Free marketing mix modeling, in your browser

Marketing mix modelling is a statistical method that explains a business outcome — revenue, orders, conversions — using how much you spent on each marketing channel over time, alongside everything else that moves the number: trend, seasonality and the demand you would have had anyway. It works on aggregate weekly data, not on individual people, so it needs no cookies, no pixel and no user-level tracking. It came out of consumer goods in the 1960s and came back when privacy changes broke click-based attribution.

You fit a regression of your outcome on channel spend, after transforming each channel twice. First adstock, which carries part of last period's spend into this one because advertising keeps working after the money stops. Then a saturation curve, which bends the response so the tenth thousand pounds does less than the first. The fitted coefficients give each channel's contribution, and whatever the model cannot attribute to media is the baseline — the sales you would have made regardless.

One row per week with a date, a spend column for every channel, and the outcome you want explained. Two years of weekly data — 104 rows — is the usual minimum, because the model has to see several seasons and several different spend levels to separate them. Fewer than about 52 rows and the estimates get unstable. Adding controls such as price, promotions, distribution or competitor activity improves the fit, but spend and outcome are the core.

Adstock is the assumption that advertising you paid for last week still sells this week. The model carries a fraction of each period's spend forward: x' = x + theta multiplied by the previous adstocked value. Theta is the decay rate, fitted per channel and between 0 and 1. A theta of 0.7 means 70% of the effect survives into the next week, typical of brand and video; a theta of 0.1 is almost immediate, which is what search usually looks like.

A saturation curve describes diminishing returns: each extra pound in a channel buys less than the pound before it, because you have already reached the people who were cheapest to reach. It is usually modelled with a Hill curve, which has a half-saturation point (the spend at which the channel is halfway to its ceiling) and a shape parameter that decides whether returns fall away immediately or after an initial S-shaped ramp. Without one, a model believes doubling the budget doubles the sales forever.

It depends entirely on your margin, and most benchmarks quoted online ignore that. A rough rule is that a channel breaks even at 1 divided by your gross margin — at a 40% margin that is a ROAS of 2.5, at a 20% margin it is 5.0. Ecommerce blended ROAS commonly lands between 3 and 5, but a subscription business with high lifetime value can profit at under 1 on first order. The number that matters for budget decisions is incremental ROAS, not the figure the ad platform reports.

ROI is the total outcome attributed to a channel divided by everything you spent on it — an average across all the money already committed. Marginal ROI is what the next pound would return at your current spend level, read off the slope of the response curve. They separate exactly when saturation bites, so a channel can have the best average ROI and the worst marginal ROI because it is already flat. Reallocate budget on marginal ROI; average ROI only tells you about money you cannot spend again.

No. A mix model finds spend patterns that correlate with your outcome, so it cannot separate a channel that drove sales from one you happened to switch on during a strong quarter, and when two channels always run together it often cannot tell them apart at all. Only experiments establish causality: geo holdouts, matched-market tests, and controlled lift studies. The right use of a model is to decide which experiment is worth running, then to calibrate the model against the result.