Given that personalization means different things to different people, it is useful to provide my understanding of the term. I see personalization as providing relevant information to an individual based on past, present and predicted behavior in a timely manner through the right channel. Information offered can range from product recommendations or next best actions, to asking a customer to give additional information, to a “Thank you for being such a valuable and loyal customer” message.
Personalization may be based on something as simple as remembering what an individual has purchased (where most systems fail, including Amazon, and Netflix) in the past so that the business does not keep recommending a product that individual has already purchased. Or it may involve using all kinds of information to predict that a certain customer will react to an offer differently than another customer – and using a targeted message to communicate with them.
For years now, companies have implemented systems that help them interact with their customers. From enhancing web sites, mobile apps and call center operations to improving sales transactions, companies are keenly aware of the need to improve customer connections in order to increase their bottom line.
The payback is clear. Increasing loyalty and keeping current customers costs less than finding new ones. Over a lifetime, loyal customers purchase more, are less expensive to sell to, and studies show they will refer other valuable customers to a business.
But how do you create loyal customers? Build a positive and favorable customer experience. Of course, making sure that on-hold times are short, that customer service representatives have access to a record of previous problems and that the Web site is easy to navigate are all critical to improving customer experience.
However, these operational systems are not being used to their full potential until a company can predict the right thing to say, or sell, to the right customer at the right time – in other words, until they begin to optimize every interaction with every customer in a multi-channel environment.
Individual Customer Experience
In today’s world, optimizing customer interactions is about individualizing the customer experience. In many ways this is like the original promise of CRM systems: replicating the old-fashioned general store experience. When one general store served everyone in a town, the owner knew everyone that came into the store. When a certain individual walked in, the owner knew that they were most likely looking for a certain product based on past buying habits and acquaintance with the persons and their circumstances.
Today we would like to accomplish the same but serving a mass of customers across many geographies. Here is the challenge: people will behave differently according to the role that they are in at a specific moment, the group processes they are influenced by, their culture, their previous (life) experience, their financial situation, and the emotions that they experience at the moment of interaction.
The key to getting back to knowing your customers is getting insight into their different behavioral patterns over time, over channels, over products and into their social and emotional affiliations using preferences, attitudes, and other ‘softer’ information. This information can then be used to maximize the experience of a customer during the relationship.
The value of personalization has two dimensions: value to the customer and value to the company. To the customer it means being recognized and served better. The transaction will be made easier for the customer: the “customer experience” will improve. Also, the customer will find that the company listens better to their needs and therefore experience reduced noise in the communication. The customer will start to see the relationship with the company as a “personal shopping experience”.
The company gets their value from an increased and more efficient conversion from browsers to buyers. It will be easier to communicate to customers the right offers which will increase the up- and cross-sell ratios and therefore the customer profitability. Because the offers are all personalized the cost involved with normal waste in communicating with the customer base will be reduced. Last, but not least, the company will see an increase in loyalty because the cost of switching to the competition for the customer will increase because they would need to restart building the history and the information exchanged with that competitor.
Approaches to Personalization
Personalization can be based on 4 types of approaches.
- Memory: Remember past interactions and supplied information
- Segmentation & Rules: Assume that customer has similar behaviors and desires of segment
- Recommendation Model: Create a predictive algorithm that will automatically generate a recommended action
- Dynamic Personalization: Multiple modeling techniques to converge on the specific offer that is right for an individual customer
The approaches are ‘embedded’ which means that ‘higher order’ approaches also (can and usually will make) use of the ‘lower order’ ones.
Cheers-effect: ‘You want to go where everybody knows your name'”
Memory & Supplied Information
The first approach, based on memory, is more operational in nature. These data can be used to facilitate repeat behavior. Examples are:
- Memory: A bartender in a hotel that can serve a guest “the usual” based on what this guest has ordered on a previous visit making the guest feel recognized and “at home”
- Supplied Information: An airline that remembers a passenger’s previously stated preference for an aisle or window seat. Or a hotel that remembers a guest’s preference for a (no) smoking room. The customer will feel that information that (s)he gives is used efficiently and there are no unnecessary repeats of questions.
This approach can be implemented using an operational package with simple business rules and can then be embedded into the operational processes. The overall effect is more loyal and satisfied customers because the company makes them feel recognized and it shows that the information the customers give is used to their benefit. It could be described as the “Cheers-effect: “You want to go where everybody knows your name”
Segmentation & Rules
This type of personalization is based partly on analysis of customer data and partly on knowledge of the business. In this approach the customers are divided into homogeneous groups or segments based on their background data like demographics or their behavior. These profiles are linked to purchasing and other behavior to create rules that define the action to be taken for personalization.
The key to this approach is defining profiles or segments and then in a top down approach the customer is linked to the profile (s)he belongs to according to the company. An example is:
- A bank has identified 3 segments in their customer base: ‘Young’, ‘Middle’, ‘Old’. It is established, either through analysis or through a business rule definition, that customers from the ‘Young’ segment are (or should be) sensitive to offer A. A rule has been defined that when a customer from this segment comes into a local office the teller should present him or her with this offer. The benefit for the customer is that (s)he should receive offers that should be valuable to them.
Defining rules based on segments identified without analyzing customer data may lead to ‘wishful thinking’ push strategies that are not accepted by the market”
This approach has worked very well in generating ROI in the past, especially in database marketing there are examples of increased response rates and revenue leading to x00% in ROI on marketing spend. It also contains some inherent risks:
- Defining rules based on segments identified without analyzing customer data may lead to ‘wishful thinking’ push strategies that are not accepted by the market. This may lead to suboptimal offers to the customers and therefore to irritation and churn.
- Having a limited set of segments may lead to ‘forcing’ customers into a profile that they do not belong to. Again, this may lead to irritation and loss of customers.
- The rules base of this approach can become very large and will take a lot of maintenance. It can therefore become cluttered with rules that are out of date or rules that are in contradiction.
These risks have been mitigated in the use of more advanced analytics, especially in the field of data mining. This allowed for data driven segment definitions and evaluation of results based on the likelihood that customers are interested in an offer creating ‘pull’ definitions instead of ‘push’. It also allowed for a multitude of (micro-) segments that allowed them to assign customers more precise. By using automatic deployment in business rules management systems of the identified segments and rules and monitoring their effectiveness the prioritization in and maintenance of the rules became more efficient, but still very burdensome and static.
With the arrival of the internet and success stories like Amazon, Google, Netflix and the way the offer suggestions to customers: “People who bought this book also bought this”, “People who searched for this also searched for this”, “People who watched this also watched this” recommendation engines/systems have become the flagship of personalization in the market. The increased availability of data and the development of more extensive (machine learning) algorithms has improved the quality of the recommendations these systems make manifold.
The big advantage of these systems is that they use a bottom up approach to detect patterns in buying behavior that are then automatically translated into dynamic rules. Based on new data a customer’s recommendation can be updated in real time and the rules do not need to be stored and maintained. In the best implementations the earlier 2 approaches are embedded but the memory and segmentation & rules input is limited to decision rules that are more stable over time and do not need frequent modification.
The biggest misconception about recommendation systems is probably that they make personalized recommendations”
There are also some big disadvantages to and misconceptions about these systems. Recent research has also shown that recommendation systems themselves have an influence as well (see “The Hidden Side Effects of Recommendation Systems”). And a lot of recommendation systems do not handle the introduction of new products well or they fail to detect changes and new opportunities in the market.
The biggest misconception about recommendation systems is probably that they make personalized recommendations. This view is usually supported by someone quoting some example of a few recommendations that were made that were useful. In the story that they tell they forget that those few recommendations are only a small % of the total recommendations and that most of them were irrelevant. Recommendation systems’ objective is to increase the number of offers accepted on average. If a company generates 100,000 recommendations per month randomly and 0.5% of those offers are accepted with an average order value of €50 they make €25,000. Using a recommendation system that will help them make at least 2% of the offers relevant they increase their revenue to €100,000 while still being wrong 98% of the time.
what , at this time, for this specific customer, through this channel is the right kind of information to provide to maximize the customer experience and the value of that customer?”
Dynamic personalization is the opposite of segmentation. It can best be described as aggregation: Based on an individual customer’s profile and the information of the interaction the customer is aggregated through a set of models to identify what, at this time, for this specific customer, through this channel is the right kind of information to provide to maximize the customer experience and the value of that customer.
The strength of this approach is that it takes the customer as a starting point and it allows for the customer to be offered different information depending on the interaction while still being treated consistently.
This requires the use of multiple models in tandem like described in “Why “Many-Model Thinkers” Make Better Decisions”. The types of models working together in different combinations to achieve the maximum effect could be:
- Need: what are our customers getting out of using our products/services?
- Value: What does the customer get out of this relationship?
- Product usage: how do customers use our products/services?
- Customer Profitability: what is the company getting from the customers?
- Risk/credit/fraud: are any customers abusing our products/services?
- Acquire: how do we effectively acquire more of our good customers?
- Cross-sell: how can we increase the profitability of our customers?
- Retain: how can we keep our profitable customers?
- Channel: How effective is a channel and how do customers behave over channels?
- Campaign: Which customers should be approached with what message over which channel to maximize the returns?
These models are intertwined. Maintaining them requires a process where these relationships are stored so a dependency analysis can identify the impact of changes. They also still need to be embedded in a (limited) set of business rules from approaches 1 and 2 that are defined by good business practice, for example, “Don’t offer a product that is out-of-stock or that the customer already bought” or “Don’t include customers in an e-mail campaign that have indicated to want to receive their communications through traditional mail”.
Most importantly, the results of the model need to be deployed to the point of interaction. Being able to predict what a customer is likely to buy doesn’t make the customer buy. The offer must be made, and the result needs to be evaluated and incorporated in updated models to ensure closing the loop.
No personalization approach will give you perfect results. We need to keep in mind is that perfect information does not exist. There will always be incomplete data, algorithms can have flaws. The best advice is given by Box & Draper:
Remember that all models are wrong; the practical question is: how wrong do they have to be not to be useful”