Is Everything We’ve Ever Learned About Attribution Really Wrong?
While there are certainly inherent problems with the typical attribution models in use today, it is possible to overcome them by asking three important questions.
While there are certainly inherent problems with the typical attribution models in use today, it is possible to overcome them by asking three important questions.
There was a lively discussion on LinkedIn a couple of weeks ago following a thought-provoking post by Gary Angel of Ernst & Young on digital marketing attribution. Angel’s core assertion in his post is that all attribution models are wrong and unlike George Box’s famous quote, he doesn’t think many of them are particularly useful, either. The reason they are “wrong” is because they don’t take into account what might have been expected to happen anyway.
If we take a step back, then digital attribution models have historically tended to fall into three types:
Single and multi-touch models are pervasive mainly because they are simple to understand and easy to implement in a piece of software. However, they are not insight-driven and the decision as to which one to use is largely random and often political.
Data-driven models are based on what the data sees rather than what we necessarily believe. However, as Angel points out, if you are trying to merely allocate conversions back to marketing channels, it doesn’t matter how smart your algorithm is, you’re trying to solve the wrong problem.
Like Angel, I came to the world of digital analytics having spent time in offline marketing analytics and so I have a lot of sympathy for his perspective. Ingrained in the offline analytics world is the notion of lift or understanding “baseline vs. incrementals.” Baseline is the sales that you would have got anyway (often attributed to the brand’s equity) and the incremental sales are those driven specifically by the marketing activities. Marketing mix modeling essentially uses econometric techniques to establish baseline vs. incremental sales. Other approaches involve the use of controlled testing and experimentation to measure the lift say in some markets where activity is taking place vs. markets where it isn’t.
So, where does this leave digital campaign attribution today? As I said, I have a lot of sympathy for Angel’s point of view and I wholeheartedly believe in the challenging of the existing industry mindset to understanding digital marketing effectiveness. Simplistic approaches aren’t going to cut it in the future and more development work needs to be done to improve capabilities in this area. At the same time, though, I believe there is a lot more that organizations could be doing to improve their attribution understanding with some of the tools that they already have by really filtering down to the nub of the problem by asking these questions:
Campaign attribution is often considered in the context of customer acquisition. But there’s a big difference between truly new customer acquisition and previous customers buying again. Different channels are going to be working in different ways for these different customer segments. With existing customers there’s a high propensity for them to buy again, probably using low-friction channels such as email. A lot of these sales might be considered “baseline” sales; they’re the ones you’re getting as a benefit of previous activity. Any attribution activity should probably distinguish between new customers and repeat customers. The results will be very different.
Understanding return on investment is the ultimate goal. But how is return being measured? Attribution approaches that ignore the value of the transaction are potentially missing some vital information. Some channels/campaigns might help to drive higher or lower transaction values than others. Some organizations I’ve worked with look at customer lifetime value as the way to understand ROI.
Buying cycles differ from industry to industry and from customer segment to segment. It’s important to understand how people buy from you to understand the context for your attribution methodology. Different types of consumers will display different patterns of buying behavior even for the same product or service.
As discussed before, existing customers will in all probability behave quite differently to new ones and buying cycles may be much shorter. If your business is largely driven by the acquisition of new customers with short buying cycles, then a single-touch or multi-touch model may be a cost-effective way of understanding where the money is going. If customer buying behavior is more complex, typically involving multiple touch points over an extended period of time, then simple approaches will simplistic as well.
So, by chunking the problem down, I think it’s possible to overcome some of the inherent problems with the typical models in use today, making them them somewhat more “useful.” The future, though, is going to be about getting to the “incrementals.”