# Marketing mix modeling: what it is, tools and real costs

URL: https://3ammarketer.com/blog/marketing-mix-modeling-what-it-is-tools-costs
Published: 2026-10-18

> Marketing mix modeling estimates what each channel adds to sales from aggregated data. The free tools cost nothing to license; the data and the weeks of work are the real price.

Marketing mix modeling (MMM) is a statistical analysis that estimates how much each marketing channel, and each non-marketing driver, adds to sales or another KPI. It works from aggregated time-series data rather than user-level tracking, which is why it has come back into use as cookies and online signals fade. Meta describes it as a privacy-friendly, highly resilient, data-driven statistical analysis that quantifies the incremental sales impact and ROI of marketing and non-marketing activities. [2]

The catch is the input. Meta's guide to its open-source tool says a robust model needs a minimum of two years of historical weekly data, and Google's Meridian documentation shows that two years of weekly data gives only 104 data points to estimate every channel, control and time effect. That limit, more than any software licence, decides whether MMM is worth running for your business. [2][4]

## How marketing mix modeling works

At its core, an MMM is regression applied to business data. PyMC-Marketing's guide puts it plainly: the model uses the level of spend for each media channel over time, plus control measurements that capture seasonality or economic indicators, to estimate the impact of each on a metric such as new customers per week. The basic version is linear regression, with one coefficient per predictor. [3]

Real models add two non-linear effects. Saturation captures diminishing returns: the first money spent on a channel can have a big impact, while further spend returns less as people get used to the message. Adstock captures carryover: the effect of advertising can lag and decay after the first exposure, because people sometimes delay action until the following weeks. Both shape the response curves that budget recommendations come from. [3][2]

Google frames the output as answers to three questions: the historical ROI and contribution of each channel, the response curve for each channel, and how to allocate future budget to maximise the business outcome. Most modern tools then run a budget optimiser on top of the fitted curves. [1]

Robyn sorts inputs into three types: paid media variables with a clear marketing spend, organic variables, and context variables such as promotions, holidays or macroeconomic factors. Meta's guide suggests mapping every variable you want to include before collecting anything, because that list becomes the project plan. Common context metrics in the guide include GDP growth, unemployment and inflation, plus dummy variables for each type of promotion. [2]

## When MMM beats attribution, and when it does not

Multi-touch attribution follows individual users across touchpoints, so it depends on online signals that privacy changes keep removing. Meta's guide contrasts the two directly: MMM does not need user level data and runs off aggregated data, while multi-touch attribution is very reliant on online signals. That also means MMM can cover offline channels such as TV, radio or direct mail alongside digital spend. [2][3]

The trade-off is granularity. Google's Meridian works at channel level only, and its documentation does not recommend running it at campaign level because MMM is a macro tool. For more granular insights on digital channels, Google recommends data-driven multi-touch attribution instead. In practice the two answer different questions: MMM sets the budget split across channels, and attribution or lift tests guide decisions inside a channel. [4]

Meta's guide adds that MMM might not be the best solution for every question; some are better answered by an incrementality study such as Conversion or Brand Lift, or by attribution. Write down the business questions first, then map each one to a method. The output is a channel split, so it pairs naturally with planning work such as [B2B marketing budget allocation benchmarks](/blog/b2b-marketing-budget-allocation-benchmarks). [2]

## Meridian, Robyn and PyMC-Marketing compared

Three open-source frameworks dominate do-it-yourself MMM. Google's Meridian is free for anyone to use and its code is fully visible and available to modify. Meta's Robyn is open-source R code from the Meta Marketing Science team, built for in-house and DIY modelers. PyMC-Marketing is a Python library built on Bayesian modeling, which lets analysts add prior knowledge and calibrate models through lift tests. [1][2][3]


| Tool | Built by | Approach | Notable feature |
| --- | --- | --- | --- |
| Meridian | Google | Bayesian, causal inference | Geo-level data, reach and frequency inputs |
| Robyn | Meta Marketing Science | Open-source R code, machine learning | Model window over a longer data set |
| PyMC-Marketing | PyMC Labs | Bayesian, in Python | Priors and calibration through lift tests |

Meridian's distinctive features are geo-level modeling, which contains more statistical information per business, and optional reach and frequency inputs. Robyn's model window lets you fit on a recent subset of data, such as 100 weeks out of a 208 week data set, while still deriving trend and seasonality from the full history. For a deeper look at how the Bayesian frameworks and Robyn differ on solvers and calibration, see [Robyn's solvers and calibrating with lift tests](#robyn-solvers-and-calibrating-with-lift-tests) below. [1][2]

## How much data an MMM needs

Google's own worked example shows why national models struggle. With 12 media channels, six controls and eight knots, the model has 26 parameters; two years of weekly data gives 104 data points, or four data points per parameter, which Google calls too low to estimate the model reliably. Cutting to three channels and two knots, dropping a control and using three years of data gives 156 data points for 10 parameters, roughly 15 data points per parameter, enough for directional information. [4]

Robyn's guide gives a rule of thumb: a recommended ratio of 1 independent variable to 10 observations. If only monthly data exists, it strongly recommends collecting more than two years, for example 4-5 years. Geo data is the other way out: Google's example of 105 geos and three years of weekly data yields 16,380 data points, because Meridian shares information across geos. [2][4]

Google is candid that these counts are rough and directional. The data points are not independent, so 1,000 data points in an MMM setting are not the same as 1,000 coin flips or 1,000 randomly assigned participants in an experiment. Google also notes that insufficient variation in media spend adversely impacts national models. [4]

- Combine small channels or drop a channel with low spend before adding controls.
- Keep true confounders; remove controls that explain the KPI but not the media.
- Add geo granularity if your sales and spend data can be split by region. [4]

## What MMM costs in money and time

Meta's guide lists three ways to run an MMM. A third party vendor runs the whole engagement and is usually the most expensive option. SaaS platforms let you build models in a point and click environment and are a cheaper alternative to a vendor. An in-house build has no ongoing costs to run aside from headcount and time, but requires analytical or data science skills and takes the most internal time of the three. [2]

The in-house route is the same decision as any other build or buy question in marketing, and the hourly trade-offs in our look at [in-house vs agency marketing costs](/blog/in-house-vs-agency-marketing-cost-hourly) apply here too. What changes with MMM is that the in-house team needs statistical skills, and the vendor route can add benchmarks against other clients, which Meta lists as a vendor advantage depending on the vendor. [2]

The time is substantial even with free software. Meta's guide sets out the phases of an MMM study and how long each typically takes, and data collection alone needs at least 4 weeks because data rarely arrives in one format from every source. [2]


| Phase | Typical time |
| --- | --- |
| Business questions | 1-2 weeks |
| Data collection | 4-6 weeks |
| Data review | 1-2 weeks |
| Modeling | 4-8 weeks (dependent on MMM scope) |
| Analysis and recommendations | 2-4 weeks |

If you use a language model to help write the modeling code or summarise the outputs for stakeholders, that cost is small next to analyst time. On our model pricing table, Anthropic's Claude Sonnet 5 lists at $2 per million input tokens and $10 per million output tokens. [5]

## Adstock decay rates and saturation curves

In geometric adstock formulations, current carryover equals current-period spend plus a decay parameter lambda multiplied by previous adstock. In traditional models, adstock rates fall along a scale from 0% to approximately 95%, describing second-week impact relative to initial media delivery. [6][7]

Adstock decay rates alter media half-life timelines nonlinearly. An adstock rate of 50 percent translates to a half-life of 1 week, a 70 percent rate equates to a two-week half-life, and a 90 percent rate extends to nearly seven weeks. [7]

Empirical carryover rates vary significantly across channels: TV and video typically exhibit adstock decay rates between 0.4 and 0.8, whereas paid search ranges between 0.1 and 0.4. [8]

Saturation is commonly represented using Hill functions with the form Effect = Spend^alpha / (K^alpha + Spend^alpha), where alpha controls curve steepness and K represents the half-saturation point. Due to diminishing returns, marginal ROI is typically lower than average ROI at current spend levels. [8]

## Robyn's solvers and calibrating with lift tests

Meta Robyn is an experimental open-source MMM package developed by Meta Marketing Science. Rather than applying full Bayesian MCMC sampling, Robyn utilizes ridge regression to regularize multicollinearity and avoid overfitting across media channels. [9]

Robyn automatically decomposes trend, seasonality, and holiday patterns by embedding the Prophet library into the pipeline. It uses evolutionary algorithms from the Nevergrad AI library to conduct automated hyperparameter optimization across adstock and saturation ranges. [9]

Budget reallocation in Robyn relies on a gradient-based constrained non-linear solver. The solver optimizes outcomes across media channels within designated budget constraints. [9]

Model calibration leverages empirical evidence such as geo lift tests, conversion lift studies, and A/B tests to supply informative priors rather than non-informative priors. Empirical lift tests serve as a calibration check to anchor marketing mix models and ensure saturation curve estimations remain reliable despite unobserved confounders. [10][6]

Google Meridian provides methodologies to support calibration directly using experiments alongside other prior information. Meta Robyn incorporates ground-truth experiments through multi-objective optimization via Nevergrad, minimizing the error between MMM predictions and validation studies such as geo-based tests or Facebook lift. [11][9]

Limit: None of the cited sources publish vendor or SaaS MMM prices, so the cost comparison here is relative, not in dollars. The data thresholds are Google's and Meta's guidance for their own tools; the right amount for your business depends on how much your spend varies, and Google says the most accurate check is to fit the model and look at the width of the credible intervals.

## Questions

**What is marketing mix modeling in simple terms?**
It is regression on business data: it relates weekly sales or another KPI to spend in each channel plus controls such as seasonality, then estimates what each channel contributed. Because it uses aggregated data, it needs no cookies or user tracking.

**What is the difference between MMM and multi-touch attribution?**
MMM models aggregated weekly data across all channels, online and offline, and needs no user tracking. Multi-touch attribution follows individual users across digital touchpoints, so it relies on online signals but can work at campaign level, where Google says MMM does not belong.

**Is marketing mix modeling software free?**
The leading open-source frameworks are. Google's Meridian is free for anyone to use, Meta's Robyn is open-source R code, and PyMC-Marketing is an open-source Python library. The real cost is the analyst time and the data work, which Meta's guide says takes the most internal time of any route.

**Can a small business use MMM?**
Only if it has enough history and enough variation in spend. Meta's guide asks for at least two years of weekly data, and Google warns that a national model with many channels on that much data is too thin to estimate reliably. A small advertiser with few channels and steady spend often gets more from lift tests and attribution; see our [Bayesian MMM frameworks comparison](/blog/marketing-mix-modeling-bayesian-frameworks-robyn) before committing.

## Sources

1. [An introduction to Meridian](https://developers.google.com/meridian/docs/basics/about-the-project)
2. [An Analyst's Guide to MMM, Robyn](https://facebookexperimental.github.io/Robyn/docs/analysts-guide-to-MMM/)
3. [Introduction to Media Mix Modeling, PyMC-Marketing](https://www.pymc-marketing.io/en/stable/guide/mmm/mmm_intro.html)
4. [Amount of data needed, Meridian](https://developers.google.com/meridian/docs/advanced-modeling/amount-data-needed)
5. [Model API prices, 3AM Marketer](https://3ammarketer.com/data/model-pricing)
6. [Marketing Mix Modeling : A Complete guide: PyMC Labs Blog](https://www.pymc-labs.com/blog-posts/marketing-mix-modeling-a-complete-guide)
7. [Adstocks: Accounting for the Lagged Effect of Advertising - Recast](https://getrecast.com/adstock-rates/)
8. [ROI of marketing mix modeling: calculating, interpreting and improving returns | Analytical Alley](https://www.analyticalalley.com/knowledge-hub/roi-of-marketing-mix-modeling/)
9. [Robyn](https://facebookexperimental.github.io/Robyn/)
10. [What is Marketing Mix Modeling (MMM)? A Complete Guide for Marketers](https://sellforte.com/blog/what-is-marketing-mix-modeling)
11. [GitHub - google/meridian: Meridian is an MMM framework that enables advertisers to set up and run their own in-house models. · GitHub](https://github.com/google/meridian)
