Marketing Mix Modeling for Mid-Size Budgets

What Is Marketing Mix Modeling and Why It Is Back

A SaaS client came to me in April spending EUR 35,000 per month across Google Ads, Meta, LinkedIn, and a branded podcast. Their attribution stack was a mess: GA4 showed one story, each ad platform told another, and Safari's seven-day client-side cookie cap meant roughly a quarter of their web traffic was invisible to analytics within a week. The VP of Marketing had been presenting GA4's data-driven attribution as ground truth, but after reading my breakdown of why attribution models disagree, she realized she was making budget calls on incomplete data.

She asked me a question I hear more and more: "Should we try marketing mix modeling?"

Marketing mix modeling (MMM) is a statistical method that measures how each marketing channel -- plus external factors like seasonality, pricing, and competitor activity -- contributes to a business outcome (usually revenue or conversions). Unlike user-level attribution, it works on aggregated data: weekly spend, weekly revenue, holidays, anything that might influence sales. No cookies. No consent banners. No cross-device identity graph. Just time-series regression with adstock and saturation curves.

The technique has been around since the 1960s, but it spent decades locked inside CPG conglomerates and media agencies with the budgets to match. What changed is the open-source tooling. Between 2022 and 2025, Meta released Robyn, Google launched Meridian, and PyMC Labs shipped PyMC-Marketing. Suddenly, marketing mix modeling software that would have cost six figures in consulting fees runs on a laptop.

But "free" and "ready" are not the same thing. Here is what mid-size teams actually need to know.

Who Should (and Should Not) Build a Marketing Mix Model

MMM is not universally appropriate. I have talked clients out of it as often as I have recommended it. The deciding factors are budget scale, channel diversity, and data history.

Minimum Requirements

Factor	Threshold	Why it matters
Monthly marketing spend	EUR 15,000+	Below this, the signal is too weak for the model to distinguish channel effects from noise
Active paid channels	3+	With one or two channels, you are better off running platform-level lift tests
Historical data	18+ months, weekly	The model needs enough variation in spend levels to estimate diminishing returns curves
Outcome data	Consistent weekly revenue or conversions	Garbage in, garbage out -- if finance and marketing disagree on the numbers, fix that first

If you meet those thresholds, MMM can give you something attribution cannot: a view of diminishing returns. Attribution tells you which channels get credit. MMM tells you where your next euro of spend will produce the most incremental outcome -- and where you have already saturated the channel.

When to Skip It

If your sales cycle exceeds six months and conversion volume is low (under 50 deals per quarter), you likely need a different approach. I wrote about this in detail in Marketing Measurement for Long B2B Sales Cycles -- the short version is that CRM-based offline conversion imports combined with incrementality tests will serve you better than aggregate modeling when you have sparse outcome data.

Open-Source MMM Tools: Robyn, Meridian, PyMC-Marketing

The three dominant open-source MMM software options each come with real tradeoffs. I have built models in all three. Here is an honest comparison.

Meta Robyn

Robyn is written in R and uses ridge regression with a multi-objective evolutionary algorithm for hyperparameter optimization. It automates a lot of the modeling choices that used to require a senior econometrician: adstock decay, saturation curves, trend decomposition.

Strengths: Fast iteration. The automated hyperparameter search reduces human bias in model selection. Good documentation. Active community.

Weaknesses: R-only, which is a hard stop for Python-native teams. The automated model selection can be a black box if your team does not understand what the Pareto-front optimization is actually doing. And being a Meta product, it defaults to assumptions that tend to favor digital-heavy media mixes.

Google Meridian

Meridian uses Bayesian causal inference and supports geo-level hierarchical modeling, which is a significant methodological advantage. If you run campaigns in different regions, Meridian can use geographic variation in spend to strengthen causal estimates.

Strengths: Bayesian uncertainty quantification (you get confidence intervals, not point estimates). Geo-level modeling. Python-based. Google provides reach and frequency data integrations for YouTube and Search.

Weaknesses: Heavier computational requirements. The Bayesian sampling is slower than Robyn's optimization. Requires more statistical expertise to configure priors correctly. And like Robyn, "open source" does not mean "plug and play." You still need someone who understands marketing econometrics.

PyMC-Marketing

PyMC-Marketing is the most flexible option -- fully Bayesian, built on top of PyMC, and designed for teams that want to customize the model rather than use a pre-packaged pipeline.

Strengths: Maximum transparency. You can inspect and modify every assumption. Supports MMM and customer lifetime value modeling in a single framework. Bayesian priors can partially compensate for shorter data histories by encoding domain knowledge, which may reduce the amount of historical data needed compared with frequentist approaches.

Weaknesses: Highest technical barrier. This is not MMM software for a marketing ops team -- it is a data-science toolkit. Expect to invest significant engineering time.

The Honest Cost of "Free"

All three tools are free to download. None are free to operate. You need someone who can clean and prepare the input data (weekly spend by channel, revenue, external variables), configure the model, interpret the outputs, and maintain it over time. In my experience, the first model build takes 40-80 hours of skilled work, and quarterly refreshes add another 15-25 hours each.

For teams without a data scientist, the total cost of running an open-source MMM -- including contractor or hire costs -- often lands between EUR 30,000 and EUR 80,000 in the first year. That is significantly less than a traditional agency engagement, but it is not free.

SaaS Alternatives for Teams Without Data Science

If you do not have a data scientist on staff and do not want to hire one, several SaaS platforms now offer MMM as a managed service. Tools like Recast, Sellforte, and Cassandra provide pre-built pipelines, automated data connectors, and ongoing model updates.

In my experience, pricing typically runs EUR 2,000-5,000 per month (as of mid-2026). What you gain is speed to first insight (days instead of months) and ongoing maintenance without internal headcount. What you lose is full control over model assumptions.

For teams spending EUR 20,000-100,000 per month on marketing, a SaaS solution often makes more sense than building from scratch. Above that, the economics shift toward in-house or hybrid approaches.

Building Your First Model: A Practical Checklist

If you decide MMM is right for your situation, here is the sequence I follow with clients.

Step 1: Audit Your Data

Before you touch any modeling tool, get your input data right. You need:

• Weekly spend by channel. Not campaign-level (too granular), not monthly (too few data points). Platform-exported spend data, reconciled with finance.
• Weekly outcome metric. Revenue, qualified leads, or conversions -- whatever the business optimizes for. Pulled from the CRM or accounting system, not from GA4.
• External variables. Seasonality, holidays, pricing changes, promotions, competitor activity. These prevent the model from misattributing seasonal spikes to your December ad push.

I cannot overstate how many modeling projects fail at this step. If you are building your analytics on shaky foundations -- broken conversion tracking, inconsistent event definitions, mismatched data sources -- fixing your measurement infrastructure comes before any modeling exercise.

Step 2: Choose Your Tool

For most mid-size teams, I recommend starting with Meridian or Robyn. If you have an R-comfortable analyst, Robyn's faster iteration cycle is appealing. If your team is Python-first or you want Bayesian uncertainty quantification, Meridian is the better fit. If you have neither, consider a SaaS platform.

Step 3: Build, Validate, and Iterate

The first model will be wrong. That is expected. The validation process matters more than the initial output:

• Compare model predictions to actual results for a holdout period. If you build the model on 2024-2025 data, test its predictions against Q1 2026 actuals.
• Check channel coefficients against business intuition. If the model says your top-performing channel has negative ROI, something is wrong with the data or the model configuration, not with your marketing.
• Run scenario analysis. Shift 20% of budget from your most saturated channel to an undersaturated one. Does the model's prediction directionally match what you would expect?

Step 4: Refresh Regularly

A marketing mix model is not a one-time project. Markets shift, channel effectiveness changes, and your spend mix evolves. I recommend quarterly refreshes at minimum, with a full rebuild annually if your channel mix has changed significantly.

MMM vs. Attribution: They Answer Different Questions

One mistake I see repeatedly: teams treating MMM as a replacement for attribution. It is not. They are complementary.

Approach	What it answers	Data type	Time horizon
Multi-Touch Attribution (MTA)	Which touchpoints contributed to this specific conversion?	User-level	Days to weeks
Marketing mix modeling	How does each channel contribute to total business outcomes?	Aggregated	Months to years
Incrementality testing	What would have happened without this channel?	Experimental	Per test

The ideal measurement stack uses all three. Attribution handles tactical, in-flight optimization. MMM handles strategic budget allocation. Incrementality tests calibrate both. For a deeper look at when multi-touch attribution earns its setup cost, see my multi-touch attribution post.

Most mid-size teams should start with clean attribution and layer in MMM once they have the data history and budget scale to support it.

Common Pitfalls I See in Practice

Overfitting to Noise

With limited data, it is tempting to add more variables to improve model fit. Do not. A model that fits historical data perfectly but cannot predict next quarter is useless. Prefer simpler models with fewer channels grouped logically (all paid social as one variable, for instance) over complex models that try to separate TikTok from Instagram with 18 months of weekly data.

Ignoring Offline and First-Party Data

MMM can incorporate offline channels -- events, direct mail, TV -- which is one of its advantages over digital attribution. But only if you actually collect and structure that data. If your events team tracks spend in spreadsheets with inconsistent date formats, that data is not model-ready. For practical steps on building a usable first-party dataset, see my first-party data playbook.

Treating Model Output as Truth

A marketing mix model gives you estimates with uncertainty bands, not facts. The model might say that Meta's ROI is 2.3x with a 90% credible interval of 1.4x-3.1x. That is useful directional guidance. It is not a precise number to put on a slide without context.

Marketing Mix Modeling FAQ

Not sure whether your measurement stack is ready for MMM -- or whether your tracking data is trustworthy enough to feed a model? Book a measurement audit. I will review what you have, tell you what is broken, and recommend the fastest path to budget decisions you can actually trust.