What Is Marketing Mix Modeling?

How Does a Marketing Mix Model Work?

A marketing mix model (MMM, sometimes called a media mix model or market mix model) lines up a history of sales against a history of spend in each channel, plus the other things that move sales, and uses regression to estimate how much of each week's sales each one explains. It runs on aggregate totals, so in the words of Meta's Robyn documentation, "MMM does not need user level data and instead it runs off aggregated data".¹

Google's Meridian documentation lists the inputs as your target KPI (often sales), media data for each channel, and any relevant control variables.²

• The outcome: sales, revenue or signed customers, period by period.
• Marketing: spend (and often impressions) for each channel in the same periods, including offline media such as TV and print.
• Everything else that moves sales: seasonality, price changes, promotions, competitor activity and the economy, so the model doesn't credit marketing for what the calendar did.

Two adjustments bring the model closer to how advertising behaves: adstock reflects that "the effects of advertising can lag and decay following an initial exposure", and saturation reflects that "each additional unit of advertising exposure increases the response, but at a declining rate".¹ The model estimates totals by channel, while the CRM record the Channel Selection step builds follows each customer from first contact to signed deal.

How Do You Do Marketing Mix Modeling?

To do marketing mix modeling, decide what the model has to answer, gather a long, consistent history of sales, spend and outside factors, fit and check the model, calibrate it against experiments, and then refresh it as new data arrives. In our view, how long it takes depends mostly on how quickly that history can be gathered and cleaned, more than on the modeling itself.

1. Write down the budget question the model has to answer, such as how to split next year's spend, because that decides which channels and how much detail you need.
2. Collect the history: Meta's guide for its Robyn tool recommends a minimum of two years of weekly data, and more than two years (for example four to five) if only monthly data exists.¹
3. Check that spend varied, because a channel whose spend never changed gives the model nothing to compare.
4. Keep the inputs in proportion to the data: the same guide recommends about one input for every ten observations.¹
5. Fit the model with a specialist or an open-source package. Google's Meridian is an open-source MMM that Google says is designed to estimate the causal impact of marketing,² and Meta's Robyn is an open-source MMM package that Meta describes as "especially suitable for digital and direct response advertisers with rich data sources".³
6. Calibrate against experiments: Meta "strongly" recommends using experimental results as ground truth to calibrate an MMM,¹ and Meridian lets you add experiment results as starting assumptions (priors).² The tests themselves are covered in how to do incrementality testing.
7. Refresh it on a schedule, because a model fitted to last year describes last year's prices, competitors and channels.

How Do You Use a Marketing Mix Model to Set a Budget?

Use a marketing mix model to set a budget by reading each channel's response curve and moving money toward the channel where the next dollar still buys the most, rather than toward the channel with the best average return. Meridian, for example, reports each channel's ROI and contribution, a response curve for each, and budget optimization recommendations.²

• Contribution: the share of sales the model attributes to each channel and to the baseline (sales you'd have had with no marketing at all, from reputation, repeat buyers and seasonality).
• ROI by channel: contribution divided by spend, with a range around it (Meridian's credible intervals) that shows how uncertain each figure is.²
• Response curves: how sales change as spend in one channel rises or falls, and where each curve flattens as the channel saturates.

The curve matters more than the average because a channel with a high average return can already be close to flat, so the next dollar buys little, which is the same added-customer logic the Marketing Action Plan step uses with a CRM. Treat the optimizer's answer as a proposal, and confirm it one channel and one spend step at a time before moving the whole budget.

Is Marketing Mix Modeling Worth It for a B2B Company With Few Deals?

Marketing mix modeling is rarely worth it, in our view, for a business that signs a handful of customers a week or fewer through a few channels, because the model has too few outcomes to separate one channel from another or from chance, while CRM tracking and controlled tests answer the same budget questions for much less.

The businesses MMM suits, in our experience, sell in volume every week, spend across many channels including offline media, and have years of history with real swings in spend. For a company with fewer deals, the routes in How Do You Measure Marketing Effectiveness? are usually the cheaper place to start, and Which Attribution Model Should You Use? explains why switching attribution models doesn’t settle the same question.

What Are the Limits of Marketing Mix Modeling?

The main limit of marketing mix modeling is that it finds patterns in history, so its estimates are only as good as the variation in that history and the factors it was given. Meta's own guide notes that regression "looks at the correlation between the dependent and independent variables", which is why it pairs MMM with experiments.¹

• Spend that moved with the season: if spend always rose in the busy months, the model can't fully separate the season from the spending.
• Too little variation: channels with steady spend, or channels that always moved together, produce estimates with wide ranges.
• Assumed lag: adstock settings decide how long an ad keeps working, and with long sales cycles a small change in that assumption can move a lot of credit.
• The past range only: response curves are estimated from the spend levels you ran, so predictions far beyond them are guesses.
• Totals, not customers: it can rank channels but can't say which customer types, campaigns or offers produced profit unless the data is split that finely, and splitting it thins the data further.
• Sales rather than profit: a model of sales doesn't know the margin or lifetime value of the customers each channel brought, so its ROI needs translating into gross profit before it moves money.

Treat the model as one estimate among three, next to your CRM record and your experiments, and look hardest at the channels where the three disagree.