Marketing Mix Modeling: Better Budget Calls

A paid-search dashboard claims credit for a sale. A social platform reports the same customer as a conversion. The affiliate network also takes a share. Then finance asks a harder question: if the company moves next quarter’s money from one channel to another, what is likely to happen to total revenue?

Marketing mix modeling is built for that budget question. It relates changes in an aggregate business outcome—such as revenue, units sold, or conversions—to changes in media and other business conditions over time and, when available, across geographic markets. Google describes MMM as an aggregate method that measures marketing alongside non-marketing factors, without cookies or user-level information (Google Meridian). Its useful output is not a reconstructed customer journey. It is an estimate of how much each channel contributed, how returns changed across the observed spending range, and where another dollar may work hardest.

That distinction determines when MMM is worth doing. I would use it to set channel-level budgets across online and offline media when the business has enough history and real variation in activity. I would not use it to choose between two headlines in a campaign launching tomorrow, explain one person’s path to purchase, or manufacture precise channel rankings from a short, nearly constant spending history. The method can cover more of the marketing mix than user-level attribution, but it pays for that reach with lower granularity, stronger assumptions, and wider uncertainty.

Marketing mix modeling answers an incremental budget question#

At its core, an MMM compares many observations of the same business under changing conditions. Those observations might be weeks for a national model, or market-weeks for a geographic model. The outcome sits on one side; paid media, promotions, price, trend, seasonality, market differences, and selected outside influences sit on the other. The model estimates which combinations best explain the outcome while representing how advertising works over time.

The word incremental matters. Revenue that happened while an ad was running is not automatically revenue caused by that ad. A brand may advertise more before a holiday precisely because demand is about to rise. Search spending may increase when consumer interest is already climbing. If the model treats every concurrent sale as a media effect, it rewards channels for demand they did not create.

Controls are intended to separate those forces. Meridian’s data guidance says control variables should focus on factors that affect both the outcome and media execution, and suggests asking planners what influenced their budget decisions (Google’s data collection guidance). That conversation is often more valuable than downloading a long list of economic series. The pricing change, distribution outage, competitor launch, product release, promotion, or planning rule that actually moved both spend and sales belongs in the discussion. An unrelated variable does not become useful merely because it is available.

MMM therefore does more than divide sales among channels. A well-specified model can estimate channel contribution and return on investment, describe response curves, and compare future allocation scenarios. Those are decision aids conditioned on the data and assumptions, not invoices proving that a channel “owns” a fixed amount of revenue. The correct management question is not “Which dashboard is right?” but “Which allocation is most defensible within the range we have observed, and how uncertain is that choice?”

The model must represent memory and diminishing returns#

Two features keep an MMM from treating advertising as a simple same-week, straight-line input. The first is carryover. Exposure this week may influence sales later, so a lag transformation spreads part of the media response across subsequent periods. The second is saturation. Moving from no presence to a meaningful level of exposure can have a different marginal effect from adding the same amount to an already heavy schedule.

Meridian’s published specification represents carryover with an adstock transformation and saturation with a Hill function. It can apply these transformations to paid and organic media while also including controls, non-media treatments, time effects, and geographic effects (Meridian model specification). The practical lesson is more important than the formula: last week’s campaign can still matter, and the next dollar does not have to perform like the average dollar already spent.

Average ROI and marginal ROI consequently answer different questions. Average ROI compares the estimated incremental outcome from a channel with its spending over a defined period. Marginal ROI concerns the return near the current spending level. A mature channel can have strong historical average ROI but weak marginal ROI because it is near saturation. A smaller channel can show a promising marginal return yet lack enough observed scale to justify a dramatic increase. For allocation, I would emphasize marginal response and uncertainty, while keeping average ROI as context.

Response curves should not be extended casually beyond the observed data. If paid social varied only within a narrow weekly band, the model learned little about what happens after doubling it. If television was always on at almost the same weight, the model had little contrast from which to separate television from the baseline. A smooth curve can still be weakly determined. The appropriate response is a restrained scenario range or a deliberately designed market test, not extra decimal places.

Data readiness matters more than the modeling package#

A credible project starts with one outcome that matches the decision. Revenue is natural when the budget owner wants financial return. Units or conversions can work when revenue is unavailable, provided the metric can be summed across the chosen time and geography. Meridian’s input documentation lists the usual structure: a KPI, media exposure, media spend, time, geography, and control variables; it also supports reach and frequency, organic media, and non-media treatments where relevant (Google’s data collection guidance).

Exposure and spend should not be treated as interchangeable without thought. Spend is needed for return calculations, while impressions, clicks, reach, or another exposure measure may better represent the media delivered. Meridian warns that using spend as an exposure proxy can mislead when media costs change: more expensive inventory can produce higher spending without more impressions. I would preserve both spend and exposure when both exist, document any definition changes, and reconcile channel totals to finance before fitting a model.

Geography creates valuable additional contrast. Instead of asking only whether national sales rose in a week when national media rose, a geo-level model can also ask whether markets with different media levels showed different outcomes. Meridian is designed as a geo-level hierarchical model but permits a national model when geographic data is unavailable (Meridian model specification). Geographic detail is not magic, however. National campaigns copied identically into every market add rows but not independent media variation, and inconsistent market definitions can create false differences.

History also needs to be long enough for the number of variables and the rhythms of the business. Meta’s Robyn analyst guide recommends at least two years of weekly data for its workflow and more history when only monthly data is available. The same guide stresses variation and sufficient observation volume, warning that unnecessary variables can overspecify the model (Meta Robyn analyst guide). That two-year figure is a framework recommendation, not a universal pass mark. Two years of steady spend across twelve correlated channels can be less informative than a shorter, well-varied geo dataset with fewer channel groups.

Before approving an MMM, I would inspect the dataset as a decision artifact. Plot the outcome and every major input by time. Mark launches, outages, promotions, price changes, tracking breaks, and unusual demand periods. Compare planned with delivered media. Check whether supposedly separate channels always move together. Count active periods and measure the spending range for each channel. These checks reveal whether the model will encounter genuine contrasts or merely be asked to divide credit among synchronized lines.

Granularity should follow both the decision and the available information. Separate search from social if budgets can move between them and their activity varies independently. Split brand and performance campaigns only if execution records, variation, and observation volume can support that distinction. Combining everything into “digital” hides useful choices; splitting every publisher, format, audience, and creative can leave the model unable to distinguish any of them. I would start at the highest level where money can actually be reallocated, then add detail only when a specific decision justifies it.

A useful MMM is designed backward from a decision#

The first deliverable should be a written decision question, not a model. “Estimate channel ROI” is incomplete. “Decide how to distribute the next quarterly media budget across search, social, television, and online video while maintaining national coverage” tells the team what outcome, channels, time horizon, and constraints matter. It also exposes questions MMM may not answer, such as which creative should run or which audience should receive it.

Next, map each candidate variable to a business mechanism. Media variables represent interventions the company can change. Non-media treatments can represent actions such as promotions or pricing. Controls represent influences that confound the relationship between media and the outcome. Trend, seasonality, and geographic terms absorb recurring or structural patterns. Meridian’s specification can represent all of these components, but including a column does not guarantee its separate contribution can be estimated (Meridian model specification). Correlated inputs can still trade contribution between themselves.

Then fit a small set of defensible specifications rather than searching mechanically for the most flattering answer. Reasonable alternatives might change channel grouping, lag assumptions, controls, or the flexibility assigned to time. The point is to learn whether the budget conclusion survives plausible choices. If one specification sends money to search and another equally reasonable one sends it to television, the decision is not stable enough for a confident reallocation.

Consider an illustrative company with a $10 million annual media budget across six channels. Suppose its model suggests moving $800,000 from two mature channels toward online video. The executive recommendation should not be “online video wins.” It should state the modeled spending ranges, the expected outcome under the proposed mix, an uncertainty interval, and the assumptions that could reverse the choice. If online video was observed only at low spend, I would stage the move in smaller increments and create geographic variation that makes the next model more informative. The initial compromise is slower theoretical improvement in exchange for less extrapolation risk.

This staged approach turns allocation into learning. A company can reserve part of the budget for controlled variation, use the model to identify where information is weakest, and refresh the model after new outcomes arrive. MMM then becomes a repeated planning process rather than an annual presentation that produces attractive response curves and disappears.

A good fit does not prove the channel split is right#

A model can track total sales well and still assign the wrong shares to correlated channels. Predictive fit asks whether the model reproduces held-out or observed outcomes. The budget decision asks whether the modeled marketing effects are credible. Meridian’s guidance treats out-of-sample prediction as a useful preliminary check, but not the primary basis for judging causal marketing effects (Meridian model-fit guidance).

Model review should therefore combine several kinds of information. The predicted outcome should follow the broad shape of the actual outcome without implausible misses. Baseline and channel contributions should be plausible in light of how the business operates. Results should remain reasonably stable under defensible changes in inputs and assumptions. Uncertainty should be visible rather than reduced to one ROI number. Most importantly, the recommendation should change when its underlying assumptions change; a result that remains perfectly certain under weak data deserves suspicion, not admiration.

Experiments can add a stronger reference point, but the comparison has to be aligned. A test of paid social in selected markets over eight weeks does not automatically confirm an annual national social estimate. Treatment, outcome, population, geography, and period need to describe the same quantity. Google’s guidance explicitly notes that a well-designed experiment must estimate the same target as the MMM before it can directly assess the modeled effect (Meridian model-fit guidance). When alignment is good, experimental results can inform model assumptions; when it is not, the discrepancy may reflect different questions rather than a failed model.

The final recommendation needs limits as well as numbers#

The clearest MMM readout begins with the decision and the safe range of action. It states which budget move is supported, the period and markets represented, and the downside if the estimates are wrong. Channel contribution, average ROI, marginal ROI, and response curves then explain the recommendation. Technical fit statistics belong behind that story, available for scrutiny but not substituted for business judgment.

I would reject a readout that ranks channels solely by point-estimate ROI. Rankings conceal overlapping uncertainty and ignore scale. They also encourage the false idea that all spending should flow to the apparent winner. A budget allocation must account for saturation, feasible spending changes, channel roles, contractual commitments, and the fact that historical data may not contain the future conditions being planned.

The strongest conclusion an MMM can support is conditional: given this outcome, these markets and periods, these controls, and these response assumptions, a specified allocation is preferable within a stated range. That is not a weakness. It is the information a budget owner actually needs. Marketing mix modeling earns its place when it converts aggregate history into a cautious, testable spending decision—and makes clear where the data is not yet strong enough to decide.