Marketing Mix Modeling (MMM) is a statistical technique from the 1960s which uses linear regression for marketing attribution. It matches spikes and dips in sales to events and actions in marketing, with the resulting model able to determine how to allocate budgets across channels. The method has seen a resurgence with Apple’s release of iOS14 which let users opt out of tracking. MMM uses aggregate data so it’s privacy friendly and future-proof to further tracking restrictions. More interest in the category came when Facebook (Meta) published their open source tool Robyn, and with Google’s equivalent LightweightMMM as well as their 2017 papers on how to modernize the method with Bayesian algorithms and geo-level data.
Yet this modernization effort has so far failed to address an age-old criticism of MMM – it doesn’t account for your advertising’s message or creative, only the amount spend or number of people reached. Nielson found that Creative accounts for 47% of advertising effectiveness with a further 15% contribution from Brand: only 22% came from Reach. Models become unreliable if you include too many variables, so practitioners recommend a ‘scorecard’ approach – what percentage of ads follow a set of best practices – but this doesn’t deliver the necessary granularity. The most it can tell you is what the performance uplift is from following a set of arbitrary ‘best practices’, which isn’t enough to make strategic creative decisions, let alone make tactical optimizations.
Message Mix Modeling is a reformulation of MMM, putting creative in its rightful place at the center of performance. To build such a model, you must collect data in a slightly different way: not by channel or campaign, but at the individual ad level. These texts and images are coded (tagged) based on what was in the ad: i.e. ‘family friendly’, ‘earn rewards’, ‘low price’, etc. As the goal is to see the relative performance of creatives, you don’t have to use aggregate sales as the dependent variable (what you’re trying to predict). Instead you can solve a simpler problem, using analytics or ad platform-reported conversions and use the model to estimate creative contribution. That way the model doesn’t have to the explain the whole marketing mix, so it doesn’t suffer as much from the noise you get in traditional MMM.
The accuracy of the model is measured in terms of how well it predicts unseen data – usually from a holdout group kept aside for such purposes – but it’s also important to calibrate the model based on plausibility of its findings. If the model disagrees with the results of prior incrementality or creative tests, it might need further iteration to get it right. Once the model is robust, realistic, and reliable, it can be used to forecast the impact of making changes to your creative mix. Insights from this analysis can immediately save money by turning off the worst performing ads, or scale up the best performers. However there’s a bigger benefit: by understanding what creative strategies drive impact, that knowledge can be fed back into your creative strategy to do more of what works.
Name | Link | Type |
---|---|---|
Facebook Robyn, Twitter Thread | Social | |
How to bring your marketing mix modeling into the 21st century | Article | |
Measurement and the Magic of Message | Blog | |
Message-Mix Modeling (Econometrics): It’s all about the Creative Message | Social | |
The Death of Marketing-Mix Modeling, As We Know It | Article | |
When it Comes to Advertising Effectiveness, What is Key? | Reference |