Marketing mix modeling and in-store media

Where should screens in independent stores show up in a marketing mix model? As their own channel, with their own spend, delivery, and geography, rather than folded into a generic "other" line. Marketing mix modeling, or MMM, is the statistical practice of estimating each marketing input's contribution to sales from historical data, and it can only credit a channel it can see clearly.

MMM has aged well. As identifier-based attribution has weakened, models built on aggregate data have moved back to the center of how large brands allocate budgets, which makes it worth knowing how an in-store channel should feed one.

What does MMM actually do?

It works backward from outcomes. Feed the model a history of sales alongside a history of marketing activity and other factors like price, distribution, and seasonality, and it estimates how much each input contributed. The output is a set of effects a planner can use to reallocate budget toward what the model says earns it.

Its character is the opposite of click attribution: aggregate rather than individual, slow rather than instant, and strategic rather than tactical. No cookies, no shopper identity, just variance in the data explained as well as the inputs allow.

What inputs should an in-store channel supply?

The model is only as good as its feed, and in-store media can supply an unusually clean one. Spend by week. Delivery in plays and modeled impressions. Flighting dates. And geography, which is where this channel shines: campaigns on NRS Digital Media activate down to zip-code level, so the model can see not just when the channel was active but exactly where.

Geographic variation is fuel for MMM. A channel that ran everywhere at once is hard for a model to separate from everything else that happened at the same time. A channel that ran in some zip codes and not others hands the model a natural comparison, and the NRS footprint of 8,500+ zip codes leaves plenty of room to create that variation deliberately.

Clean input files sound mundane, but modelers spend much of their time repairing them. A channel that delivers tidy weekly data by zip earns better treatment in the model than one that arrives as a lump sum and a date range.

Why does granular sales data improve the model?

Because the outcome side of the equation matters as much as the input side. A model fed only national monthly sales has few observations to learn from. Store-level and region-level transaction data, of the kind generated across a network processing 1.9 billion transactions annually, multiplies the observations and lets the model detect effects that national aggregates would smooth away.

This is where a POS-anchored channel contributes twice. It's a media input to the model, and its scan-data analytics can sharpen the sales history the model learns from, particularly for the independent retail channel that national datasets often cover thinly.

What can't MMM see, and what pairs with it?

MMM estimates from history, so it struggles with channels that are new, small, or unvarying. A brand's first modest in-store flight may be invisible to a model trained on years of heavy national spend. That's not a verdict on the channel; it's a limitation of the method. Model results also arrive with a lag, usually read quarterly or annually, so they steer next year's allocation rather than this month's flight.

The pairing that works is MMM for the long view and store-level experiments for the sharp one. A test-versus-control read on a specific flight gives you fast evidence that approaches causal — as close as matched-store designs honestly get; the model then incorporates the channel's accumulating history over quarters. Brands that run both stop arguing about which is right and start using each where it's strong.