How to use multiple MMMs to make better paid media decisions

After some initial setup and tinkering, any marketing mix model (MMM) can produce a convincing chart with channel decomposition, response curves, a clean R², and a recommended reallocation. But the chart won’t show how much the data shaped the result versus how much the model’s built-in assumptions did.

Adstock and decay windows determine how long a channel’s effect lingers. Saturation curves affect how quickly returns diminish, driving every reallocation the model recommends. Priors and regularization (Bayesian or ridge) encode beliefs about plausible effect sizes, while seasonality and control variables influence how much lift the model attributes to the calendar versus the channel.

Change these factors, and the same history can produce a different story.

As a result, a single model is a first opinion, not a verdict. Running multiple models against the same inputs reveals how different assumptions change the recommendation before you make a seven-figure budget decision.

My process uses multiple MMMs to surface uncertainty, validate recommendations, and make paid media decisions more reliable.

Run a multi-model MMM comparison

Before comparing MMMs, it helps to separate model validation from experimental validation.

Incrementality tests provide a stronger causal check. A geographic lift, holdout, or on/off test directly measures whether a channel caused additional outcomes. But it covers one channel at a time, costs more, and requires planned variation.

MMM works at a broader level. It infers causal contributions across every channel at once, including those you can’t easily test, and it reruns cheaply to prove results within days or weeks. But it depends entirely on its own assumptions.

Ideally, the two complement each other. Models generate and rank hypotheses, experiments confirm which ones matter, and confirmed results feed back into the models as priors.

Because you can’t test every channel every quarter, second, and third models become the next line of defense against any one model’s blind spots. Skip this step, and a reallocation from a single model inherits that model’s blind spots. A misallocation can reach six or seven figures before anyone notices.

Top tools for running MMM

My solution is to run MMM through three separate tools with different modeling approaches and assumptions:

• Robyn (Meta, open source) uses ridge regression with an evolutionary search across hyperparameters. It’s fast to run and accessible to marketing teams without a Bayesian background, which makes it a strong initial baseline.
• Meridian (Google, open source) is Bayesian and geographically hierarchical, with reach, frequency, and upper-funnel effects within its scope. Regional spend variation adds statistical signals a national-only model doesn’t have, which makes it well suited to geographical data and brand channels.
• PyMC-Marketing (open source, Python, built on PyMC) is fully Bayesian with user-defined priors, model structure, and indirect-effect paths between channels. It gives you full control over assumptions, but it needs someone who can defend those assumptions.

Robyn relies on R, while Meridian and PyMC-Marketing are Python-native. If your team is R-first, I recommend keeping Robyn as the fast in-house baseline and either building light Python wrappers for the other two or budgeting time to work across both ecosystems. The data prep is shared across all three.

Dig deeper: Not all MMM tools are equal: Meridian, Robyn, Orbit, and Prophet explained

How to run a multi-model comparison

Use this workflow to run and compare multiple MMM models, from inputs through testing:

• Inputs: Use identical spend, outcome, and control variables for all three tools. While the initial model is expensive, every additional model reuses the same prepared inputs. As a result, the marginal cost of models two and three is small.
• Models: Run all three with defaults first and tuning second. Resist the urge to hand-tune model one before model two has even run. You want to see where the defaults disagree before you start explaining the disagreement away.
• Comparison: Compare channel decomposition and response curves across models. Fit statistics can be a distraction here. For instance, a high R² means the model fits history, not that its causal story is correct. Every MMM fits reasonably well, but they still disagree with each other. Budget recommendations depend more on where each model sees diminishing returns on the saturation curve and how consistently channels rank across models than on small differences in revenue share.
• Triage and testing: The comparison turns into a decision.

Analyze MMM output where the models disagree

Convergent results support action, while divergent results require investigation. A comparison makes any uncertainty visible before it turns into a budget mistake.

When the models converge, the finding has held up across three different modeling approaches and their underlying assumptions. At this point, you can defend reallocation, close the debate, and use the result as a prior for the next model refresh.

When models diverge, no winner is declared. The root cause is almost always a confound, collinearity problem, or data gap. Confidence gets downgraded, and the next experiment focuses on the biggest disparity.

Why MMM models often diverge

Here are recurring reasons the models disagree:

• Channel collinearity: Two channels scale together, so each model divides the credit between them differently. You can’t identify the split from observation alone. A holdout test settles it.
• Seasonal confounds: A channel that always spends into peak season will absorb calendar lift under weak controls. If its credit collapses once seasonality tightens, it’s likely the calendar (not the channel) was driving the result.
• Flat spend history: Always-on budgets have no experimental variation, so the models are extrapolating saturation from functional form rather than from data. To fix it, introduce deliberate spend variation.
• Adstock window sensitivity: Slow-building channels show near-zero contribution under a short decay window but meaningful contribution under a long one. This tells you the effect is slow-building and long-tailed.
• Data gaps and tracking breaks: Divergence localized to one region or period is often the fastest way to catch a tracking gap. Ideally, you spot it before it distorts a budget decision.

An example of divergent MMM models

As an example, using a synthetic dataset, suppose a direct-to-consumer (DTC) brand spends roughly $1.5 million a month across four channels. There are 2.5 years of weekly data to run through all three models.

Branded search was over-credited by roughly 2x

With no priors pulling it back, Robyn’s ridge regression assigned 41% of revenue to paid search. Ridge credits whatever correlates most tightly with conversions, which was branded search in this example.

Both Bayesian models treated search as partly downstream of existing demand and cut that credit to 19%-22%. A geographical holdout test on branded search settled it at 17% incremental, close to the Bayesian estimates.

The TV estimate followed the decay window

Robyn’s short geometric adstock left TV at 3%. An eight- to 10-week decay window in Meridian and PyMC-Marketing put it at 14%-16%. This disagreement indicates that the true effect is slow-building and long-tailed.

The Meta/Google Shopping split isn’t discernible from this data

Meta and Google scale together every peak season, so each model split their combined ~40% share differently (24%/11%, 31%/9%, 18%/22%), each with its own level of confidence. Side by side, the splits show that it’s impossible to recover the allocation from observational data alone.

The following patterns tend to recur in other datasets, too:

• Seasonal credit swings.
• Small channels that flip sign between models (a sign of a fragile, noise-driven estimate).
• Geographical heterogeneity that a national model averages away.
• Halo paths where upper-funnel video gets credit for contributing to search performance in models that allow indirect effects.
• Uncomfortable convergence, where all three models independently agree that a long-favored channel is underperforming.

Dig deeper: How to avoid marketing mix modeling mistakes that derail results

Turn triangulated results into budget decisions leadership will trust

From here, analysis becomes a framework for decision-making. A single-model finding can turn into a debate about the model itself. But a triangulated finding, with three independent methods in agreement and ideally confirmed by a geographical test, focuses on weighing evidence instead.

Close the loop with testing, not averaging. Model disagreement is a ranked list of genuine unknowns, and every major divergence is a question the observational data can’t answer on its own.

Rather than developing an experimentation roadmap based on intuition, rely on a prioritized list from a model comparison. One or two targeted geographic lift tests can resolve more uncertainty than a quarter of scattershot experimentation.

In the example above, this played out three ways in one quarter. First, the branded search holdout moved the budget on a confirmed 17% incremental number instead of Robyn’s 41% estimate. The TV finding changed the media plan with longer flights and a longer evaluation window, rather than cutting spend based on an adstock artifact.

The Meta/Google Shopping ambiguity became the target of the next experiment instead of triggering a budget cut based on unstable results. And Meridian and PyMC-Marketing can also incorporate experiment results as priors. This way, each test can improve the next model refresh.

A three-month MMM implementation plan

Follow this plan to implement MMM in three months:

• Assemble one dataset: Gather weekly spend, outcomes, and controls, along with two or more years of history. This step is time-intensive, but all other models will reuse the dataset.
• Run Robyn as the baseline: This is the fastest path to a working decomposition. Treat it as a first opinion, not the final answer.
• Add Meridian or PyMC-Marketing as a genuinely different second opinion: Bayesian priors change what receives credit. Geographical structure adds signal to regional data.
• Compare decomposition and response curves, not fit statistics: Log the convergences and the divergences. Consider the latter to be your findings list.
• Plan one geographical test to explore the largest divergence: The result will become the prior that narrows the next model refresh. The disagreement should visibly shrink on the rerun.

Dig deeper: Why marketing mix modeling is still hard to get right

Use multiple models to expose uncertainty

A single MMM can produce a confident point estimate even in the presence of uncertainty. Running three models against the same data makes the uncertainty visible and shows where further testing is required.

The result is an evidence-backed budget recommendation with a clear distinction between findings supported across models and those that still need validation.