Marketing mix modeling is the oldest of the three ways to measure marketing, and the one that needs no tracking. It works from weekly totals, so limits on tracking leave it intact. Meta calls it privacy-friendly and resilient to signal loss (Meta Robyn). Google and Meta now both publish open-source versions.

What is marketing mix modeling?

Marketing mix modeling, or MMM, is a statistical model of what each kind of marketing adds to sales. Meta’s Robyn guide calls it a model of the extra sales that marketing and other drivers bring in (Meta Robyn). It covers paid ads, owned channels, price and deals. It also covers outside forces, such as the season and holidays.

The name comes from Neil Borden of Harvard Business School. He got the idea from a colleague, James Culliton. In 1948, Culliton called a business leader a “mixer of ingredients.” Borden then began to use the phrase “marketing mix” in his teaching (Borden).

How does a marketing mix model work?

An MMM relates a time series of sales to time series of spend and other drivers. Plain regression misses two facts about advertising. Its effect lingers after the ad runs, and each extra dollar buys less than the one before. Google researchers built both into a Bayesian model in 2017 (Jin et al., 2017).

The first fact is called adstock. Robyn describes it as the theory that ad effects can lag and decay after the first exposure. The second is saturation: each added unit of ads raises response, but at a falling rate (Meta Robyn). With a 50% weekly decay, a $100,000 burst in week one carries $50,000 of effective pressure into week two and $25,000 into week three.

The output is a set of response curves, one per channel. From them, the model reports each channel’s return on spend and its marginal return, the payoff on the next dollar. Google’s Meridian adds budget plans built on those curves (Google Meridian).

Bayesian models like Meridian start from prior beliefs about each channel and update them with the data. They report a range around each return, which Meridian calls a credible interval (Google Meridian). A narrow range means the data pinned the return down. A wide one means it did not.

What data does a marketing mix model need?

Robyn says a sound model needs at least two years of weekly data (Meta Robyn). Google’s Meridian guide shows why. A national model with 12 media channels, six controls and eight time knots has 26 parameters. Two years of weekly data gives 104 data points, or four per parameter, which is too few. Cut to 10 parameters over three years, and 156 data points give about 15 per parameter (Google Meridian).

Data by region helps. Meridian can model each region on its own, and Google says this data holds more information (Google Meridian). Its guide counts 105 regions over three years of weeks as 16,380 data points (Google Meridian). A program with members in many markets can build its weekly series by region, which adds up fast.

InputExamplesThe loyalty program version
Goal (KPI)Weekly sales or sign-upsMember sales, new members or active members
Paid mediaSpend and impressions by channelAds that drive enrollment or visits
Organic mediaActivity with no direct media costMember emails, push and app messages
Non-media treatmentsPrice changes and promotionsBonus points events, tier changes, earn rate changes
ControlsSeason, holidays, the wider economySchool holidays and peak travel weeks
Experiment resultsLift tests used to calibrate the modelHoldout tests on member offers

Meridian treats owned activity with no media cost as organic media. It treats promotions and price as non-media treatments, which have “calculable incremental impacts” (Google Meridian). Spend also has to vary. Robyn warns that if TV spend stays flat while sales move week to week, the model has trouble seeing what TV did (Meta Robyn).

How is marketing mix modeling different from multi-touch attribution?

Multi-touch attribution tracks each person’s clicks and views. MMM works from totals. Robyn puts it plainly: attribution relies on online signals, while MMM “does not need user level data” and runs on aggregated data (Meta Robyn). So MMM can measure TV, print, radio and store sales that leave no click. It also holds up as browsers and phones restrict tracking.

The cost is detail and speed. MMM reads whole channels. It cannot rank single campaigns or ads, and it needs years of history. Attribution answers daily questions inside digital channels. Running both, and settling disputes with experiments, covers each method’s blind spot.

How do you know if a marketing mix model is right?

Check it against tests. Robyn strongly urges teams to use test results as the ground truth to tune the model (Meta Robyn). Meridian turns lift test results into the model’s starting beliefs, called priors (Google Meridian). If a model’s return for a channel clashes with a clean holdout test, look again.

Read the range around each estimate as well as the estimate. In the 2017 Google paper, priors had a big impact when the sample was small, and the best media mix had a large variance (Jin et al., 2017). A wide range on a channel’s return means the data cannot yet tell its true value. Our incrementality entry covers the tests used to narrow it.

How do loyalty programs use marketing mix modeling?

Program activity belongs in the model as drivers. Leave it out and the model folds it into the baseline, the sales it credits to no marketing at all. That baseline can be most of the business. Ulta Beauty reported more than 46 million members and about 95% of total sales coming from members in fiscal 2025 (Ulta Beauty 10-K). In a business like that, member behavior is the sales series.

Pick the goal to match the decision. A model of total sales weighs media against the program. A model of new members shows what drives sign-ups. A model of active members shows what keeps members buying.

Feed the model the program’s levers week by week: member messages, bonus points events and earn rate changes. Then check its answers with a control group test on the largest lever. The loyalty program ROI guide turns the result into profit.

For illustration, the model finds that paid social adds $300,000 of sales on $100,000 a quarter, and $450,000 on $200,000. At $200,000, the average return is 2.25. The second $100,000 returned only 1.5. If the model gives a bonus points event a marginal return of 2.0, the next dollar earns more there.

How GRAVTY handles marketing mix modeling

An MMM is only as good as its weekly inputs. GRAVTY logs each earn, burn, tier change and offer response by member, so program levers can be summed by week and market. On lji.io, Agentic AI Compass brings KPIs, partner data, customer sentiment and market signals into one place. AgenticTest measures lift by segment. Lift results like these are what Robyn and Meridian use to tune a model.

Frequently asked questions

How much data do you need for marketing mix modeling?

At least two years of weekly data, by Meta’s Robyn guidance (Meta Robyn). The real test is data points per parameter. Google’s Meridian guide counts 104 weekly points against 26 parameters as too few, at four per parameter (Google Meridian). Fewer channels or regional data raise the count.

What are adstock and saturation in MMM?

Adstock is the carryover of an ad’s effect into later weeks. Saturation is the drop in return as spend rises. Robyn says ad effects can lag and decay after the first exposure. It also says each added unit of ads raises response at a falling rate (Meta Robyn).

Are Google Meridian and Meta Robyn free?

Yes. Google describes Meridian as an open-source MMM that is free for anyone to use, with code you can read and change (Google Meridian). Meta’s Marketing Science team built Robyn as open-source R code for in-house modelers (Meta Robyn). Both still need a skilled analyst and clean data.

How often should a marketing mix model be refreshed?

Refresh it each time a useful block of new data arrives, and rebuild it when most of the data is new. Robyn’s refresh function adds new weeks to a chosen model. Its guide says a full rebuild is better when, say, 80 new weeks join a 100-week model, or when new variables are added (Meta Robyn).

Can marketing mix modeling measure a loyalty program?

Yes, if the program’s levers go into the model as their own variables. Bonus events, earn rate changes and member messages each need a weekly series. The model can then estimate what each added. The program’s total value is better measured with a long-run holdout of members.