Marketing mix modeling and multi-touch attribution

How much does each marketing channel contribute to sales, and where should the next euro go? Two complementary approaches answer this. Marketing mix modeling (MMM) works top-down on aggregated sales and spend over time. Multi-touch attribution works bottom-up on individual customer journeys.

Marketing mix modeling

An MMM is a regression of sales (weekly or monthly) on media spend by channel, plus price, promotions, seasonality and trend. Two transformations make it realistic:

• Adstock: advertising has a carry-over effect. Part of this week’s TV effect continues next week. A decay rate of 0.70 means 70% of the effect carries over each week; the half-life is the number of periods until half the effect has faded.
• Saturation: each extra unit of spend has a smaller effect than the last (diminishing returns). A concave transformation captures this.

Reading the MMM

The example model explains 84.5% of the variation in weekly sales (adjusted R² = .830) with a mean absolute percentage error of 3.5%, and the Durbin-Watson statistic (2.21) shows no residual autocorrelation. Media drive about 24.3% of sales; the remaining 75.7% is base demand from trend, seasonality, price and other factors. This split is typical and important: most sales would happen without advertising in the short run.

Search has the highest return, 4.94 in sales per unit spent, and its effect is immediate (no carry-over). TV has a long carry-over (half-life of almost two weeks), which last-click measurement would miss entirely. Marginal ROI, the return on the next unit of spend, is what matters for budget decisions: it is lower than average ROI because of saturation.

Budget reallocation

Moving budget from low to high marginal ROI channels, keeping the total budget fixed and each channel within ±50% of its current level, the model suggests shifting spend towards search (+49%) and away from social (−42%) and TV (−28%), for an expected 133,283 in extra sales (9.2% more media-driven sales). Treat this as a hypothesis: MMM estimates are correlational and should be validated with geo experiments or holdout tests before large moves.

Multi-touch attribution

Attribution assigns credit for each conversion to the channels in the customer’s journey. Rule-based models are simple but arbitrary: last touch gives all credit to the final channel, first touch to the first, linear splits it equally, position-based gives 40% each to the first and last and 20% to the middle, and time decay favours recent touches. The Markov chain model is data-driven: it models journeys as transitions between channels and measures each channel’s removal effect, the drop in conversions if the channel were removed.

From 7,974 journeys and 453 conversions (5.68%), removing paid search would cut conversions by 33.4%, organic by 29.3% and email by 26.9%; display has the smallest removal effect (15.0%). Normalised, paid search gets 23.4% of conversions, organic 20.5%, email 18.9%, direct 14.0%, social 12.7% and display 10.5%. Compare with last touch: display gets 9.1% there and 10.5% in the Markov model, and social 9.7% versus 12.7%. Upper-funnel channels that start journeys are undervalued by last-touch attribution.

MMM or attribution?

Mature teams use both, and calibrate them with experiments: MMM for strategic budget allocation, attribution for tactical optimisation within digital channels, and lift tests to check that both reflect real incremental effects.