Written for the marketing, finance and analytics leads at brands in Singapore and Australia whose spend now runs across search, social, video and offline channels. It is an owner-side overview; every data requirement quoted here comes from the open-source tools' own documentation, linked below.
What marketing mix modelling measures, and what it can't
Meta's Robyn documentation describes MMM as a statistical analysis that quantifies the incremental sales impact and ROI of marketing and non-marketing activities, running on aggregated data instead of user-level data. Google's open-source Meridian frames the same job as three questions: the historical ROI and contribution of each channel, the response curve for each channel, and how to reallocate budget.
- What it can tell you: each channel's share of outcomes, where extra spend stops paying back (the response curve), and a budget split to test next.
- What it can't: which ad, keyword or creative worked. MMM works at channel level, week by week. It also cannot see effects the data doesn't vary, so a channel that has always run at the same spend is hard to read.
MMM vs attribution vs lift tests
| Method | Data it uses | Question it answers | Main limit |
|---|---|---|---|
| Marketing mix modelling | Aggregated weekly spend, outcomes and outside factors | How much does each channel contribute, and where does extra spend stop paying back? | Needs long, varied history; channel-level only; slow to update |
| Platform / multi-touch attribution | User-level clicks and views | Which campaigns, ads and keywords got credit for a conversion? | Signal loss from privacy changes; each platform credits itself |
| Lift (incrementality) tests | A test group vs a holdout group or region | Did this channel cause extra sales during the test? | One channel and one period at a time; the holdout costs sales |
The three work together. Robyn's guidance treats experiments such as Conversion Lift and GeoLift as the ground truth used to calibrate the model, and Meridian supports calibration with geo experiments in the same way.
The data a model needs
- History: "a minimum of two years of historical weekly data," per Robyn; four to five years if only monthly data exists.
- Granularity: weekly as best practice, and regional where you have it.
- Enough rows for the variables: Robyn's rule of thumb is roughly 7 to 10 times as many rows as variables. Every channel, promotion and outside factor you add needs history to support it.
- Variation: spend that changed over time. Flat budgets give the model nothing to learn from.
- A consistent outcome: the same definition of a conversion or of revenue for the whole period.
Signs you're not ready yet
- Your conversion tracking changed partway through the history, for example a pixel rebuilt, a form replaced or a CRM switched. The model will read the break as a marketing effect.
- Most conversions are recorded only in platform dashboards, with no server-side events or CRM record behind them.
- You have fewer than two years of weekly data, or only one or two channels with meaningful spend.
- Offline sales, such as a showroom, a clinic or a sales team, never make it back into the data.
What to fix first
Server side tracking and clean outcome data are the foundation for every method in the table above, not only MMM.
- Server-side conversion events (for example the Meta Conversions API) alongside the browser pixel. See the Conversions API guide, and check your current set-up with the pixel and CAPI checker.
- Offline and CRM conversion imports, so qualified leads and closed sales flow back to the platforms and into your weekly record. Our offline conversion imports guide covers the WhatsApp-led version.
- Stable campaign naming and a weekly spend log by channel and market, kept in one place you own.
- One lift test on your largest channel, so the first model has a ground-truth point to calibrate against.
Where we fit
We set up tracking before spend scales: server-side conversion events, qualifying lead forms, and hand-offs into the client's CRM or sales team. That is the data foundation an MMM is built on. See how we work with large brands in Singapore and Australia, or our performance marketing practice in Singapore and Australia.
Questions
How much data does marketing mix modelling need?
Meta's Robyn documentation says a robust MMM needs at least two years of historical weekly data, or four to five years if only monthly data is available, and roughly 7 to 10 times as many rows as model variables. Spend also has to vary enough for the model to separate one channel's effect from another's.
Is marketing mix modelling a replacement for platform attribution?
No. MMM is resilient to signal loss but slow and channel-level. Attribution is fast and granular but credits the platform that reports it. Lift tests show causation for one channel over one period. Most large advertisers use all three and calibrate the model with the tests.
What should we fix before commissioning an MMM?
The outcome data: a consistent weekly record of conversions or revenue by channel. That usually means server-side conversion events, offline or CRM conversion imports and stable campaign naming first.
Sources
- Meta Robyn — An Analyst's Guide to MMM — checked 2026-10-01
- Google Meridian — About the project — checked 2026-10-01
How this page is maintained
The data requirements on this page are quoted from the Robyn and Meridian documentation as published. If either project changes its guidance, tell us and we will update the page.