A useful marketing mix modelling project starts with a decision, not a promise to recover every conversion lost to tracking restrictions. MMM uses aggregate data to estimate how marketing and other factors relate to a business outcome. The practical work is deciding what the data can support, checking the assumptions and turning uncertain estimates into a controlled budget decision.
Substantively reviewed and updated: 22 October 2026. The original 2023 discussion anticipated a broadly “cookieless” future. That framing was too sweeping. First-party and third-party cookies differ, as do browser policies. Google’s 22 April 2025 announcement said Chrome would maintain its then-current approach to third-party-cookie choice. This dated decision is historical context, not a claim about every browser’s policy today.
1. Specify the decision and outcome
Write down the business question: should a defined budget move between channels next quarter, and what outcome should improve? Revenue, orders, subscriptions and contribution are not interchangeable. If modelling orders, explain how order value and margin enter the eventual financial decision.
Set a reporting grain appropriate to the business, such as weekly observations by geographic market. The available history, variation in activity and number of independent markets constrain the detail you can estimate. A spreadsheet with many rows does not necessarily contain enough independent information to distinguish similar channels.
If you need the conceptual overview first, see what MMM can and cannot measure.
2. Build a consistent dataset
- Align the outcome, media activity and spending to the same geography and time periods. Keep currency, tax treatment and attribution definitions consistent.
- Record relevant non-media factors, including prices, promotions, availability and distribution. Seasonality and economic conditions may affect both demand and advertising decisions.
- Audit missing values, tracking changes, duplicate records and changes in channel definitions. Record genuine zero activity separately from unknown activity.
- Inspect whether channels move together. If two channels always rise and fall together, separating their effects may depend heavily on modelling assumptions.
Choose controls with a causal rationale. A variable influenced by advertising may be part of the mechanism you want to measure rather than an appropriate control. More columns do not automatically remove bias.
3. Represent lag and diminishing returns
Advertising may affect outcomes after the week it runs, and additional spending may yield smaller incremental returns. MMMs commonly represent these through lag or adstock functions and saturation curves. The functional choices, parameter constraints and priors can materially affect the answer.
Google’s Meridian introduction describes a Bayesian workflow with controls, lagged effects, saturation, posterior estimates and budget optimisation. It is a concrete implementation to study, not evidence that any fitted model is correct. Record model choices so someone else can inspect and reproduce the analysis.
4. Challenge the estimates
Check fit, residual patterns and whether predictions generalise to held-out periods where feasible. Compare plausible specifications and control sets. Examine whether channel contributions or response curves change sharply under reasonable assumptions. A good prediction score alone does not establish that the estimated channel effects are causal.
Use relevant experiments to inform or challenge the model when available. Match the tested population, period and intervention to what the MMM estimates; a lift result from one campaign is not automatically transferable to every future campaign. Investigate disagreement rather than adjusting the model only until it reproduces a preferred result.
Report intervals and the limits of identification. When the data cannot separate two tactics reliably, report a broader grouping or uncertainty rather than presenting precise rankings.
5. Turn a scenario into a bounded test
Suppose a model suggests moving part of the next quarter’s budget from one channel to another. Compare the expected marginal outcome at the proposed spending levels, not just historical average ROI. Include operational constraints, minimum commitments and the possibility that creative quality or market conditions will change.
Treat optimisation as conditional on the fitted response curves. Large moves beyond observed spending ranges are extrapolations. Start with a feasible change, monitor the business outcome and use an experiment where practical. Keep a record of the forecast, assumptions and realised result so the next review can distinguish a useful recommendation from a confident forecast that failed.
MMM is one part of measurement, alongside experiments, customer research and operational knowledge. For a narrower financial question, measuring social media ROI explains incremental contribution, programme cost and the assumptions behind counterfactual estimates. The goal is a better-supported decision, not a model that appears to explain every sale.
Leave a Reply