These are some of the key questions marketers have and should be able to answer in a post-campaign analysis. Looking at one campaign’s data isn’t, however, a feasible way to get accurate & reliable results, especially if there’s multiple overlapping offline and online media channels at play (which usually is the case with larger advertisers). In order to measure individual media channel’s sales & profit uplift, you need to look at marketing effectiveness on a much longer time period.
Traditional MMM projects have fallen for one cardinal sin which ruins the possibility to measure parallel media channels’ impact accurately: Utilizing aggregated and/or inadequate sales data in the modeling. The reason for this is often that the client doesn’t want to (or even can’t for statutory reasons) disclose this type of data to external party. As a result, the modeling will have to include a lot of assumptions in allocating each factor’s effect with high-level data. In other words, it’s like trying to explain new educational programme’s impact on single student’s study performance by using class average grades.
“Traditional MMM projects have fallen for one cardinal sin which ruins the possibility to measure parallel media channels’ impact accurately: Utilizing aggregated and/or inadequate sales data in the modeling.”
Another shortcoming is in the subheading: projects. We all know that no model is perfect. But we can get close to perfect by practicing, or technically put training the model with data. This possibility is partly lost with the project-format, as the modeling process can start with different set of premises, hypotheses, business assumptions and data sets. In the worst-case scenario, each modeling starts from scratch, making it impossible to build on earlier learnings. Combine time and money-consuming projects with insufficient data quantity and quality, and you’re in for a disaster.
If you were to analyze ongoing campaign’s media uplifts with no additional data, it would be impossible to get reliable and accurate results (unless you’re pure-ecom player with only online media channels in use). Additionally, if there has been same campaigns with exactly same media mixes at the same time of the year, it is impossible to attribute sales and profit to individual media channels. But this is rarely the case. Below’s what you need to measure individual media channel’s sales and profit impact:
Getting reliable media-level results requires typically data from at least the past 2-3 years. Anything less will start to decrease the credibility of the results, as the model won’t be able to measure how e.g. seasonality or weather affects the results. Moreover, this extensive dataset usually includes decent amount of variation in advertising exposure and media items, which both help the model to attribute the sales/profit uplift accurately.
With sufficient historical dataset, analyst/data scientist can apply time series analysis to compare daily sales uplifts when specific product is not in any promotion or media, when it’s in promotion but not in any media, and when the product is promoted in specific media. More often than not products are promoted in multiple media channels simultaneously, but as the media mix varies within the time period, the model starts to recognize individual media channel’s impact on sales and profit.
In addition to getting enough data, data granularity (how detailed the data is) is the other prerequisite for successful modeling and reliable results. The sales should be on item-location-day level (receipt data is the best format for advanced modeling), whereas the marketing data should be on as granular level as possible (example below).
Knowing what products have been sold, when the sales have taken place and how the results are divided into different geographical areas enables the model to find the link between specific media activities and commercial metrics. Combine granular marketing and sales data with abovementioned timeline, and the model starts to recognize for example:
· Media’s regional impact
· Seasonal changes in specific media’s ROI
· Spillover/synergy effects between different channels
Recognizing and concretizing these metrics will not only help you make more informed business decisions in the future, but they also improve the model over time (assuming that the company has passed the project-phase in the modeling and analytics department).
So, decent data quantity and quality will not only help marketers to separate parallel media channels’ uplifts from each other, but they also make the results much more accurate and reliable compared to traditional methods (let’s keep in mind that no model can deliver 100% accurate results. We’re striving for that 99% accuracy level). Whether you know it or not, you already have all the data that’s required for the analysis already available: Sales data comes from your internal systems, and marketing data (which is the trickier one) can come from your in-house team or media agency. Adding external data sources (such as weather data, competitor prices etc.) is optional, and it turn counterproductive after specific point when the model starts to overfit the results.
If you skimmed through the article, here's a checklist what you need to start measuring the sales lift of an individual media channel:
1. You need sales and marketing data from at least past 2-3 years
2. The data needs to be as granular as possible. Item-location-day level sales data and media metrics + media items with marketing data are the bare minimum
3. Analyzing single campaigns, quarters or even years without wider dataset leads to inaccurate results. You need to look at the big picture to understand the smaller details
4. Continuous analysis = continuous development. Marketing Mix Modeling projects are a thing of the past, with many flaws and heavy price tag on them
In addition to getting more actionable insights and information about your marketing activities, separating parallel media channels’ uplifts enables you to simulate different scenarios more accurately. As the model has learned how specific media drives sales in specific context, it can now calculate how planned media plan will play out in the future in terms of sales and profit (which wouldn’t be possible if the model utilizes aggregated data).
This type of predictive modeling works with novel media channels and media items as well, as there’s usually correlation between media types and product categories to some extent. Example-wise, separating media channels is like unraveling a puzzle to pieces: After you’ve done it successfully for a couple times, you know how to assemble a new one with similar image and pieces, even if they aren’t exactly the same as last time.