FMCG (fast-moving consumer goods) brands invented marketing mix modelling. But the sector has changed: more channels, weaker panel data, and faster product cycles mean models need to be built differently than they were a decade ago.
Marketing mix modelling has its deepest roots in FMCG. Procter and Gamble, Unilever, and Nestle were using econometric models to plan TV spend in the 1990s. The principle is unchanged: take sales data, media spend data, and contextual factors, and use statistics to separate out what drove the numbers. What has changed is the data environment. Panel data (surveys of households about what they buy) is weaker than it used to be. Retailer data is more available but more fragmented. And the media mix has expanded from TV and print to include dozens of digital channels.
The core FMCG measurement challenge
FMCG brands rarely sell directly to consumers. They sell through supermarkets, convenience stores, and online retailers. That means they do not own the transaction data. To measure the effect of marketing on sales, they rely on retailer EPOS data (electronic point of sale), which records what scans through the checkout each week, or on consumer panel data from providers like Kantar or NielsenIQ. Both have limitations. EPOS data shows what sold but not who bought it or what influenced them. Panel data shows purchasing behaviour but is based on a sample and has a time lag.
MMM sits above both. It takes whatever sales data is available, combines it with media spend, pricing, distribution, and promotional data, and uses statistical techniques to estimate each factor's contribution to volume and value.
The typical FMCG media mix
- TV: still the dominant brand-building channel for mass-market FMCG, with proven links to long-term market share
- Video on demand and streaming: growing share, particularly for household brands targeting families
- Digital video (YouTube, social): shorter formats, effective for younger demographics and new product launches
- Out-of-home: high reach, effective near point of purchase for impulse categories
- Retailer media: in-platform ads on Tesco, Sainsbury's, Boots, and Amazon, increasingly measurable via closed-loop retailer data
- Shopper marketing: in-store displays, end-of-aisle features, coupon redemptions
- Price promotion: deep discounts and multi-buys, the single largest driver of short-term volume for most FMCG brands
Why price promotion is the biggest variable
For most FMCG categories, price promotion drives more short-term volume than any media channel. A 25 percent price cut or a buy-one-get-one deal can triple weekly sales in the promoted retailers. This creates a serious measurement problem. If your model does not properly account for promotional depth and mechanics, it will attribute the promotional volume uplift to whatever media was running at the same time. This makes your media look artificially effective in weeks when promotions are running and ineffective in weeks when they are not.
Model promotional mechanics at a granular level. Do not just use a binary "promotion running yes/no" flag. Include the promotional depth (percentage discount or deal type) so the model can estimate the price elasticity accurately.
Seasonality and category factors
FMCG seasonality varies enormously by category. Soft drinks peak in summer. Hot chocolate peaks in winter. Baby products have their own demographic seasonality linked to birth rates nine months prior. Confectionery has multiple seasonal peaks: Easter, Halloween, Christmas. The model needs to capture these patterns, either through explicit seasonal variables or by using weekly dummy variables that the model learns from the data.
Category-level trends also matter. If the entire snacking category is growing because of post-pandemic changes in eating habits, your brand benefits from that tailwind. The model needs a category growth variable to prevent it from attributing category-driven growth to your media investment.
Distribution as a key driver
Gaining a new distribution listing, or losing one, has an enormous effect on FMCG sales that has nothing to do with marketing. If your brand gains a listing in Aldi or gets delisted from Tesco, your weekly volume changes immediately. The model needs weekly distribution data (number of stores selling the product, or weighted distribution score) so that distribution changes are properly separated from media effects.
Using MMM to defend TV investment
One of the most common uses of FMCG MMM is defending TV investment against finance teams who question its short-term ROI. TV's effects are often longer-term than those of promotion or digital, meaning they do not show up clearly in weekly sales data. A well-specified model with appropriate adstock decay rates can show the cumulative effect of TV over time, including the "long multiplier" effect that IPA research shows typically doubles TV's measured short-term impact when measured over a full year.
We sell through multiple retailers with different promotional calendars. How do we handle that?
Model at the retailer level rather than at a total brand level. This takes more data and more time, but it gives you retailer-specific estimates of media and promotional effectiveness. You can then aggregate back up to a total brand view while retaining the ability to plan at a retailer level.
We have just launched a new product. Can we run MMM on it?
Not until the product has at least 18 to 24 months of sales history. MMM needs historical variation to learn from. For new products, use controlled experiments (geo tests or split retailer trials) in the first year, then transition to MMM once you have enough data.
How do we value retailer media separately from our own media in the model?
Treat retailer media spend as a separate channel with its own adstock and diminishing returns curve. If you have sales data at the retailer level, you can even link retailer media spend directly to that retailer's sales, which gives you a much sharper read on its incremental contribution than modelling at a total market level.