MMM with Meridian and Robyn: data comes before the model

The renewed interest in Marketing Mix Modeling is easy to understand. Less reliable cookies, weaker attribution, privacy constraints, fragmented channels and budget pressure push companies to look for a more stable reading of marketing impact.

Open source tools like Google Meridian and Meta Robyn have made the topic more accessible. The code is visible, the methodology is documented and adoption does not require a closed platform by default. For a company, that is real value.

Accessible does not mean simple. The risk is thinking that installing a framework is enough to get a credible answer about budget, ROI and channels. An MMM does not repair weak data. It exposes it.

What the model does not know about your business

Meridian presents itself as a statistical framework for answering business questions about budget, performance and allocation. Robyn describes MMM as an econometric model used to quantify the incremental impact of marketing and non-marketing activities on a KPI.

These definitions matter because they clarify the goal: not producing an elegant chart, but estimating causal effects well enough to inform budget decisions.

That goal connects MMM to the same discipline behind attribution and Measurement Protocol. Each tool answers a different question. Attribution describes how credit is assigned inside a configured system. Event tracking describes observed behavior. MMM tries to estimate broader contribution across channels and non-marketing factors. Confusing those questions is how teams end up with more reports and less clarity.

To get there, inputs must make sense: coherent media data, reliable KPIs, seasonality, promotions, pricing, distribution, external events, product changes, stock availability and campaigns outside platforms. If these elements are missing or treated as noise, the model can still produce an output. That does not mean it will produce a better decision.

Causality is not validated by a nice fit

Meridian’s documentation on model fit is clear about a point that teams often underestimate: directly validating causal inference quality is difficult. Well-designed experiments would be the stronger reference, but they are not always practical or available. Teams rely on indirect measures, diagnostics and checks.

This matters for operators making real decisions. A model can fit historical data well and still be weak for deciding what to do next. Descriptive accuracy is not enough if relevant variables are missing, channels are collinear or the available granularity does not capture what matters.

Robyn also gives space to calibration, model refresh and experiments. The practical message is the same: MMM needs governance. It should not be consumed like an automatic report.

For a smaller company, the first step is more modest

For many startups, ecommerce teams or scaleups, the first step is not building an MMM immediately. It is preparing the conditions under which a model could make sense later.

That means keeping a coherent history of spend and revenue, documenting campaigns, cleaning naming conventions, separating promotions from media, tracking business events with discipline and preserving major product or pricing changes. It is not glamorous work, but it turns measurement from a collection of dashboards into decision memory.

Only then does it make sense to ask whether Meridian, Robyn or another approach can help. At that point MMM becomes a lens, not an oracle. It can show patterns, tensions and scenarios, but the decision remains with the team.

The value, to me, is not a more sophisticated model. It is enough order for marketing, product and operations to discuss the business from the same base of reality.

Before opening a notebook, I would ask for a plain data readiness check: do spend, revenue, product changes and campaign calendars share a reliable timeline? Are offline effects and promotions documented? Can the team explain where attribution is known to be weak? If those answers are missing, the first milestone is not a model. It is decision memory good enough for a model to deserve attention.

That preparation is not wasted if the team never runs a full MMM. Cleaner campaign history, named assumptions and reconciled business events already improve weekly operating decisions. The model is only one possible consumer of that discipline.

That is the pragmatic threshold: if the preparation would not improve a normal business review, it probably is not mature enough to feed a causal model either.