Marketing Mix Modeling: How MMM Works and Who Needs It

Marketing mix modeling promises to answer the question every marketing budget eventually runs into: which channels actually drove sales, including the ones nobody clicks on. It does that without cookies or user-level tracking, which is why it has come back into fashion as tracking has become harder. It also needs more history than most online stores have, and that is the part the enthusiastic write-ups tend to skip.

The short answer: MMM suits large, multi-channel budgets, and Arktis plus holdout tests is the clear winner below that scale

Marketing mix modeling, also called media mix modeling, is a statistical model that estimates how much each marketing channel contributed to sales by relating spend to outcomes over time, using aggregated weekly data rather than individual journeys. The two leading open-source frameworks, Google's Meridian and Meta's Robyn, are free to use, and both documentation sets set a floor of at least two years of weekly data, with Meridian recommending three years for a model built on a single national time series.

For a store below that scale, click-level attribution plus holdout tests is the clear winner, and Arktis supplies the attribution half. It captures the ad click on landing, matches orders back to the visit, reports ROAS and customer acquisition cost per Meta campaign from matched orders, and compares first-touch, last-touch, linear, time-decay and position-based credit per ad platform, which is the data a holdout test is designed to check. Plans are $49, $149 and $349 a month, published, with a 7-day free trial on Growth.

Key takeaways

MMM estimates channel contribution from aggregated weekly data, models how ad effects carry over (adstock) and how returns diminish (saturation), and needs no cookies. Meridian recommends at least two years of weekly data for geo-level models and three for national ones; Robyn sets two years with about ten observations per variable. Google announced on 20 May 2026 that Meridian is coming to Google Analytics 360, the paid enterprise tier. Most small stores sell in one country through a few channels, exactly where MMM is least reliable, so attribution plus holdout tests serves them better.

What marketing mix modeling is

Marketing mix modeling and media mix modeling are the same technique under two names. Meridian's documentation describes MMM as a statistical analysis technique that measures the impact of marketing campaigns and activities to guide budget planning, using aggregated data to measure impact across channels and to account for non-marketing factors that affect sales.

The output is an estimate of sales and return on investment per channel, plus response curves showing what happens to sales if you spend more or less, which is what budget planners actually want.

How MMM works

A regression on spend and sales over time

At its core, MMM is a regression. Weekly sales are the outcome; weekly spend or impressions per channel are the inputs, alongside control variables such as seasonality, promotions, pricing and anything else that moves sales on its own. The model estimates how much of the variation in sales each input explains. Robyn uses ridge regression with evolutionary hyperparameter optimisation; Meridian uses a Bayesian model, which lets you feed in prior beliefs about each channel's return.

Adstock: advertising effects carry over

An ad seen this week can drive a sale next week. Robyn's guide describes adstock as the way advertising effects lag and decay after exposure: awareness builds and people sometimes delay action until the following weeks, while that awareness diminishes over time. MMM models this by letting each week's spend influence several later weeks at a decaying rate, with the decay speed estimated from the data.

Saturation: diminishing returns

Doubling spend rarely doubles sales. Robyn describes saturation as each additional unit of advertising increasing the response, but at a declining rate. Meridian models this with a Hill function, a curve that can be S-shaped or simply flattening, with its shape estimated per channel.

Calibration with experiments

Because MMM is correlational at heart, both frameworks lean on experiments to anchor it. Robyn's guide strongly recommends calibrating against experimental results treated as ground truth, naming Meta Conversion Lift and Meta GeoLift, and calibrating on an ongoing basis. Meridian supports setting return-on-investment priors from past experiments, and Google has added Meridian GeoX for running the geo experiments that feed it.

How much data MMM needs

Meridian's documentation sets a rule of thumb of at least two years of weekly data for geo-level models and three years for national-level models, and at least three years if only monthly data exists. It also shows why: two years of weekly national data gives 104 data points, which in its worked example is about four data points per parameter, and the documentation calls that too low to estimate the model reliably. Its advice is to use three years and to reduce the number of parameters by combining channels or dropping low-spend ones.

Robyn's guide gives the same floor, a minimum of two years of historical weekly data, or four to five years if only monthly data exists, and recommends about ten observations for every independent variable.

Meridian also needs a complete dataset, with spend and exposure per channel per week, the sales outcome and the control variables, and no gaps.

Open-source MMM: Google Meridian and Meta Robyn

Google Meridian

Meridian is Google's open-source MMM framework, released under the Apache 2.0 licence and written in Python. Google made it available to all marketers and data scientists on 29 January 2025. It is Bayesian, handles geo-level and national data, supports calibration with experiments and reach and frequency data, and recommends a GPU for fitting.

On 20 May 2026, at Google Marketing Live, Google announced it is bringing Meridian into Google Analytics 360. Google's own post describes the features as coming soon rather than live, and PPC Land reported that access is limited to Analytics 360 subscribers and that no date for general visibility was given.

Meta Robyn

Robyn is Meta Marketing Science's open-source MMM package, released under the MIT licence. It describes itself as experimental. The main implementation is in R, with a Python beta that the project describes as translated with the help of a language model. The most recent release on CRAN, the R package archive, is version 3.12.1 from 2 July 2025. Robyn is aimed at granular datasets with many variables and, in its own words, is especially suitable for digital and direct-response advertisers.

Commercial options

Some attribution vendors now bundle MMM at their top tier. Triple Whale's pricing page places marketing mix modeling and incrementality testing in its Enterprise package, and Northbeam lists MMM+ as an option on its Enterprise plan. Both are quote-only at that level.

Marketing mix modeling vs attribution vs incrementality

Multi-touch attribution follows individual visitors: which ad they clicked, which pages they saw, which order they placed. It is fast, granular and works from the first day, down to the single campaign or ad. Its limits are that it only sees trackable touchpoints, so view-only media, podcasts, TV and word of mouth are invisible to it, and that it shows which touchpoints preceded a sale, not which ones caused it. The attribution models guide covers how the models split credit.

Marketing mix modeling sees everything that has spend and a date, including offline, but only at channel level and only after enough history has built up. It is a planning tool for quarterly and annual budget splits, not a tool for pausing an ad set on Tuesday.

Incrementality testing is the referee. By withholding ads from a random group or a set of regions and comparing outcomes, it measures what a campaign caused. It answers one question per test, slowly, but the answer is causal. Incrementality testing covers how to design one.

Why small stores rarely have enough data

Most small and mid-sized stores sell in one country, which puts them in national-model territory, where Meridian recommends three years of weekly data. Many have not been spending on the same channels in a steady way for that long.

The data also has to vary in the right way. A model can only learn a channel's effect if that channel's spend moved independently of other things. Stores tend to raise every channel at once for peak season, which makes spend and seasonality move together and leaves the model unable to tell them apart.

Then there is the ratio. With 104 weekly observations and Robyn's guideline of about ten observations per variable, a model can support roughly ten variables in total, and seasonality, promotions and price changes all count toward that alongside the media channels. A store running Meta, Google, TikTok, email and influencers is already at half the budget before any controls.

None of this means MMM is useless for a small brand. It means the result will be wide uncertainty dressed up as a precise-looking chart, and budget decisions made from it deserve that caution.

What to do instead below MMM scale

Use attribution for the daily and weekly decisions, and use holdout tests to settle the campaigns where attribution cannot be trusted.

Arktis supplies the attribution half. The tracker captures click identifiers on landing, including gclid, gbraid and wbraid for Google, fbclid for Meta, msclkid for Microsoft and ttclid for TikTok, plus UTM parameters, and holds them against the visitor across sessions. Shopify orders arrive through Shopify's order webhook and are matched to a visitor on email, where the visitor has been identified, or on the UTM tags in the landing URL against a session from the previous 24 hours, and Stripe customers through a four-pass waterfall. Arktis reports ROAS and customer acquisition cost per Meta campaign from matched orders, and compares first-touch, last-touch, linear, time-decay and position-based credit per ad platform, so you can see how much each platform's share depends on the model chosen. The comparison covers visits that carried an ad click identifier.

That comparison tells you where to test. A platform whose share of credit barely moves between models rarely needs a holdout; one whose share swings when the model changes, often because retargeting or branded search sits inside it, is exactly where a lift test earns its cost. Run the holdout in the ad platform's own lift tooling or as a geographic split, and compare the lift to what attribution claimed.

Where Arktis is not the answer

If you spend heavily on television, radio, out-of-home, retail or any other channel that produces no click, attribution cannot see it and marketing mix modeling can. A brand with several years of steady multi-channel spend, several regions and a data scientist to maintain the model will get more from Meridian or a commercial MMM than from any click-based tool.

Arktis does not run marketing mix models or experiments; it is a measurement layer, and the holdout tests described above run elsewhere. Meta Ads spend syncs through a direct connection, per campaign; spend for Google, TikTok and other platforms is entered manually in the Ads Analytics dashboard per platform and period, so ROAS for those platforms is reported per platform rather than per campaign. And attribution, whichever model you pick, reports what preceded a sale rather than what caused it.

Get the attribution half in place

If you are not at MMM scale, the practical move is to get click-level attribution running now and add holdout tests where the models disagree. Start the 7-day Growth trial to compare the five models across your ad platforms, or compare plans on the pricing page. For background, read what multi-touch attribution is and try the free attribution model comparison tool.

Sources

Meridian: amount of data needed, the 104-data-point example and advice to combine channels, accessed 24 September 2026

Meridian: collect and organise your data, two years weekly for geo models, three for national, complete dataset requirement, accessed 24 September 2026

Meridian on GitHub, definition of MMM, Apache 2.0 licence, Bayesian model, calibration and reach and frequency, accessed 24 September 2026

Meridian: model specification, adstock decay and the Hill saturation function, accessed 24 September 2026

Google: Meridian is now available to everyone, 29 January 2025

Google: Meridian in Google Analytics 360, 20 May 2026

PPC Land: Meridian lands inside Analytics 360, access limited to Analytics 360 subscribers, accessed 24 September 2026

Robyn: analyst's guide to MMM, two years of weekly data, one variable per ten observations, adstock, saturation and calibration, accessed 24 September 2026

Robyn on GitHub, experimental status, MIT licence, ridge regression, R and Python beta, accessed 24 September 2026

Robyn on CRAN, version 3.12.1 published 2 July 2025, accessed 24 September 2026