When I ask organizations about their marketing roadmaps, they often stress the need for personalization capabilities in digital experience delivery as one of their top organizational objectives. The problem with this focus, however, is that implementing personalization capabilities is not an end in itself.
While the constant barrage of messages from marketing technology vendors implies that installing their tech solutions will produce better marketing, in reality, technology is always a means to an end. Technology may be an important enabler of good experience strategy but it doesn’t create it, and without good experience strategy — and data and content — technology alone cannot deliver results.
Machine Learning for Content Personalization
For the last year, I have been working with several marketing teams on the implementation of algorithmic content personalization using a machine learning approach. While the concept of machine learning may sound exactly like letting your technology do your marketing for you, this approach actually requires continuous refinement of the underlying consumer experience strategy and a very strong experience design methodology to make the machine’s efforts pay off in performance results.
In describing machine learning, I first compare the differences between the rules-based experience delivery that people are accustomed to developing through their marketing automation tools and the adaptive experience delivery made possible by machine learning.
Predefined Versus Redefined Targeting
Simply put, rules-based targeting applies predefined decisions about what content to present to different categories of consumers, while machine learning continually redefines its targeting decisions based on observations about consumers and their anticipated responses to the combinations of content it has available to present to them.
In both cases, consumer categories are defined by some combination of CRM/sales data and look-alike modeling, a Data Management Platform or some other consumer profiling data such as referring media campaigns. Consumer categories are typically related to age, gender, income and purchase history, and are sometimes further refined through RFM (recency, frequency and monetary value) and expected LTV (lifetime value) analyses.
Rules-Based Versus Learning Approaches
In deciding on a personalization strategy, marketers must determine whether a rules-based-targeting or a learning approach is the right strategy for their needs. Rules-based approaches are very good at maintaining consistency in targeting, while adaptive approaches are best in situations where the alignment between consumer categories and the most effective content is less clear.
Rules-based methods tend to focus on a fixed number of distinguishable consumer categories, while adaptive approaches will continuously refine the composition and number of audiences that can be distinguished by both unique characteristics and their response to different experiences.
Designing Personalized Customer Experiences
With a decision in hand about the right targeting approach for experience personalization, the next non-technological challenge is to design the experiences that will be shown to each distinct customer category. While machine learning can uncover categories of consumers, it can only do so in the context of the content it has been given to generate responses. In both rules-based and adaptive, if there are only two versions of content, A and B, then there can, by definition, be only two audiences.
Whereas rules-based targeting would have predetermined how large each of these audiences would be in advance — based on the proportion of consumers in the segments pointed to each respective version — adaptive targeting will discover the size of each audience based on actual responses to A and B by different types of consumers.
An Example of Rules-Based Targeting
In discussing how many options for personalization need to be created and how to determine the content of those options, I often resort to the analogy of a restaurant owner deciding on a menu.
For some reason, this restaurant owner has decided on a liver-themed menu of three dishes: liver on a roll, liver on rye, and liver on cucumber.
For a few weeks, everyone coming through the door is offered one of these options. If the restaurateur is using a rules-based approach, then people wearing jeans will always be offered liver on a roll, people in khakis will be offered liver on rye and people in any other outfit will be offered liver on cucumber. Some of each group will accept and eat but many more will leave without eating the menu choice they have been offered.
Taking the Superior Adaptive Approach
If that same restaurant owner uses an adaptive approach instead, (s)he begins by randomly offering a different dish to each type of person and then begins to form theories about what characteristics might predict which dish they will be most likely to eat. As the restaurateur forms various theories, (s)he begins targeting based on those theories, occasionally retesting the theory randomly to confirm that the emerging pattern still outperforms a random distribution.
Given how poor the fabric of pants would likely be as a predictor of liver preference, we would expect that the adaptive approach would be better at uncovering the actual measurable predictors of liver dish preference than rules-based targeting.
The Suboptimal Strategy Dilemma
However, the fact that adaptive targeting can more accurately predict a preference for one of three liver dishes and optimize their sales should not be construed as proof that liver dishes are the best food to be offering to everyone who walks through the door. This is a classic case of the dilemma I call the optimization of suboptimal strategy.
While this owner has successfully applied a learning technology to maximize conversion against the “content” the restaurant is offering, it should be apparent that there are better food options to show to passersby. The restaurant may have maximized its conversion among diners who will eat liver of some kind but it has no content to motivate people who don’t like liver.
Blinded by Confirmation Bias
Marketers who come at personalization from the standpoint of technology run the same risk as this restaurant owner: They get caught up in optimizing conversions for varieties of content that have distinguishable audiences, yet miss the fact that those audiences represent only a fraction of their potential consumers.
The self-limiting decision to offer only three types of liver results from taking an inside-out approach to experience design where the brand decides how it want consumers to engage with it and then offers only that experience.
Then, just as our restaurant owner saw improved conversion among a small sample of foot traffic as validation of the decision to focus on liver, brands that test only one way of engaging consumers become blinded by the fog of confirmation bias.
Shifting to Outside-In Thinking
Moving to the recognition that consumers might want to diversify into pizza or falafel dishes cannot emerge from an inside-out perspective. Those insights require a shift to the kind of outside-in thinking that unlocks the full potential of targeting technologies.
My next and final post in this three-article series will focus on those approaches — you can catch up with my first article in the series here.
Editor's Note: Scot Wheeler will dive deep into how to unlock the full potential of targeting technologies during his presentation at CMSWire's DX Summit in Chicago, Nov. 14 through 16 at the Radisson Blu Acqua in Chicago.