Is marketing mix modeling realistic for a $5 million brand?

The software is accessible: Google's Meridian and Meta's Robyn are free. The data requirements are harder. A useful model needs substantial clean weekly history and relatively few variables. At $5 million in revenue, start by comparing total sales with total advertising spend, then test the largest channel. The $5 million revenue figure is a fictional USD illustration, not an eligibility threshold or a currency conversion.

The benchmark

Marketing mix modeling, or MMM, uses sales and spending series to estimate each channel's incremental contribution. The source compared these open-source libraries as of October 20, 2026:

The software comparison retains the original release-history context. The Meridian repository and Robyn project page were rechecked on October 29, 2026; package versions change and are not readiness criteria.

The constraint is observations relative to parameters. Robyn's analyst guide recommends roughly one variable per ten observations. With 104 weeks, that leaves around ten variables for channels, promotions, prices, and seasonality combined. Meridian's example of twelve channels, six controls, and eight trend knots leaves four points per parameter over two years: too little for reliable estimation. Three channels, five controls, two knots, and three years give roughly fifteen, enough for directional guidance. Controls represent nonadvertising influences; knots shape the underlying trend.

The rule

Before counting weeks, check whether budgets changed. With nearly constant Google and Meta budgets, the model cannot cleanly distinguish effects that never varied. It may return assumptions dressed as findings.

The Club contributors consider MMM more reasonable with two clean years, three or four material channels, and meaningful budget variation. That is editorial judgment, not a vendor-published revenue threshold.

Before then, an incrementality test can answer the more direct question: what would we lose if this channel stopped? Meta's GeoLift and Google's Meridian GeoX are free tools. Google Ads geographic Conversion Lift is another route, subject to account eligibility and access through a Google representative, as confirmed in current Google Ads help. Test results can then inform Meridian's prior assumptions.

The steps

1. Track the whole business. Weekly, divide total sales by total advertising spend across platforms. In an outdoor-brand audit, a contributor recommended this blended view before channel detail. Connect it to margin through the profit-and-loss guide.
2. Inventory the history. Assemble weekly sales or margin, channel spend, promotions, average price, and stockouts for two years. Reconstruct gaps using platform invoices.
3. Count parameters. Use points per parameter = weeks of history ÷ (media channels + control variables + knots). Below Robyn's ten-observation guideline, combine small channels or build more history. Meridian also evaluates uncertainty through credible intervals around channel effects.
4. Inspect variation. Identify weeks when a channel's budget rose or fell materially. If none exist, design a change rather than assuming the model can invent identification.
5. Run a test. Start with the largest channel and geographic groups. GeoLift and GeoX require daily geographic data. GeoX asks for pretest history at least three times the test duration; GeoLift calls for at least a buying cycle. Where regional volumes are too small, the incrementality guide describes a four-week campaign pause. Account for seasonality and competing changes when interpreting it.
6. Model with a decision in mind. With sufficient history and one or two experiments, involve a data scientist. Write down which allocation decision follows each possible finding before running the model.

The pitfall

Free software still requires specialist time and disciplined collection. A poorly specified model can display implausibly precise ROI figures. Meridian's cited guidance works at channel level, not as a verdict on individual campaigns.

In the outdoor-brand audit, paid advertising gained 1,100 orders while organic channels lost 1,400. The contributor described communicating vessels. Channel reporting alone could not show which orders were new and which merely moved attribution columns. Experiments, then a carefully specified model, help separate those effects.

What to tell your team

The following example is fictional:

“We have only seventy clean weeks and a nearly constant Meta budget. MMM would not be dependable today. We will test Meta across regions for four weeks, then consider Meridian next fall with the experimental result as an input.”

Sources

• Google: Meridian is available to everyone
• PyPI: google-meridian releases
• Meta: Robyn official project
• CRAN: Robyn package
• GitHub: Google Meridian
• PyPI: robynpy releases
• Meridian: Frequently asked questions
• Meridian: National-level models
• Robyn: Analyst guide to marketing mix modeling
• Meridian: How much data is needed
• Meta GeoLift: Walkthrough
• Google: Meridian GeoX
• Google Ads: Geographic Conversion Lift setup
• Meridian GeoX: Prepare pretest data