Marketing mix modelling, or MMM, is a statistical method that estimates how much each marketing channel contributes to sales, using aggregated weekly or daily data instead of user tracking. AI-assisted tools, including open-source projects such as Google’s Meridian and Meta’s Robyn, make it easier to build models. Small businesses can start if they have about two years of consistent spend and sales data.
Why is MMM back in fashion?
Click-based attribution has become less reliable. Browser privacy features, consent requirements and app tracking changes mean many conversions cannot be linked to an ad click. Platform reports also tend to credit themselves. MMM sidesteps this by looking at the relationship between spend and outcomes over time, without personal data.
The method is decades old, but it used to require expensive consultants. Open-source libraries, cheaper computing and AI coding assistants have lowered the barrier. A marketer comfortable with spreadsheets and some Python, or working with an analyst, can now run a basic model.
Is MMM realistic for a small business?
It depends on data, not company size. MMM needs variation: weeks where you spent more on Meta and less on Google, a pause on one channel, a seasonal push. If you have spent the same amount on the same two channels every week for a year, the model has little to learn from.
- At least 18 to 24 months of weekly data is a common rule of thumb
- Spend by channel per week, from each ad platform
- Sales or leads per week from your store or CRM, not from ad platforms
- Known events: promotions, price changes, stock-outs, holidays
- Ideally some natural variation in spend between channels
For a business spending a small amount across two channels, a simpler approach may be better: planned on-off tests or geographic tests on one channel at a time. MMM becomes worthwhile when there are several channels and the question of where to move budget is real.
Which tools can you use?
Meridian is Google’s open-source Bayesian MMM framework, written in Python. Robyn is Meta’s open-source package, built in R with a Python interface available. Both are free to use and documented on GitHub. There are also commercial SaaS tools that wrap the modelling in a dashboard, aimed at e-commerce brands.
The AI assistance sits in several places. Coding assistants such as ChatGPT, Claude or Gemini help write the data cleaning and model scripts. The libraries themselves use automated optimization to try many model settings. And language models can help explain results in plain terms, though a person must check the explanation against the numbers.
How do you build a first model step by step?
- Step 1: Export weekly spend per channel from Google Ads, Meta Ads Manager and any other paid source
- Step 2: Export weekly revenue or leads from Shopify, WooCommerce or your CRM
- Step 3: Add columns for promotions, holidays and price changes
- Step 4: Check the data for gaps and obvious errors, such as a week with zero sales due to a tracking issue
- Step 5: Run a basic model with the library’s default settings and documentation example
- Step 6: Review the channel contribution estimates and their uncertainty ranges
- Step 7: Compare results with anything you already know, such as the effect of a past pause
- Step 8: Plan one test to check the model’s biggest claim
What mistakes should beginners avoid?
The first mistake is treating the output as truth. MMM gives estimates with ranges, and those ranges can be wide for small data sets. If the model says one channel returns three times its cost with a range from below one to six, the honest conclusion is ‘we do not know yet’.
The second mistake is ignoring things that are not marketing. A price drop, a new competitor or a viral social post can drive sales. If those are missing from the data, the model will credit them to whatever channel happened to spend more that week.
- Do not use ad platform conversion counts as the outcome
- Do not include channels with almost no spend variation and expect clear answers
- Do not make big budget shifts based on one model run
- Do record what changed in the business every week
How do you turn results into decisions?
Use the model to form a hypothesis, then test it. If MMM suggests YouTube contributes more than click attribution shows, increase YouTube spend in a controlled way for a set period, or run a geographic test, and see whether total sales respond. Meta and Google both offer lift study tools for some advertisers that can help validate findings.
Rerun the model every quarter as new data arrives. Over time, the combination of model estimates and experiment results gives a far more reliable view than either one alone.
How Alien Road applies AI-assisted marketing mix modelling
In a typical engagement for a smaller advertiser, we begin with a data readiness check: how many weeks of clean spend and sales data exist, which channels vary enough to model and which business events need recording. If data is too thin, we recommend structured tests first and set up the data collection so MMM becomes possible later.
When the data is ready, our analysts build a model with an open-source framework, document every assumption and present results with uncertainty ranges rather than single numbers. The recommendations always include a validation test. Quarterly reporting updates the model, compares predicted and actual outcomes and adjusts the budget plan.
Keep the first model deliberately small. Three or four channels, a handful of control variables and clear documentation will teach you more than an ambitious model that nobody on the team can explain to the finance department.
Where to start this week
Create a spreadsheet with one row per week for the last two years. Add columns for spend per channel and for revenue from your own store or CRM. Fill in what you can and mark the gaps. If you end up with clean data for most weeks and at least two channels that varied, you are ready for a first model.
Keep reading: Advertising · AI Advertising Optimization