The main risks of AI-assisted marketing mix modelling are confident results built on too little data, models that cannot separate channels that always move together, and budget decisions made on a single run without calibration. Governance should define minimum data standards, require experiments to calibrate results and set who signs off before a model moves real money.
Marketing mix modelling estimates how much each channel contributes to sales using aggregated time series data, such as weekly spend and revenue. Open source tools such as Google’s Meridian and Meta’s Robyn made it accessible, and AI assistants now help teams prepare data, run models and explain outputs. Accessible is not the same as safe.
Why has marketing mix modelling come back?
Privacy changes weakened user-level tracking. Browser restrictions on third-party cookies, Apple’s App Tracking Transparency and consent requirements all reduce what attribution tools can see. Mix models do not need user-level data, so they survive those changes.
Meridian is a Bayesian framework from Google, released as open source, that supports priors from experiments and geographic data. Robyn from Meta uses automated hyperparameter search and ridge regression. Both are serious tools. Both can also produce misleading answers in inexperienced hands.
What are the biggest risks?
Too little data
A mix model typically needs two to three years of weekly data, or geographic splits, to estimate several channels reliably. With one year of data and eight channels, the model has far more unknowns than evidence. It will still produce numbers, and they will look precise.
Channels that move together
If search and social budgets always rise in the same weeks, the model cannot tell which one caused the sales lift. Statisticians call this multicollinearity. The result is often a large share credited to one channel and almost none to the other, depending on small changes in settings.
Model shopping
Robyn produces many candidate models, and teams choose among them. If the chosen model is the one that confirms what the team already believed, the exercise becomes decoration. AI assistants can make this worse by producing persuasive explanations for whichever result is picked.
- Short history or too few geographic units for the number of channels.
- Spend that never varies independently across channels.
- Missing variables: promotions, pricing, stock-outs, competitor activity, seasonality.
- Choosing a model because its result is convenient.
- Treating point estimates as exact rather than ranges.
What can AI assistants safely help with?
AI assistants are useful around the model: cleaning and joining data, writing and debugging code, documenting assumptions, and translating results into plain language. They are risky when asked to choose priors, decide which model is right or judge whether results are plausible without context.
- Safe: data preparation scripts, checks for gaps and outliers, chart generation.
- Safe with review: drafting the methodology note, explaining Meridian or Robyn settings.
- Risky: choosing priors, picking the final model, writing the budget recommendation alone.
- Avoid: asking a chatbot to estimate channel ROI from a pasted spreadsheet without a proper model.
The last point deserves emphasis. A language model asked to “calculate ROI by channel” from raw data may produce plausible tables with no statistical basis. Use proper modelling tools and let the assistant support them.
What are the real limits of mix models?
Even a well-built model gives channel-level, average effects over time. It cannot tell you which ad creative worked, which audience responded or what happened last Tuesday. It is a strategic budget tool, refreshed every few months, not a daily optimization engine.
Mix models also struggle with new channels, since there is no history, and with brand effects that build over years. Adstock and saturation settings help, but they are assumptions, and different reasonable assumptions give different answers.
Which governance rules should you set first?
- Data standard: minimum history, required variables and a written data dictionary.
- Calibration rule: at least one lift test or geo experiment per major channel feeds the model as a prior or check.
- Model selection: criteria agreed before results are seen, such as fit, stability and agreement with experiments.
- Uncertainty reporting: every channel estimate shown as a range.
- Budget limit: no channel moves more than an agreed share of budget, for example 20 percent, on model results alone.
- Sign-off: marketing and finance both approve recommendations.
- Refresh cadence: rerun quarterly or after major market changes.
The budget limit rule is the most practical protection. It lets the model guide direction while experiments confirm the size of the change before larger moves.
How do experiments and mix models work together?
Experiments, such as Meta and Google conversion lift studies or geographic holdouts, measure the incremental effect of one channel at one time. Mix models estimate all channels over a long period. Each covers the other’s weakness. Meridian explicitly supports using experiment results as priors, which pulls the model towards measured reality.
A reasonable cycle: run the model, identify the channel with the most uncertain or surprising estimate, test it with an experiment, feed the result back and rerun. Over a year, the model becomes more trustworthy channel by channel.
How Alien Road applies AI-assisted mix modelling
A typical engagement starts with a feasibility check: we review how much weekly or geographic data the client has, which channels varied independently and which outside factors were recorded. Sometimes the honest conclusion is that the data cannot support a full model yet, and we recommend experiments first.
When it is feasible, we prepare data with scripted, reviewed steps, agree selection criteria with the client in writing, then run Meridian or Robyn depending on data shape. AI assistants help with code, checks and documentation, never with final judgement. Reports present ranges, show agreement with any experiments and propose capped budget shifts, followed by a test to confirm the largest one.
Where to start this week
Export two years of weekly spend by channel and weekly revenue into one sheet. Plot them together and look for weeks where channels moved differently. Those weeks are the evidence any model will rely on.
- Count how many weeks of clean data you have.
- List promotions, price changes and stock-outs by week.
- Mark which channels always moved together.
- Plan one lift test for the channel you are least sure about.
If the plot shows every channel rising and falling together, design budget variation before building any model.
Keep reading: Advertising · AI Advertising Optimization