MMM need decision boundaries, not attribution theatre

Marketing mix modeling is becoming attractive again because the old promise of user-level attribution is weaker than many teams admit. Privacy constraints, fragmented journeys, offline effects, retail media, marketplaces, and brand demand all make the neat last-click story less credible. A good MMM can help a founder, CMO, or growth lead see how channels appear to contribute to business outcomes over time. But that does not make it a budget oracle.

The thesis is simple: marketing mix models need decision boundaries, not attribution theatre.

Attribution theatre happens when the team treats a model as a more sophisticated way to justify decisions it already wanted to make. The deck has curves, priors, confidence intervals, and channel decomposition, but the operating question remains vague: what are we actually allowed to decide from this output? Without that boundary, MMM becomes another dashboard ritual. It looks mature, but it does not change how budget moves.

This matters because MMM is not a replacement for a tracking plan, experimentation, or ownership. It is a decision-support layer. It should say where evidence is strong enough to act, where evidence is too weak, and where another measurement method is needed. That is the same governance problem behind Measurement Protocol is not a tracking plan: instrumentation can collect events, but it cannot decide what the organization is responsible for learning.

The model is not the decision

Google describes Meridian as an open-source MMM framework designed to help advertisers measure marketing effectiveness and optimize budget allocation. Meta Robyn is also positioned as an open-source MMM solution, with features for adstock, saturation, model selection, and budget allocation. Those are useful capabilities, and they are serious tools. They still do not remove the need for a decision contract. See Google Meridian and Meta Robyn for the technical basis.

A model can estimate relationships between spend, time, controls, and outcomes. It can show that a channel appears saturated, that another channel may have room to scale, or that a short-term cut may carry a longer-term cost. But the model does not know your cash position, sales capacity, inventory constraints, board expectations, creative pipeline, or risk appetite.

That distinction is not academic. If the model says paid social has a lower marginal return than search, does the team immediately move 20 percent of the budget? Does it run a four-week test? Does it wait because the creative mix just changed? Does finance approve the shift or does the growth lead own it? The model cannot answer those questions unless the organization has defined the boundary around the decision.

The right question is not: can MMM tell us the truth? The better question is: which decisions can this MMM safely inform, given the quality of the inputs, the stability of the business, and the consequence of being wrong?

What decisions can MMM actually support?

MMM is strongest when the decision is about directional allocation across channels, geographies, markets, or time periods, especially when the business has enough variation in spend and outcomes to learn from. It can support questions such as: should we keep investing in a channel that looks saturated? Is the next euro or dollar likely to work harder in brand, search, retail media, or lifecycle? Are we underweight in a region where demand responds differently? Should we protect a baseline budget even if short-term attribution looks weak?

That does not mean the model should decide everything. MMM is usually a poor tool for judging one ad, one landing page, one keyword cluster, or one audience segment in isolation. It is also dangerous when the data is too smooth, too sparse, or too politically curated. If every channel always increases spend at the same time, the model has little independent variation to learn from. If promotions, pricing, stockouts, seasonality, and sales changes are missing, the model may give marketing credit or blame for effects it did not create.

This is why the decision boundary must be explicit. Before the model is run, write down the action classes it can influence:

1. Reallocation decisions, such as moving budget between channels within a controlled range.
2. Protection decisions, such as defending spend that supports delayed or upper-funnel demand.
3. Investigation decisions, such as requiring a geo test, lift test, or data audit before budget changes.
4. No-action decisions, where the signal is too uncertain to justify movement.

The fourth class is the one mature teams include and theatrical teams avoid. A no-action zone is not indecision. It is a guardrail against overfitting the business to a beautiful chart.

Where does attribution theatre begin?

Attribution theatre begins when the organization asks MMM to produce certainty instead of forcing a better decision process. The symptoms are easy to spot.

The first symptom is retroactive framing. The model is built after a budget debate has already hardened, and the output is used to decorate the winning argument. The second is metric laundering. The team changes from platform ROAS to MMM contribution without changing the quality of the underlying data or the decision owner. The third is selective sophistication. People debate adstock curves and priors but ignore missing cost data, inconsistent campaign taxonomies, or offline constraints.

A fourth symptom is ownership ambiguity. If the model says a budget shift is justified, nobody knows who can approve it. Marketing says finance owns the budget. Finance says marketing owns performance. Product says the conversion rate changed because the onboarding flow changed. Sales says the leads were different. The model becomes a shared object that everyone can cite and nobody has to own.

This is where MMM connects to operating model design. A measurement system only becomes useful when decision rights are clear. The same principle applies in product work, as argued in Product operating models start with decision rights. If nobody owns the action, the insight is not operational. It is commentary.

Build the boundary before the model review

The practical artifact is a decision-boundary map. It can be simple: input data, model output, decision boundary, budget action or no-action zone.

Start with input assumptions. Which outcome is being modeled? Revenue, contribution margin, pipeline, activated accounts, qualified demand, or something else? Which channels are included? Which non-marketing variables are included, such as seasonality, pricing, promotions, distribution, sales coverage, product launches, or macro conditions? Which data is trusted, partially trusted, or known to be weak?

Then define model outputs in action language. Do not stop at channel contribution. Translate outputs into decision prompts: increase, decrease, protect, test, audit, or wait. Each prompt needs a threshold. For example, a budget decrease might require repeated evidence of saturation, alignment with experiment results, and no known data-quality blocker. A budget increase might require available creative capacity and a defined downside limit.

Then assign ownership. The model owner is not always the decision owner. Analytics may own the model. Marketing may own the recommendation. Finance may own the approved budget range. The founder or CMO may own exceptions. Write this down before the meeting, not after the disagreement.

Finally, define the no-action zone. This is the range where the model is informative but not decisive. It may trigger a data audit, an incrementality test, a taxonomy cleanup, or a decision to hold spend steady until the next refresh. IAB Tech Lab measurement work is a useful reminder that measurement depends on standards, definitions, and interoperability, not only analytical ambition. See the IAB Tech Lab Measurement standards for the broader standards context.

A useful MMM meeting has fewer opinions

A good MMM review should not feel like a courtroom where each channel manager defends their budget. It should feel like an operating meeting with predefined moves.

For each recommendation, ask five questions. What decision is this output allowed to inform? What evidence would make us reverse the decision? What risk are we taking if the model is wrong? Who approves the budget action? When will we review the result?

Those questions make the model less magical and more useful. They also reduce politics. If the boundary says that weak signals go to a test, the team does not have to argue whether a chart is beautiful enough to move money. If the boundary says that high-confidence saturation can trigger a limited reallocation, the team knows the size and owner of the move.

This is also where MMM and tracking work meet. Event instrumentation, tagging, and platform data still matter. A broken measurement foundation will not become reliable because the model is advanced. For teams using MMM with modern tools, MMM with Meridian and Robyn: data comes before the model is the companion principle: before asking the model to decide, make the data explainable enough to trust.

From measurement theatre to budget action

The value of MMM is not that it replaces judgment. The value is that it disciplines judgment. It gives leadership a structured way to discuss uncertainty, trade-offs, delayed effects, saturation, and budget movement without pretending that every customer journey is fully visible.

But MMM only earns that role when its decision boundaries are written down. What can it decide? What can it not decide? Which inputs are assumed to be reliable? Which outputs trigger action, investigation, or no action? Who owns the budget move?

If those questions are missing, the model may still be impressive. It may still be technically valid in parts. But it will drift toward attribution theatre: a sophisticated explanation layer that lets the organization avoid responsibility.

The better path is more operational and less glamorous. Build the boundary. Name the owner. Define the no-action zone. Then let the model do what it is good at: support better decisions, not pretend to be perfect truth.