Marketing Mix Modelling is only as good as the data you feed it. A data audit before you start saves weeks of model-building time and prevents the embarrassing situation of a model that cannot be trusted because of fixable data issues.
Marketing Mix Modelling (MMM) works by finding statistical relationships between your historical marketing spend and your sales. If your data has gaps, errors, or inconsistencies, the model will fit those flaws into its output. The result is a model that looks impressive but gives you wrong answers. Auditing your data before you start is not optional. It is the most important step in the whole project.
What an MMM data audit covers
An audit checks three things: completeness (do you have all the data you need?), consistency (is the data formatted and categorised the same way across all periods?), and accuracy (does the data reflect what actually happened?). Problems in any of these categories will affect your model output, so all three need attention before modelling begins.
The data you need for MMM
- Weekly marketing spend by channel, ideally going back 2 to 3 years (104 to 156 weekly data points minimum).
- Weekly or daily sales data: units sold, revenue, or orders, depending on your business model.
- Media impressions or GRPs (Gross Rating Points, a measure of advertising reach and frequency) for TV or OOH campaigns.
- Pricing data: average selling price or any promotional price changes over the period.
- Distribution data: number of stores, listings, or active markets, if relevant.
- External variables: competitor activity data if available, search volume trends, economic indicators for your category.
- Promotional and event data: bank holidays, sales events, major sporting events that affect your category.
Common data problems and how to fix them
The most common data problem is gaps. A channel that went dark for three months, data pulled from a platform before a tracking change was made, or simply data that nobody kept records of. Flag every gap before modelling starts. Some gaps can be imputed using seasonality patterns or benchmarks. Others require a shorter modelling window that avoids the gap entirely. Knowingly modelling over an unaddressed gap produces unreliable results for the whole channel.
The second most common problem is inconsistency. Spend data that has been split differently across agency invoices and platform exports. Campaigns that were reclassified mid-year from one channel to another. A change in how you count conversions that makes pre-change and post-change data incompatible. Reconcile all of these before handing data to your modelling team.
One of the most overlooked data problems in MMM preparation is price and promotion data. If your model cannot see when you ran a sale or changed your price, it will attribute the associated sales spike to whatever marketing was running at the time. This systematically overstates the impact of any channel that happened to be active during a promotional period.
How to structure your data audit
Create a data log that lists every data source, the date range available, the granularity (daily or weekly), the owner of that data, and any known quality issues. Assign a RAG status (Red for serious problems, Amber for minor issues, Green for clean data) to each source. Resolve all Reds before you start the model. Amber issues need to be documented as limitations and factored into how you interpret results.
How long does a data audit take?
For a business with 5 to 10 marketing channels and organised data infrastructure, a thorough audit takes 2 to 4 weeks. For businesses with fragmented data across multiple agencies, ad servers, and internal systems, allow 6 to 8 weeks. The time investment is worth it. A model built on audited data typically takes less time to run and produces more reliable outputs than one built on unaudited data that then requires remedial work mid-project.
We only have 12 months of data. Can we still run MMM?
It is possible but not ideal. MMM needs sufficient variation in your spend and sales data to identify patterns reliably. With only 12 months of weekly data (52 data points), results will be sensitive to individual weeks and the model may struggle to separate seasonal effects from channel effects. If possible, prioritise gathering older data before starting the project. If 12 months is all you have, the model will still produce outputs but treat them as indicative rather than definitive.
Do we need spend data from every channel, including very small ones?
Include every channel where spend is more than 1 to 2 percent of your total budget. Very small channels can be grouped into a combined other category. Leaving them out entirely risks attributing their effect to whichever channel the model uses as a catch-all for unexplained variation. The more complete your spend data, the cleaner the model.
Our agency holds our media spend data and has been slow to provide it. What should we do?
Escalate this immediately and make media spend data access a contractual requirement in all future agency agreements. Specify data format, granularity, and delivery timeline. Agencies who are slow with data either have internal systems issues or, less charitably, prefer that you cannot independently verify their reported performance. Either way, you need this data to run your business effectively.