How do MMMs work? What happens between your data and your budget recommendation

A sound engineer can be handed a finished recording of a full band and asked a deceptively hard question: how much did each instrument contribute to what you're hearing? There's no separate track for the drums or the bass, only the final mix. So the engineer works backward, listening for patterns that reveal which instrument was doing what. It gets tricky fast. Two instruments playing the same notes at the same moment blur together, and the bass line changes how the guitar sounds even when the guitarist doesn't change a thing.

Marketing mix models do the same job with your revenue. (NOTE: Technically, MMM stands for marketing mix modeling, but since people often refer to marketing mix models as MMMs, we'll use that terminology here, too.) They start with the finished recording (your total sales) and work backward to estimate how much each channel, campaign, and outside force contributed. The way a model does that separating shapes every number it hands back, and those numbers shape your budget decisions. Marketing leaders who understand what happens between the data going in and the recommendation coming out are in a much better position to know when to trust a result and when to push on it.

Key takeaways

• Marketing mix models work backward from aggregated data, like total spend and total sales, to estimate what drove revenue, so they don't depend on user-level tracking.
• Traditional marketing mix modeling runs a regression on two to three years of historical data, with adjustments for how long ads keep working and how returns flatten as spend grows.
• Open source marketing mix models like Robyn, Meridian, and PyMC-Marketing made the software free, but they still need data scientists, clean data, and regular upkeep.
• An analyst's choices about carryover, saturation, and starting assumptions can change the results, which is why two marketing mix models built on the same data can disagree.
• Prescient's MMM uses machine learning to model marketing as a connected system, which is what makes campaign-level reads, daily updates, and halo effects possible. This means it does not run on a regression like traditional MMMs.
• The most useful question to ask about any marketing mix model is how its accuracy gets checked against the revenue that actually came in.

What every marketing mix model is trying to do

Marketing mix modeling (MMM) is a statistical method for estimating how much your marketing activities and outside forces each contributed to business outcomes like sales. It works from aggregated data, like daily or weekly spend by channel and total revenue, so it never needs to follow an individual shopper.

That second point explains why the method has moved back to the center of data-driven marketing. Privacy regulations and browser limits on third-party cookies have chipped away at user-level tracking, which multi-touch attribution depends on. Marketing mix models don't touch user-level data, so they keep working however those rules change. They can also measure marketing that clicks can't reach, including TV commercials, out-of-home, and other offline channels.

What goes into a marketing mix model

The math varies from one approach to the next, but every model needs the same broad categories of inputs.

The last row differentiates models quite a bit from one another. A model that only sees your website's sales can't give a campaign credit for the purchase someone made in a store a week later.

What comes out of a marketing mix model

Once the inputs are in place, you can expect a few standard model outputs:

• Contribution: The incremental sales each channel or campaign added, alongside base sales, which is what you would've sold with no marketing at all
• Efficiency: The return on your marketing investment in each channel
• Response curves: How revenue is expected to change as you spend more or less
• Forecasts: What's likely to happen under different budget allocation scenarios

Every approach to marketing mix modeling aims for these outputs. Where they split is in how they get there.

How traditional marketing mix modeling works, step by step

Traditional MMM dates back to the 1960s, and its core process hasn't changed much since. A consultancy or an in-house data science team usually builds it, and the entire process can take months. Here's what happens along the way.

1. Collect data. The team gathers two to three years of historical performance data, usually weekly, though some models still run on monthly data. Data collection tends to be the slowest step because advertising costs, pricing, and sales records live in different systems.
2. Adjust the media inputs. Raw advertising spend doesn't go into the model as is. First it's adjusted for adstock, also called carryover, because an ad keeps working after it runs. Then it's adjusted for saturation, because the model assumes returns flatten as spend climbs.
3. Fit the regression. Regression is a statistical technique that looks at how sales moved as each input moved, then gives each input a weight. The result is an equation: base sales plus a separate contribution from every channel.
4. Test the model. Data scientists compare the model's predictions to actual sales, ideally for a stretch of weeks the model never saw. That check on predictive accuracy is the first sign of whether the model can be trusted.
5. Break down the results. The model splits past revenue into base sales and the incremental sales tied to each channel.
6. Simulate budgets. The team runs scenarios to find the optimal allocation for a given marketing budget, and that feeds planning for future campaigns.

Where analyst judgment shapes the answer

Those six steps sound mechanical, but people make consequential choices at almost every one. Someone has to decide how quickly the effect of TV commercials fades, what shape each saturation curve takes, and which external factors make it into the model. In Bayesian versions of marketing mix modeling, analysts also set "priors," which are starting beliefs about each channel's performance that the model then updates with data. (Yes, this means there are assumptions baked into your MMM.)

None of this is cheating. It does mean two skilled teams can take the same performance data and produce different MMM results, because statistical rigor in the math can't remove the judgment calls in the setup. It's a big reason practitioners so often describe MMM estimates as "directionally correct."

Where the regression approach runs into trouble

Most of traditional MMM's limits come from how much it asks of a small dataset.

• Too many questions, too few rows. Three years of weekly data is only 156 data points. That's thin once you're trying to separate TV, radio, paid search, social, pricing, and seasonality.
• Channels that move together. If you raise marketing spend everywhere for the holidays—as most brands do—the model struggles to tell which channel did the work, or whether the season would've lifted sales anyway.
• Fixed effects. Most regression models assume a channel's marketing effectiveness stays the same across the whole time window, even though creative, competition, and platforms keep changing.
• Channels treated as soloists. The equation adds up separate contributions from different marketing channels, so it has a hard time showing how one channel lifts another.
• Slow-building brand effects. Some advertising effects build over months, and a simple decay setting tends to miss them. That can make direct response channels look stronger than brand campaigns by comparison.
• Coarse, slow reads. Results usually stop at the channel level and refresh a few times a year, which is hard to act on if you manage marketing performance week to week.

How open source marketing mix models work

Open source tools took the traditional approach to marketing mix modeling and packaged it as free code that any marketing analytics team can inspect. Meta released Robyn in 2021, and Google and PyMC Labs followed with packages of their own.

The foundation will look familiar: adstock, saturation, and a model that assigns credit to marketing channels. The tools mainly differ in the statistical models underneath, which fall into two camps.

• Frequentist: The model learns only from the data you give it.
• Bayesian: The model starts with priors and updates them with your data. It reports a range for each result, which gives you a read on statistical confidence. (We have a plain-language guide to Bayesian linear regression if you want to understand that more in depth.)

Here's how the three best-known open source marketing mix models compare.

What open source changes, and what it doesn't

Open source removed the license fee and made the math visible, which opened up marketing mix modeling to brands that could never have paid for a consultancy build. The work around the model—and most of the cost—is still yours, though:

• Expertise: You need data scientists who can set up, tune, and interpret the model.
• Data quality: Clean, consistent inputs still decide whether the output is usable.
• Upkeep: The model needs refreshing as your channels and marketing tactics change.
• Assumptions: The built-in ideas about carryover and saturation come with the package.

Teams implementing MMM this way often find the software was the cheap part. A data scientist in one practitioner forum described watching a company hand its model to an intern because leadership wanted a tool that would simply output an answer, and that's a fast route to confident, wrong budget decisions.

How Prescient's marketing mix model works

Prescient's MMM starts from a different premise. Where regression fits a single equation and adds up separate channel effects, Prescient uses machine learning to model marketing as a connected system that changes over time. That's closer to understanding how the band plays together than to measuring each instrument's volume.

That design shows up in what a marketing team actually sees:

• Campaign-level reads: Campaigns and tactics drive revenue differently, so you see performance for each one and not only a channel average.
• Daily updates: The models refresh every day, so your view keeps pace with the decisions you're making.
• Learned decay: How long a campaign's effect lasts is learned from your own data, campaign by campaign.
• Saturation that isn't assumed: The model doesn't push every campaign onto the same flattening curve, so it can surface more than one point of efficient spend.
• Efficiency that moves: A campaign's effectiveness can shift with seasonality, events, and other external factors, and the model accounts for that.
• Halo effects: This is revenue a campaign drives indirectly through another channel, such as when a paid social ad leads to a later purchase through branded search, on Amazon, or in a retail store. It's estimated from patterns in spend and sales, with no individual tracking involved.

Prescient also splits each campaign's revenue into halo effects and base revenue, which comes from direct engagement with the ad. One note on vocabulary: base revenue isn't the same thing as the base sales in a traditional model.

How the model earns your trust

Any model deserves skepticism until it proves itself, so the checks are built into the platform. Forecasts come with a confidence band—the shaded range of likely outcomes—and a separate model confidence score. If you've run an incrementality test, Prescient shows you the model's accuracy with and without that test included, and you decide whether to use it.

From there, the Media Forecaster helps with budget allocation by showing what's likely to happen if you scale a channel's spend or shift budget into it from other channels.

Three approaches to marketing mix modeling at a glance

If you're skimming, this table sums up how the three approaches differ.

Common misconceptions about marketing mix modeling

A few beliefs come up again and again in practitioner forums, and each one can lead to expensive budget allocation mistakes.

• "The model tells you where to spend, and you just follow it." A recommended split is only as good as the assumptions behind it, and it's least reliable at spend levels you've never tried. Treat it as one input to strategic budget allocation, alongside what your marketing team knows about the business.
• "More history always makes a better model." Extra years do add data points. They also describe a business with different products, creative, and marketing strategies than the one you're running today.
• "You build it once." Implementing MMM is the start of the work. Models drift as you add marketing channels, change promotions, or rethink your media mix, and one that was built last year and never refreshed is describing last year.
• "An incrementality test always makes the model more accurate." A test captures one channel, in one place, during one window. That makes it accurate locally, but the result may not hold across the rest of your marketing activities, so check whether adding it actually improves accuracy.

Questions to ask about how any MMM works

Marketing leaders don't need to follow the math to evaluate a marketing mix model. Whether you're comparing MMM vendors or sizing up one your own team built, a handful of questions will show you how it really works.

• What does it assume about saturation, and how would you know if that assumption were wrong for a campaign?
• How detailed are the results: channel, tactic, or campaign?
• How often does it refresh?
• How is accuracy checked against actual revenue?
• What happens when a test result and the model disagree?
• Does it see sales beyond your website, like Amazon and retail?

Clear answers tell you whether an MMM analysis can turn into actionable insights and data-driven decisions, or whether it'll mostly produce charts.

Where Prescient comes in

Prescient built its MMM for omnichannel brands that need marketing measurement they can act on every day. The model works at the campaign level, updates daily, and measures halo effects across your website, Amazon, and retail. That means you can see what each campaign really contributes to your business outcomes before you move any of your marketing budget.

You also don't have to take the model's word for it when a marketing investment is on the line. Confidence indicators show how much weight a forecast can bear, and you can see for yourself whether an incrementality test makes the model more or less accurate. Book a demo to see how it works and all the features the platform has to offer.

FAQs

How much data do you need for marketing mix modeling?

Most traditional approaches to marketing mix modeling ask for two to three years of historical data so the model can see more than one full seasonal cycle. Variety matters as much as volume, though. A model learns from change, so past data where marketing spend barely moved teaches it very little, and daily data gives it far more to work with than weekly or monthly data. Requirements also differ by approach, so ask any provider what its minimum is and why.

How accurate is marketing mix modeling?

It depends on the model and on how its accuracy is tested. Marketing mix models can't be graded against a perfect answer key, because nobody can observe exactly what sales would've been without a given ad. What you can check is how well a model predicts revenue it hasn't seen yet, and whether the MMM results stay stable from one refresh to the next. A model that passes both tests can support real budget decisions, and one that's never been tested shouldn't be trusted with them.

What's the difference between marketing mix modeling and media mix modeling?

People often use the two terms interchangeably, and both get shortened to MMM. Strictly speaking, marketing mix modeling (MMM) covers everything that shapes business outcomes, including price, promotions, and distribution, while media mix modeling focuses on paid marketing channels. Media-only models use fewer variables, so they're usually faster to build and refresh. The trade-off is that they can miss the effect of something like a price change.

Can you build a marketing mix model in-house?

Yes, and open source tools have made it far more realistic. Building an MMM in-house still takes data scientists who understand the method, engineers to keep the data flowing, and time to maintain the model as the business changes. The software is free, but the people aren't. Brands without that bench usually reach useful marketing analytics faster with a platform.

Do marketing mix models use machine learning?

Some do. Traditional marketing mix modeling is built on regression, a classical statistical technique, although newer tools use machine learning to automate parts of the setup. Other approaches, including Prescient's, use machine learning inside the model itself to learn patterns such as how long each campaign's effect lasts. The label covers a wide range, so it's worth asking where machine learning sits in any model you're evaluating.

What's the difference between MMM and multi-touch attribution?

Multi-touch attribution follows individual users across touchpoints and splits the credit for each conversion among them. Marketing mix models work from totals and never need user-level data. That leaves multi-touch attribution more exposed to privacy regulations and blind to marketing channels without clicks, like TV, along with the brand effects those channels build over time. Many performance marketing teams grew up on click-based reporting, and they turn to marketing mix modeling when those reports stop adding up.