evidence exists and allowing the model to extrapolate more confidently elsewhere.
Over time, and with the right data foundation, this process can be automated. Experiments become embedded in normal campaign delivery. MMM becomes a living system that is continuously refreshed, challenged, and corrected, rather than a static report produced once or twice a year.
That is what keeping a hand on the handlebars looks like.
Before building an MMM, advertisers need to lay the groundwork.
Media mix modelling can be a powerful guide. But like any ride worth taking, it is safest and most effective when someone keeps at least one hand firmly on the wheel.
