Follow our blog and discover the latest content and trends in artificial intelligence, intelligent automation, and sales and marketing analytics.
One of the most common problems encountered in sales and marketing analytics is determining what customer actions are important for reaching a goal (e.g., converting a lead or closing a sale), and deciding how to weight and value of each of those actions.
Consider a customer that first learned about your company from a social media post, represented in the image below. In this example, the customer clicked on the post, which drove him to your company's website where he viewed a total of three blog posts before leaving your website. Because the customer viewed several of articles, your marketing team re-targeted the user with advertisements, which eventually led the customer to view additional content. Because the content was good and kept your company at the top of the customer's mind, the next day the customer actually searched for your company and viewed additional content, and then converted to a lead (or perhaps a sale).
Although this graphic only considers the medium (or channel) that your company interacted with the customer, it is still illustrative of the problem: how do we measure the value of each action in the customer's journey so that we can focus efforts on what works best?
In his groundbreaking article nearly sixty years ago, Marvin Minsky (one of founders of Artificial Intelligence) coined the term the Credit Assignment Problem (Minsky, 1961) to describe problems like the one we have in measuring actions on our customer's journey. Generally, the Credit Assignment Problem concerns itself with determining how the success of a system’s overall performance is due to the various contributions of the system’s components. It is a problem that we will encounter throughout our analytics and artificial intelligence efforts (particularly, reinforcement learning).
More specifically for our scenario, the Credit Assignment Problem is about how we should weight each of the actions in the customer's journey so that we do more of what works and less of what doesn't. For this process, we typically start at the goal (the conversion), and work our way backwards to determine how important each of the customer's activities (the blog views, the ad click, and the organic search) was to the customer's decision to fill out a form or purchase something from your company.
Such attributions allow us, over a number of similar conversions, to see what is working (and what is not). As we shall see in future posts, how we decide to solve the credit assignment problem will have a major impact on what we consider to optimize each customer's journey. For now, though, let's look at the most common method used in marketing analytics: multi-touch attribution.
Multi-touch attribution is a way of measuring marketing effectiveness across a complex customer journey. It can be applied to determine the value of a particular medium (or channel), as in our examples above. And it can be equally applied to measure which campaigns, sources, referrers, content, or even actions by your company are most effective.
Multi-touch attribution generally replaces the more traditional first-touch and last-touch approaches. Under each of those approaches, the entire value of a goal is attributed to a single event (the first or last event, respectively) and we ignore the other events. In our example customer journey above, first-touch attribution would assign the entire value to the social media posts, and would value the advertisements and search as worthless to achieving the goal.
Multi-touch attribution, on the other hand, would use a model that attributes partial value to each of the activities based on some model.
Several of the more common multi-touch attribution models are shown in the image above:
There are certainly other models for marketing attribution (for example, Click360 implements several others), but fundamentally each of the models seeks to solve the Credit Attribution Problem by using certain assumptions. Those assumptions, as we shall see, have issues of their own.
The most common problems with single- or multi-touch attribution models are:
Of these problems, the first may be the most important for our purposes. We will consider each in turn.
The biggest problem with multi-touch attribution models is that they contain assumptions that do not perfectly reflect reality, because the assumptions are not based on learning from data. That is because most often the decisions on which model to use are based on our gut feeling--they are usually not made after a thorough analysis of the data to determine the model that best reflects the reality of the customer journey.
This problem is part of a larger problem that Click360 helps marketers and salespeople solve: in order to best understand reality, we should not just impose our assumptions on the world, but rather we should learn from the data and information we have available to make the best decision possible. Click360 uses machine learning and broader artificial intelligence techniques to learn from the data and discover how things work. We believe this is the future as it allows frequent automated experimentation and simulations that simply cannot be reasonably done with attribution models.
That being said, multi-touch attribution models still have a place in the world because they provide a simple, explainable way to measure results. The trick with multi-touch attribution models, if used, is to carefully select the model that best reflects reality.
Another problem with attribution models is that they usually do not take into account the context of a customer or the customer's actions.
For example, a customer in a particular country looking for a particular product may have viewed a general page that was not really relevant to them and then finally found one that was what they were looking for. Under our multi-touch attribution models, those types of factors are usually not distinguished (although you could create thousands or hundreds of thousands of models to account for the many variables).
On the analysis of sequences, you may similarly find that a customer that when a customer does the action sequence "A-B-C" that "C" is considerable more influential to the customer than when the customer just does the action "C".
(Marketers usually understand more simply: that different segments respond better to different marketing. Similarly, salespeople understand that different questions and actions are required for different stages of the sales cycle.)
However, in most attribution models, each event is treated distinctly instead of as part of a sequence. By considering the customer's activities as a sequence we can measure not just the value of an event, but of a sequence of events. Which is typically more how your customer's think about things. For each of those problems,
While multi-touch attribution still has a place for descriptive analytics (because people can more easily understand it), for predictive analytics and personalizing each customer's journey we've found that uses context-aware AI models that are trained from the data and can experiment and change over time offer far better performance and outcomes.
A final major problem with multi-touch attribution models is that they usually fail to distinguish between incremental and recurring revenue.
In reality, many businesses have an established based of customers that feeds a portion of their recurring revenue, and have value in their brands that can influence future decisions. However, multi-touch attribution models frequently do not take that reality into account. That is, they do not take into account the context of a particular user's activities or state (e.g., has the customer previously bought from us?).
The inability to distinguish between existing and new business often limits growth because the metrics may suggest that what drives new/incremental business (which may be necessary for growth) is just a small portion of the whole. A company may conclude that what is actually driving growth is not valuable compared to focusing on existing customers, and and therefore under-invest in is actually working simply because the incremental revenue will always be smaller than the recurring revenue.
A considerably more accurate approach is to combine marketing and click-stream data with sales data from the CRM, and measuring each step of the end-to-end process. Furthermore, simulations and experiments on the effect of adding or removing certain activities from some customers' journeys allow each activity to be effectively measured.
The important thing when deciding which attribution model to use is to understand your business goals for using the attribution model, and how it can best improve the outcomes of your sales and marketing activities. Each of the multi-touch attribution models will be considerably better than single-touch attribution (or no attribution at all), and will provide insights that were not otherwise available.
Click360 supports the multi-touch attribution models discussed above for descriptive analytics, and further provides opportunities for context-aware predictive analytics to personalize and optimize each customer's journey. If you would like to learn more how Click360 can improve your sales and marketing outcomes contact us today!