AI-assisted marketing mix modelling uses statistical models, now often open-source tools such as Google’s Meridian and Meta’s Robyn, to estimate how each channel contributes to sales, then helps plan budgets. The workflow runs in five steps: collect and clean weekly data, set up the model, calibrate it with experiments, review results with the business and turn them into a budget scenario. AI speeds up data prep, model search and reporting.
What is marketing mix modelling?
Marketing mix modelling, often called MMM, estimates the effect of marketing activities on an outcome such as sales or sign-ups using aggregated data over time. It does not track individual users, which makes it useful when cookie and device tracking is limited. It accounts for factors such as seasonality, price changes and promotions, and it estimates how ad effects carry over time and how returns flatten as spend rises.
Two open-source tools are widely used. Robyn, from Meta, uses automated hyperparameter search to produce many candidate models. Meridian, from Google, is a Bayesian framework that lets analysts include prior knowledge and experiment results. Both need analyst judgement; neither is a button that produces truth.
Where does AI help, and where does it not?
“AI-assisted” here means two things. The modelling tools automate parts of model search and estimation. Separately, general AI assistants help with the slow surrounding work: writing data cleaning scripts, documenting assumptions, explaining outputs and drafting reports.
What AI does not do is decide whether a result is believable. An assistant can produce a plausible chart from a broken model. The analyst and the business still judge whether the numbers fit what they know.
Step 1: collect and clean the data
MMM needs weekly (sometimes daily) data, usually covering two years or more so the model can see seasonality.
- Outcome: sales, revenue, leads or another business result per week.
- Media: spend and, where possible, impressions per channel, such as Google Ads, Meta, TikTok, YouTube, TV and out-of-home.
- Controls: price, promotions, distribution, holidays, competitor activity where known, and macro factors such as inflation in high-inflation markets.
- Organic signals: email sends, organic search sessions, PR spikes.
Export data from each platform, align it to the same weeks and currency, and check for gaps. AI assistants are handy here for writing pandas or SQL scripts that merge files and flag missing weeks, but each script should be reviewed and run on a copy of the data.
Step 2: set up the model
Choose the tool
Use Robyn if your team works in R or Python and wants many candidate models to compare. Use Meridian if you want a Bayesian approach with priors and you have experiment results to feed in. Some teams test both.
Define the structure
- Which channels are modelled separately and which are grouped.
- Adstock (carryover) settings and saturation curves per channel.
- Control variables and seasonality terms.
- Priors in Meridian, based on past studies or experiments, set conservatively.
Write these decisions down. An AI assistant can turn rough notes into a clear assumptions document that the client can read.
Step 3: calibrate with experiments
MMM results improve when checked against causal tests. Geo experiments, conversion lift studies from Meta or Google and holdout tests give ground truth for some channels.
- Run or collect at least one lift test for a major channel.
- Feed results into the model as calibration (Robyn) or priors (Meridian).
- Compare model estimates for that channel with the experiment; large gaps mean the model needs revisiting.
Step 4: review results with the business
Model output is a set of channel contributions, return curves and confidence ranges. Before anyone acts on it, review it in a working session.
- Do the contributions match what the team knows about the market?
- Are the uncertainty ranges wide enough that decisions should wait?
- Are there channels with too little spend variation to estimate well?
- Did any one-off event, such as a stock-out, distort results?
AI can draft the first version of the summary, but the review itself should involve people who know the business.
Step 5: turn results into a budget scenario
Use the model’s budget optimiser to compare scenarios: current split, a shift toward channels with better marginal returns and a cautious middle option. Apply limits that reflect reality, such as minimum brand spend or maximum weekly spend per channel.
Then plan a follow-up test. If the model suggests moving budget from one channel to another, move part of it and measure, rather than moving everything at once.
How Alien Road applies AI-assisted MMM in advertising work
A typical engagement starts with a data readiness check: how many weeks of data exist, which channels have enough spend variation and what controls are available. If the history is too short, the team starts with experiments and a simpler model, and plans a fuller MMM later.
Setup involves building the merged weekly dataset, running Robyn or Meridian, and writing the assumptions document with help from an AI assistant. Results are reviewed with the client’s marketing and finance contacts.
Measurement continues with scheduled model refreshes, usually quarterly, and with lift tests that check the main recommendations. Each refresh report shows what changed and why.
Common workflow mistakes to avoid
A handful of errors repeat across projects. Teams often model too many channels separately when some have little spend variation, which produces wide, meaningless ranges. Others forget major control variables, such as price increases in an inflationary market, and the model credits media for sales that came from pricing.
Another mistake is treating one model run as final. MMM is a cycle: data, model, test, review and refresh. The value grows with each round, as experiments sharpen the estimates and the business learns which signals to trust. Finally, avoid presenting point estimates without uncertainty; a channel return shown as a single number invites decisions the data cannot support.
Where to start this week
Make a list of every channel you spent money on in the last two years and check that you can export weekly spend for each. Note gaps, currency changes and campaigns with no records. That inventory decides whether a model can be built now or whether the first job is fixing data collection.
Keep reading: Advertising · AI Advertising Optimization