Marketing mix modeling vs incrementality testing for multi-location brands

Marketing mix modeling and incrementality testing answer different questions. MMM estimates how every channel, plus price, promotions and the season, contributed to sales over the last couple of years, so you can plan next year’s split. An incrementality test changes spend in some markets and not others, then measures what that change added at your locations. One is a map of the whole territory. The other is a survey stake driven into one spot.

MMM is back in fashion, helped by free open-source code from Google (Meridian) and Meta (Robyn). That’s useful. It also means plenty of multi-location brands are about to buy a model that can’t see their locations. This guide covers what each method does, where each breaks for a brand with dozens or hundreds of locations, a worked example where they disagree, and how to run the two together.

What marketing mix modeling does (and what it needs)

MMM is a regression on time series. You give it weekly sales and weekly spend by channel for two to three years, plus the other things that move sales: price changes, promotions, holidays, weather, store count. It estimates how much each channel contributed, how long its effect lingers after the spend stops (adstock, in the jargon), and where extra spend starts to buy less. The output is a set of response curves and a recommended budget split.

The data needs are real. Google’s Meridian documentation recommends at least two years of weekly data for a model built on regional data and three years for a national one (Source: Google Meridian documentation, 2025). The model also needs your spend to have moved. If paid social has been $30,000 a month every month for two years, there’s nothing for the model to learn from, and it falls back on whatever assumptions you gave it.

• Every channel at once. Including the ones you can’t switch off by market, like national TV, sponsorships and brand search.
• Media versus everything else. It separates ads from price, promotions and seasonality, if you’ve logged them.
• The annual split. Roughly how much should go to each channel next year.

Where MMM struggles for multi-location brands

• Channels move together. You raise search, social and radio in the same spring weeks, so the model can’t tell which one did the work. It splits the credit by assumption.
• Spend follows demand. You spend more in peak weeks because they’re peak weeks. Unless the model controls for the season well, it credits media for the weather.
• It answers at brand or region level. Even a model built on regional data rarely gives a trustworthy answer per location, and your operators manage to locations.
• It’s slow. Most teams refresh quarterly or twice a year. A channel that stopped working in May shows up in the October readout.
• The ranges are wide. And they often fall off the slide before it reaches the budget meeting.

What incrementality testing does

A test is an experiment. Pick some markets, change spend there (switch a channel off, add one, double the budget), keep comparable markets steady, and compare customers at your locations over the same weeks. Measure the gap against a weighted blend of the steady markets. Because the only planned difference between the groups is the spend change, the gap was caused by the spend, not just seen alongside it.

Tests are narrow by design: one channel, one spend level, one period. They don’t tell you how to split $4 million across eight channels. They tell you, with a range, whether $8,000 a week more of paid social in these markets added customers, and what each one cost. For the mechanics, see our guide to geo holdout tests.

The costs are different too. A test isn’t free: holdout markets give up the channel for a while, or test markets get extra spend that may not pay. It answers one question at a time, and the answer ages as creative, prices and competitors change. But it needs weeks of data rather than years, and you can explain it to a CFO in one sentence.

MMM vs incrementality testing, side by side

The third column is the one most brands forget. A multi-location brand changes spend in different markets all the time: a new store opens, a region gets a radio flight, a franchisee tops up the local budget. Each change is a small natural experiment. Comparing those markets with a weighted twin of the steady ones gives you a test-like read without designing a test. It’s less clean, because nobody randomised who got the change, but it’s continuous and it’s per location. Our synthetic control guide explains the method.

The question MMM answers badly: what will the next dollar buy?

MMM reports an average return for each channel: all the extra customers it credits to paid social, divided by all the paid social spend. Budget decisions are about the next dollar. On any channel that saturates, the next dollar buys less than the average one did. And the model’s curves are least certain exactly where you want to move: beyond the spend levels you’ve run before.

A heavy-up test measures the next dollar directly. Run the channel at the new level in some markets, keep the rest at today’s level, and the lift you see is the return on the extra spend. That’s why a model and a test so often seem to disagree. They’re answering different questions.

A worked example: when the model and the test disagree

This example is illustrative. A 36-location casual-dining group spends about $1.3 million a year on media. Its annual MMM, built by an agency, credits paid social with 11,600 extra guests on $360,000 of spend: $31 per extra guest, with a range of $14 to $90. The agency recommends adding $100,000 to social next year.

Before signing off, the marketing lead runs an eight-week heavy-up: $8,000 a week of extra paid social in six markets holding nine restaurants, with 15 steady markets as the comparison.

The two numbers aren’t in conflict. The model’s $31 is an average over all social spend, cheap first dollars included. The test’s $49 is what the next $8,000 a week bought. Both can be true.

The decision changes, though. Say a guest is worth about $11 in contribution margin per visit and a new guest comes back three more times in the year: roughly $44 in all. At $31, social clearly pays. At $49, the extra $100,000 buys about 2,040 guests worth about $90,000, a loss of roughly $10,000. Across the test’s range the result runs from a $33,000 gain to a $52,000 loss.

So the group holds social where it is, puts the test result into next year’s model, and uses the $100,000 to test a channel it knows less about. The model set the agenda. The test made the call.

How to use MMM and incrementality testing together

1. Start with the spend changes you already made. Read the markets where spend changed over the last year against twins of the steady ones. It costs nothing and usually turns up the obvious wins and the obvious duds. (This is the read Countlift runs every week from your sales and spend data.)
2. Test the big, uncertain channels. The ones that carry the most budget and where the model’s range is widest. One or two planned tests a quarter is plenty for most brands.
3. Feed the tests into the model. Both Meridian and Robyn can take experiment results as an input, so the model’s estimate for a tested channel is anchored to what the test measured instead of floating free.
4. Use the model for the split and tests for the moves. The model sets each channel’s rough annual envelope. Tests and market reads decide where inside that envelope the money goes, and when to pull it.
5. Re-test what matters. Re-run a big channel’s test about once a year, or after a major change of creative, offer or agency.

Which should a multi-location brand start with?

For most multi-location brands, testing. Three reasons.

• You already have the raw material. Spend varies by market and you count customers by location. A single-site online brand with one national budget has neither.
• Tests are easier to defend. “We switched radio off in five markets and guests there fell 2% against their twin” survives a CFO meeting. “The model attributes 14% of sales to radio” starts a twenty-minute argument about assumptions.
• MMM needs history you may not have clean. Two years of weekly data, with openings, closures, menu changes and price moves all logged. For most brands the log is the hard part.

MMM earns its keep once media is spread across five or more channels, some of which you can’t switch off by geography, and the budget runs to millions. Below roughly $1 million a year of media, the model’s ranges are usually too wide to change a decision. We’d spend that money on tests.

Common mistakes when choosing between MMM and testing

• Treating the model’s point estimate as a fact. Ask for the range on every channel. If a channel’s range runs from break-even to brilliant, the model hasn’t decided, and neither should you.
• Using an average return for a next-dollar decision. Covered above. It’s the most expensive mistake on this list, because it always errs toward spending more.
• Building the model on platform conversions. The outcome has to be what your locations count: guests, members, patients, jobs. Platform-reported conversions bake in the overcounting you were trying to escape. More on that in why click attribution misses in-store sales.
• Running tests too small to read. A test that can only see a 6% lift, when you expect 2%, will come back “inconclusive”. Size it first with the minimum detectable effect.
• Letting the channel’s seller grade the channel. Platform lift studies are fine for comparing creative inside the platform. They’re not the place to decide whether the platform gets next year’s budget.
• Averaging away a disagreement. When a test and the model disagree, that’s the most useful thing you’ll learn all year. Find out why before you split the difference.

Questions people ask about MMM and incrementality testing

Models are good at the big picture and bad at the next dollar. Tests are the reverse. Plan with one, move money with the other.

Written by The Countlift team at First Byte in The Woodlands, Texas. Questions or corrections? Email hello@countlift.com.