Click-based attribution is degrading and will keep degrading. Third party cookies are largely gone, mobile app tracking requires consent most users decline, and a growing share of conversions arrive with no usable path attached. The response from most teams has been to keep reading the same last-click reports while quietly accepting that the numbers no longer add up to revenue. Marketing mix modeling ai is the practical alternative: instead of following individuals, it models aggregate spend against aggregate outcomes and estimates what each channel contributed.
The approach is old. Consumer goods companies were doing regression on media spend and sales decades before anyone had a tracking pixel. What is new is that the modelling is now within reach of a mid-sized advertiser. Open source libraries handle the statistical machinery, cloud compute makes Bayesian estimation cheap, and the model layer helps with the parts that used to require a specialist: specifying priors, checking fit, and translating output into a budget decision.
What the model actually does
At its core you are fitting weekly or daily outcome data against spend per channel, with two adjustments that matter more than the regression itself. The first is adstock, which models the fact that advertising keeps working after it stops: a television burst or a display campaign has an effect that decays over weeks rather than ending at midnight. The second is saturation, which models diminishing returns, because the tenth thousand euros in a channel does not buy what the first thousand bought.
Add controls for everything else that moves sales: price changes, promotions, competitor activity if you can observe it, seasonality, holidays, stock outages, and any large one-off event. Omitting a control does not produce an error message. It produces a confident, wrong coefficient, usually assigned to whichever channel happened to be running during the event.
What marketing mix modeling ai can and cannot answer
Be clear about the resolution. A mix model answers channel-level and sometimes campaign-level questions over weeks. It does not answer keyword-level or creative-level questions, and it never tells you which individual customer came from where. Anyone selling you a mix model that produces user-level attribution is selling you something else.
The questions it answers well are the expensive ones. Is our brand search budget incremental or harvesting demand we already created? What happens to total revenue if we cut display by thirty per cent? At what point does our best channel stop scaling? Those are the decisions worth millions and they are precisely the ones last-click reporting cannot touch, because last-click credits the final touchpoint and is structurally blind to the channels that created the demand.
The data you need before starting
The requirements are less demanding than people assume, but they are strict on one point: history. You need enough periods to fit the parameters, which in practice means at least two years of weekly data, and ideally variation in that spend. If every channel has run at a flat budget for two years, the model has nothing to learn from. Deliberate variation is an input, not a nuisance.
- Weekly spend per channel, gross, including production and agency costs if you want a true efficiency read.
- Weekly revenue or another single outcome metric the business agrees on.
- Price and promotion history, at the same weekly grain.
- Known external events: launches, outages, PR spikes, competitor campaigns.
- Impressions or reach per channel where available, which usually fits better than spend alone.
Most of the project time goes into assembling this table. The modelling itself, once the data is clean, is a matter of days. Teams that budget the reverse are always surprised.
Validate against an experiment
A mix model is a set of estimates with uncertainty attached, and it can be confidently wrong. The only reliable check is a real experiment: turn a channel off in a set of regions, leave it on elsewhere, and compare. If the model predicted a twelve per cent revenue drop and the holdout shows two per cent, the model needs work. Run at least one geo holdout per year and use it to calibrate, not just to confirm.
This is also the honest answer to the accuracy question. Mix models and click attribution disagree, often substantially, and neither is ground truth. Incrementality experiments are the closest thing to ground truth available, and both other methods should be judged against them.
Use the output as a direction, not a verdict
The most common misuse is treating the model’s ROAS figures as precise numbers and rebuilding the media plan around the top of the table. Look at the credible intervals. If two channels overlap heavily, the model is telling you it cannot separate them, and shifting budget between them on the basis of a point estimate is noise chasing.
Where it earns its cost is directional and repeated. Refresh the model quarterly, track how the response curves move, and use it to set the broad allocation between channels while keeping click data and experiments for the decisions inside each channel. Marketing mix modeling ai does not replace your existing reporting. It answers the question your existing reporting was never built to answer, which is what advertising is worth in total when nobody can be followed.
Keep reading: AI Advertising Optimization