A business has several years of spend and revenue data and wants a model to recommend next quarter's channel allocation. The history is a useful starting point, but its length does not establish that the model can separate the effects of channels that always changed together.
An MMM readiness review asks whether the dataset contains enough relevant information for the intended decision. It should identify what can be modeled credibly, what needs repair, and which uncertainties will remain even after the data is clean.
Define the allocation question
Write the decision the model will support. Is the business comparing broad channels, estimating the effect of a new medium, or choosing a spend range for an established channel? The required detail differs.
A dataset that supports broad channel analysis may not support separate estimates for every campaign. Avoid demanding a level of granularity that the information cannot identify.
Also specify the outcome: revenue, qualified leads, subscriptions, or another business measure. Use the metric dictionary to keep its definition consistent across the modeled period.
Build a source inventory
Google's Meridian data preparation guide describes aligned media, spend, outcome, and control inputs, with time and geographic structure. It also explains why variables that influence both media decisions and outcomes can matter to causal interpretation.
For each source, record ownership, coverage, granularity, currency, timezone, extraction method, and known definition changes. Include agencies and historical accounts whose spend may be absent from the current dashboard.
The inventory should reveal whether “total media spend” actually includes every channel in scope. Missing channels can leave the model trying to explain their effects through other variables.
Inspect variation, not just volume
Plot each channel's activity over time and across regions where available. Look for meaningful variation that can help distinguish effects.
If two channels always rise and fall together, the model may struggle to separate them. If a channel ran at nearly the same level throughout the history, estimating its response to very different future spending levels is a demanding extrapolation.
Meridian's amount-of-data guidance emphasizes that data sufficiency depends on the information and model complexity, not just a count of rows. More granular observations are not equivalent to independent randomized observations.
Check the outcome series for breaks
Review changes in ecommerce platforms, revenue definitions, CRM stages, service coverage, and reporting filters. A sudden step in the outcome may reflect measurement rather than demand.
Determine whether historical periods can be harmonized honestly. If not, mark the break and evaluate a narrower scope. Do not silently stitch incompatible definitions into one apparently continuous series.
For revenue, decide how discounts, refunds, tax, and currency are represented. A model cannot recover a consistent economic meaning from a target that changes definition halfway through the dataset.
Prepare a confounder discussion
Ask the people who planned budgets why spending changed. Promotions, product launches, expected demand, distribution expansion, and supply limits may help explain both media choices and business outcomes.
Not every available variable should be added as a control. Some variables may be consequences of advertising or part of the treatment the business wants to estimate. Their role requires a causal argument, not a correlation ranking.
Document the proposed relationships in plain language before fitting the model. This makes the assumptions reviewable by people who understand the business even if they do not write the model code.
Classify missing values correctly
| Missingness case | Appropriate investigation |
|---|---|
| Channel was genuinely inactive | Confirm the true zero from source records |
| Export failed for a week | Recover the missing data rather than assume no spend |
| Region definition changed | Build a documented mapping or revise scope |
| Outcome source was unavailable | Assess defensible treatment and uncertainty |
| Historical channel detail is absent | Consider aggregation or a narrower question |
Keep an audit trail for imputations and exclusions. A completed rectangular dataset is not necessarily a more truthful dataset if its gaps were filled casually.
Evaluate geographic granularity
Regional data can add useful variation, but only when spend and outcomes are assigned consistently and the business process makes the geography meaningful. Check regional coverage, sparse outcomes, and nationally distributed media that cannot be observed at the same level.
Do not manufacture regional precision by dividing national totals across regions without considering the modeling implications. If an allocation is necessary, document its assumptions and test sensitivity.
The geographic experiment guide provides related planning questions about regional measurement and spillover, although an MMM and an experiment are different methods.
Gather calibration evidence with context
Previous experiments can inform model assumptions. Preserve the tested channel, population, dates, spend range, outcome, and uncertainty rather than importing only a point estimate.
Meridian's calibration guidance notes that translating experiments into priors requires judgment because experiments and models can estimate different effects under different conditions.
An old test is not automatically irrelevant, but its transfer to the current business needs an explanation. The same applies to professional expectations used where experimental evidence is limited.
Set a readiness outcome that guides action
Classify the project as ready for an initial model, ready only for a narrower question, or requiring specific data repairs. Name the repair owner and the decision it will enable.
After fitting, inspect model health, out-of-sample behavior where appropriate, uncertainty, and sensitivity to assumptions. A good fit alone does not establish a correct causal allocation.
Use the attribution and incrementality guide to place the model alongside other evidence. The useful deliverable is a decision supported by documented data and assumptions, with uncertainty visible enough to constrain how confidently the business acts.