Marketing mix modeling for multi-location businesses
In short
- A marketing mix model reads total ad spend against total results over time and splits the results by channel. It needs spend that moves on its own, years of history, and ads that bring in a real share of customers.
- Most chains of 5 to 50 locations change spend with the seasons and have a handful of channels. A model fed that data can give confident answers that are far off.
- What works at this size is what the industry calls unified measurement, or triangulation: read spend against all new clients, tracked attribution and what new clients say side by side, and run holdout tests where they differ. Each shows part of the picture; together they show the whole.
- Locations make good test groups. Pause or raise a channel at half of them and compare with the rest.
What a marketing mix model does
A marketing mix model, or MMM, is a regression on your history: weeks or months of spend by channel against sales or new customers, with seasonality and price taken out. It estimates what each channel added and how returns fall as spend grows. Google's Meridian, Meta's Robyn and PyMC-Marketing are free and open source, and large brands buy the same thing from vendors and agencies.
It needs three things to see clearly. Spend has to move for reasons other than demand, or the model can't tell the ads from the busy season. It needs enough history for those moves to repeat. And ads have to drive a large enough share of results to stand out from everything else. Google's documentation for Meridian has a page on how much data a model needs, and recommends calibrating it with experiments.
What we found on made-up chains
We ran Meridian on made-up chains of 6 to 14 locations with up to 143 weeks of history, where we set the true answer for every channel. None passed our bar: the answers were too uncertain to act on, or far off. When spend followed busy weeks, as it usually does, the model was confident and far off: it credited brand search with about 90 times its true return. Adding results from holdout tests fixed the channels those tests covered.
Fewer locations widened the ranges only a little. What decided what the model could see was how much spend moved on its own, and how big a share of new clients the ads brought in.
What works at this size
Three signals, each showing part of the picture, read side by side. Where two agree, plan on it. Where they differ, a test fills in the rest.
| Signal | What it sees | What it leaves out |
|---|---|---|
| Spend against all new clients | Every dollar and every new client, by month and location | Which channel did the work |
| Tracked attribution | Clicks, booking links and promo codes | Ads people only saw, and anything offline. Counts people already looking for you as brand search |
| How did you hear about us? | Word of mouth, billboards, events and ads people remember | Ads people forgot. Leans toward what they see every day |
| A holdout test | What one channel really adds | Small effects, at a small number of locations |
The survey answers are called self-reported attribution. They work best when people can tick every channel that played a part and then pick the one that decided, so credit can be shared. A person moved by a Meta ad, a Google search and a friend still counts as one new client.
A holdout test across locations
Pair your locations by size and flip a coin in each pair. Pause or raise one channel at the winners and leave the rest alone. The difference between the two groups is what the channel adds.
Before you start, measure how much the groups differ when nothing changes: run fake tests on your own history, splitting the locations the same way. A test is only worth running when its possible answers land further apart than those fake results. Write down the result that would move the budget, run it for the full length, and read it once.
Before you buy a model
- Count how often each channel's spend changed on its own, not with the season.
- Check what share of new clients your ads bring in. If it is small, a model has little to see.
- Ask for each channel's range, and what the model assumed before it saw your data.
- Ask how people heard about you the same way at every location, with every channel listed.
- Run one holdout test on the channel tracking loves most. It is usually brand search.
Want the three signals for your locations?
We read your ad spend, tracking and sign-up answers side by side for each location, and plan the tests that settle where they disagree.
See marketing measurement