Marketing Mix Modelling: Measuring Online and Offline Media Together

A model should answer a spending question

Print, outdoor, paid search and social media all run during the same quarter. Sales rise, but the individual platform reports collectively claim more revenue than the business earned. Marketing mix modelling offers an aggregated way to investigate how media and other factors relate to outcomes. It can help with allocation decisions when the data and assumptions support the work. The first question is therefore practical: what budget decision needs an answer? A model built mainly to decorate a dashboard can produce precise-looking numbers without resolving the decision that matters. For a wider view of the options, explore our marketing frameworks and measurement guide.

Choose the outcome and assemble its history

Start with a consistently measured outcome, such as weekly net revenue or fulfilled orders, at the geographic level the design can support. Gather media delivery, spend and relevant business conditions for the same periods. Google’s Meridian documentation describes aggregated outcome, media and control inputs and supports geographic modelling. ¹ Check reporting changes, missing weeks and retrospective corrections before fitting anything. A sudden jump caused by a new accounting rule should not be interpreted as advertising success. Keep data provenance and transformation decisions available to the analyst and the business reviewers.

Separate advertising from conditions that affect demand

Price, promotion, stock availability, distribution and economic conditions can influence sales. The model needs a considered approach to these influences rather than a pile of every available spreadsheet column. Variables that are caused by advertising can require different treatment from variables that influence both advertising and sales. For example, treating every brand search as an independent driver may remove part of the pathway the model is meant to assess. The selection needs causal reasoning and specialist judgement. Ask the analyst to explain what each variable represents and why it belongs in the specification.

Recognise delay and diminishing returns

Advertising effects may continue after delivery, and more spend may yield progressively smaller additional outcomes. Meridian models can represent lagged and saturation effects. ² Marketing mix modelling should make those assumptions inspectable, not bury them in a technical appendix nobody reads. Ask whether the available variation can identify the proposed response curve. If two channels always increase together, separating their contributions can be difficult. A curve that looks smooth is not proof that the business has learned the precise return at every possible spend. Recommendations outside the observed range deserve particular caution. Attribution vs incrementality is explained in a separate guide.

Read contribution and marginal return differently

Average modelled return describes the estimated outcome associated with a period’s spend under the model definition. Marginal return concerns the additional outcome expected from a small further spend change. Those are different budget questions. A channel with a strong historical average can have a weak marginal return at its current saturation level. Report the output units clearly: revenue, contribution or another KPI. Include credible or other appropriate uncertainty intervals and sensitivity to major assumptions. An optimiser cannot remove poor identification; it can only calculate using the model it receives. For the wider picture, see our guide to ROI vs ROAS.

Require validation before accepting a recommendation

Inspect fit, residual patterns, holdout performance where appropriate, plausibility and sensitivity to assumptions. A good statistical fit can coexist with misleading contribution estimates when variables move together. Compare findings with credible experiment evidence and investigate disagreements instead of averaging them blindly. Meridian supports experiment-informed calibration, but the matching of study definitions still matters. ³ Quiamo can help organise a measurement brief, media taxonomy and reporting questions, working with a qualified modelling partner where required. The commercial deliverable should be a decision narrative and an evidence record, not only a beautiful response curve. Marketing mix modelling is only as credible as the validation the team can inspect. If that is your next question, read our guide to geo experiments.

Keep the model useful after the presentation

Agree refresh timing based on decisions and data changes. Preserve consistent media definitions, document major creative changes and retain the financial cost basis used in reporting. A channel label may contain several very different campaigns; aggregation choices should fit the decision. Avoid interpreting a historical estimate as a permanent characteristic of a medium. Category conditions, distribution and creative can change. At each review, ask what the model resolved, which allocations remain uncertain and what research would be most useful next. A model becomes more useful when it helps prioritise better evidence, including the possibility that the available data cannot answer a desired question.

Put the measurement plan into practice

Before commissioning the model, request an example of the decision report rather than only a sample chart. It should explain the outcome definition, modelled period, cost basis, uncertainty and how the proposed allocation was selected. Ask how recommendations will change if assumptions about delay or baseline demand change. Also ask what the modeller will do when two channels cannot be separated with the available data. A credible response may be to combine them or recommend a study, not to assign each a precise return anyway. This procurement discussion makes the analytical scope concrete and protects the business from paying for a result that cannot support its intended decision. Retain the raw data export and the cleaned modelling dataset separately, with a change log. That record makes later reconciliation possible when sales systems or media taxonomies change.

Before the campaign goes live

Use this checklist to organise evidence about model feasibility, correlated channels and uncertainty. Assign owners to the tasks and keep the supporting evidence with the campaign record.

• Define the budget decision and the outcome unit.
• Audit time-series history and geographic consistency.
• Create a stable media taxonomy and relevant cost basis.
• Review causal assumptions and correlated activity with the analyst.
• Require validation, uncertainty and sensitivity analysis.
• Link model refreshes and experiments to actual allocation decisions.

Questions people ask

Bring the measurement question into the brief

The model is useful when it narrows an allocation decision and identifies where better evidence is needed. Marketing mix modelling should preserve assumptions, response limits and uncertainty alongside its charts. Bring Quiamo the spending question, media history and commercial outcome you need to understand. We can help shape the brief and identify the specialist analytical scope. If the data cannot separate two channels convincingly, that limitation should guide the next study rather than disappear inside an apparently exact recommendation. Explore our campaign strategy and measurement support to see how the work can be structured. You can discuss your campaign measurement brief with Quiamo.

Sources

1. Google Meridian introductionMMM uses aggregated outcome, media and relevant control data; geo-level analysis is supported.
2. Google Meridian media saturation and laggingThe model represents delayed response and diminishing marginal returns.
3. Google Meridian introduction to GeoXGeographic experiments compare marketing interventions and require historical geographic data and counterfactual analysis.

Worked examples on this page are fictional and labelled. Search phrases and figures from external sources are cited; nothing here describes a Quiamo client result.