David Raab Article Archive

“Dear Current: Picture yourself and the entire Resident family on a fabulous, all-expense-paid vacation….”

Such are the embarrassingly plebian ancestors of personalized marketing. From its early days as an attention-grabbing gimmick on pitches for magazines and shady real estate deals, personalization has grown in respectability to become the bedrock of modern customer management. Indeed, some would argue that personalization pretty much IS customer management.

Like any grand success, personalization has attracted hangers-on who are eager to use its popularity for their own purposes. Today, nearly every marketing-related system lists the ability to help plan, execute or analyze personalized interactions as one of its main benefits. (Come to think of it, those pitches themselves are rarely delivered in a personalized fashion. But let’s not dwell on what this might imply.)

Of course, marketers rarely bother to define their terms with any precision. As a result, so many conflicting claims being made about personalization that it faces an identity crisis. So, without getting bogged down in the theory of why personalization is important in the first place, it’s worth taking a look at what different kinds of personalization systems actually do. (For a range of opinions on some of the broader issues, try the excellent Web site at www.personalization.com.)

The most ambitious usage of personalization treats it as synonymous with interaction management–a process whose goal is to generate the most appropriate response to any customer action. Obviously the key to this definition is the meaning of “most appropriate”, which really boils down to the question of “most appropriate in terms of what?” That is, what factors are taken into consideration when trying to identify the most appropriate response?

Part of the answer has to do with measurement–what, exactly, are we trying to maximize? Is it short term profit? Net present value? Customer lifetime value? Retention rate? Market share? Customer satisfaction? Loyalty?

Personally, I favor a measure of return on investment–making sure that each dollar yields the greatest possible long-term profit. This isn’t particularly controversial, although it does mean accepting less-than-maximum profits from relationships with some customers. What’s really important is the definition treats every decision as an investment–even decisions that involve no incremental out-of-pocket expenditure, such as allocating limited resources within a call center or field force. How to actually perform this sort of measurement is a knotty problem. Fortunately, personalization can be deployed without solving it, so long as marketers are willing accept whatever imperfect approximations are available.

The other answer to “in terms of what” has to do with which factors are considered in selecting a reply. This is where today’s personalization systems vary most broadly.

The first factor to consider is information about the customer herself. This might seem pretty darn obvious, but in fact it’s possible to create “personalized” promotions that treat all customers the same–as in the “Dear Current” letter itself. It’s also possible to treat customers differently without knowing anything about them as individuals–consider a Web site that shows different pages depending on the visitor’s domain name or incoming link. Even systems that do identify individuals and treat them differently can vary significantly in the types of data they access. These range from a simple segment code appended during a batch process, to real-time look up of detailed historical information such as past purchase transactions, to access to information provided during the current interaction.

In fact, the breadth of available customer data is a significant differentiator among today’s customer management systems. For integrated Customer Relationship Management products like Siebel, a single customer transaction database is touted as a key advantage over single-purpose call center or sales automation systems that do not share data across touchpoints. Among Web-based personalization products, vendors who can integrate data from non-Web “legacy” systems take great care to distinguish themselves from products whose personalization is limited to profile data gathered within the system itself.

The second factor that systems might consider is the current interaction itself. Again, this may seem trivial: surely if the purpose of a personalization system is to respond appropriately to the current interaction, then the system must take that interaction into account? Yet, again, there are current products that do no more than distribute messages chosen in advance–say, by running a nightly batch process that produces a list of messages to be sent to individual customers if they should appear the following day.

This is nowhere as silly as it might sound. In fact, if each customer interaction is considered an opportunity to deliver a free advertisement, then delivering no message during those interactions would be as wasteful as buying a minute of TV time and broadcasting dead air. Viewed from this perspective, even a slightly off-target message is better than nothing–so long as some mechanism ensures the message isn’t so inappropriate in the context of the transaction that it actively antagonizes the customer.

Still, it does seem much smarter to deliver messages that are directly related to the current transaction. Obvious examples are recommendations for related products (cross sell) or more profitable alternatives (up sell). Many systems provide this sort of personalization, although they often develop the recommendations without considering the individual involved. For example, systems frequently present a list of products that are most often purchased together, without checking whether different recommendations might be appropriate for different customer segments, which products this particular customer has purchased or whether the customer has a history of accepting such suggestions at all.

If there’s no hope of a customer accepting the recommendation, then making it is a poor investment of company resources. Of course, the expected return on the investment depends on the expected revenue compared with the cost, and cost itself varies by channel: an email is much cheaper than the time of a call center operator. Then again, even a “free” Web message may carry a cost in terms of customer annoyance. As already noted, resolving this with absolute precision requires advanced metrics that do not really exist. But it’s still possible to set intelligent policies without those metrics–so long as the system provides access to the relevant data.

Systems also differ in exactly how much information they take into account about the current interaction, and in how broad a range of responses they consider. For example, some Web-based systems can distinguish the behavior patterns of a leisurely browser from that of a hurried buyer, and treat each appropriately. Or, to be a more precise, some Web-based systems capture enough data to make these distinctions based on rules defined by a marketer. Even the best of today’s Web-based systems couldn’t discover and react appropriately to these categories by themselves. Similarly, some personalization systems can consider a wide range of responses, such as whether it’s better to make a cross sell offer or to send the customer a thank-you note. This calls for more sophistication than just choosing the best cross sell offer, since it’s harder to find a common basis to use in ranking the alternatives. Once again, this is an issue of metrics and once again, today’s systems would mostly rely on rules created by marketers rather than on rankings the systems develop autonomously.

A third factor considered by some personalization systems is the customer’s status in current marketing campaigns. Many of today’s marketing systems assign customers to one or more campaigns that involve a sequence of multiple messages. Part of the challenge faced by any marketing system is to limit and coordinate these assignments so the customer is not overwhelmed by too many messages or sent messages that conflict. Like personalization, this poses the challenge of ensuring that the most productive of all possible choices are made. But campaign assignment is generally done from a campaign rather than customer orientation. (That is, the question is which customers to assign to a given campaign, not which campaigns to assign to a given customer.) This is not itself a personalization issue.

Where personalization and campaigns do intersect is when choosing the specific message to deliver during an interaction. If a customer is active in multiple campaigns, then the customer may be eligible to receive several different messages at the moment an interaction occurs. Since the system can presumably send just one message at a time (or, at least, needs to rank the different messages by priority), the personalization system must choose among them. This may be the knottiest of all the analytical problems associated with personalization, since it requires estimating not just the value of message by itself, but the incremental value of one message within a sequence. In fact, a truly thorough system would go further to estimate the value of repeating the same message compared with switching to a different message later in during the same interaction, and also somehow take into account the number and timing of future message delivery opportunities. It’s not even clear that a sound theoretical basis for making these decisions exists today. Certainly no personalization system can make such decisions in practice.

Still, decisions must be made. One common approach is to prioritize the campaigns when they are set up, using a single sequence that is the same for all customers. This has the virtue of simplicity, although it’s possible that a particular campaign will be of greater value for a specific customer than some other campaign that is ranked higher over all. In fact, it seems almost certain this will happen in some cases. When differences among the values of messages within the same campaign are added into the mix–for example, the sixth renewal offer to a magazine subscriber is worth less than the first–choosing solely on the basis of campaign priority seems inevitably suboptimal.

Some of these problems can be overcome through setting different priority sequences for different customer segments. If these segments are defined with enough precision, then the campaigns should be ranked correctly for almost every individual. Most systems that accommodate campaign rankings can be made to support this approach, although it might take some awkward tricks–such as creating several versions of the same campaign with different rankings and selection rules. A refined scheme along these lines is likely to be an administrative nightmare.

A more manageable approach would be to define campaign priorities on an individual basis. This could most easily be done when the customer is first assigned to a campaign. After all, if the system is well designed then it should already calculate the value as part of the selection process. Ideally the value would be recalculated over time. This could be done at fixed points during the life of the campaign–say, after every message–which would take into account the change in expected returns. Or, the value could be recalculated on-the-fly when messages need to be chosen. Systems that can execute scoring models in real-time can perform these sorts of selections. Sometimes the scores are used within a larger rule-based framework that makes broad choices–say, between retention campaigns and cross sell campaigns–before letting the system make the final specific selection. Since a customer may be eligible for many different campaigns, some sort of rule-based filter may in fact be necessary to limit the number of score calculations during each step in an interaction.

Still simpler methods of choosing among campaigns are also employed, and may even be appropriate when meaningful campaign values cannot be calculated. For example, systems that display Web page banner ads often look at a pool of messages for which a customer is eligible and select from this pool at random or in a rotating sequence. These systems often support general business rules such as the number of times a message can be displayed to a customer, the number of days each message is active, and the interval to wait before showing the same message again. Other approaches select campaign messages based on their scheduled delivery date–choosing whatever message is due soonest. This assumes the campaign system generates a list of future messages, which only some of them do.

Personalized messages also often vary depending on the medium in which the message is to appear. Ideally, this would be based on a measure of the effectiveness of different messages in different media. For example, a complicated offer that can be explained in detail on a Web page may be unsuitable for the one-line display of an automated teller machine. Similarly, some media can easily present multiple offers simultaneously while others cannot.

Few, if any, systems explicitly consider medium in their message selection mechanism. But separate versions of the same message must generally be prepared for different media, so marketers can implement a crude form of medium-based differentiation by simply not creating message versions for some media. Much of the value of this strategy depends on what happens when a message version is not available for the current medium. If the system goes back and selects the next-most-valuable message, there is a reasonable chance that something approaching maximum value will be achieved. But if the system simply presents a standard default message–or no message at all–the value is considerably less. Of course, the worst possible situation is if a missing message causes the system to crash altogether. Given the inherent complexity of managing an advanced personalization process, it’s essential that missing versions be handled smoothly and automatically. Of course, all this is a poor substitute for explicit measurement of message effectiveness–in a given situation, even a marginally effective version of the right message may be more valuable than the alternative.

This all assumes a central system is choosing messages to be delivered in different media. In fact, many systems today are limited to a single touchpoint such as Web pages or call centers. These systems often are highly sophisticated in some of their personalization abilities, but of course cannot coordinate messages across media. At best, it’s up to the marketers in charge to create rules and messages for different media that will generate similar outputs in similar situations. But the administrative effort required for such an approach makes it doomed under all but the simplest circumstances. This is a particularly important issue for systems that service just one side of the Internet/non-Internet divide. Unless customers conveniently limit themselves to one side or the other–which most won’t, given the option to use both–maintaining consistent customer management policies will likely prove impossible.

Channel workload is still another factor that some personalization systems consider. The most common examples are inbound telephone call centers, where managers routinely turn off options such as cross sell recommendations during busy periods to reduce the time spent on each call. Ideally, the system itself would decide the degree of personalization used for each call, balancing the value of maintaining a high service level with the value expected from extending a particular interaction. It’s quite likely that an accurate calculation would show that the value of personalization in some calls–say, where there is a very high probability of a lucrative incremental sale–outweighs the loss in revenues from other customers. As usual, a precise on-the-fly calculation is hard to develop but rules can provide a reasonable approximation. This is why elite members of frequent flyer programs have a special phone number for reservations.

Personalization can also be balanced against service levels at retail point of sale, kiosks, automated teller machines, and even high-traffic Web sites. Outbound marketing campaigns sometimes take into account other channel capacity issues, such as the number of leads that can be handled by a field force. In this case, the system has to determine which leads will yield the greatest value and automatically assign the rest to alternate channels such as telemarketing or email. This can be a tricky problem, since there may be other constraints such as target sales levels for individual products or a required mix of business for individual salespeople or dealers. In non-real-time situations, the system must also consider future capacity–it may make sense to hold off sending some leads to telemarketing today, if there is a high probability that a field sales person will have time to handle them tomorrow.

Finally, personalization systems must also adjust to other business considerations such as inventory positions, demand levels, and special promotions. For example, it makes little sense to cross sell a product in short supply–profits will be higher if the company offers other products and saves the product in limited supply for customers who request it by name. Conversely, it makes considerable sense to push a product that is overstocked, since it otherwise may end up as excess inventory. Airline yield management programs are one example of promotions that are dynamically adjusted to market conditions, although these are generally not tailored to individual customers. Better examples are programs that offer discounted tickets on excess seats to customers who have indicated an interest in flying to a particular city. While the details vary depending on the exact nature of the business consideration being accommodated, the general principle is that the personalization system needs to access information about that consideration and somehow factor this information into its value calculation.

By now, a few points should be clear. The first is that even though rule-based approaches are by far the most common today, they cannot manage the complexity of a system that considers all the listed factors–customer, current interaction, campaigns, medium, channel workload, and business constraints–in its personalization decisions. The only practical solution is a value calculation that produces a simple ranking of alternatives. While current value calculations are admittedly imperfect, they at least have the potential to be improved. By contrast, the rule-based approach is ultimately a dead end.

The second point is that there is no fundamental distinction between the personalization requirements for real-time, interactive systems and for outbound, batch-oriented campaigns. Both must make decisions based on a view of the entire customer relationship; in fact, each must take into account actions by the other. Similarly, there is no fundamental difference between “proactive” systems that scan transactions for opportunities and “reactive” systems that guide interactions initiated by the customer. Even though different technologies are needed to execute personalization through the different types of systems, they must share the same core decision functions.

A third point is that personalization systems must include mechanisms to help marketers refine their business strategies. These will range from standard test capabilities such as random sampling and champion/challenger comparisons, to automated data analysis, relationship discovery and model creation. They also include extensive storage of historical data, to allow the reconstruction of the conditions under which each interaction took place. Many of today’s personalization systems are very weak in these areas. But personalization cannot meet its promise without a systematic process to improve decision making. And as a practical matter, it will be virtually impossible to deploy such a process unless it is built into the personalization system itself.

The final point is that personalization systems will evolve into specialized products that are distinct from the operational systems which deliver the personalized messages. This is a substantial change from today’s situation, where most touchpoint systems provide their own personalization capabilities. Core capabilities of the new breed of personalization systems will include access to data from multiple sources, sophisticated value calculations, business strategy optimization, and communication with message delivery systems. Primitive versions of such solutions are beginning to appear, but it will be some time before a complete implementation is available.

* * *

David M. Raab is a Principal at Raab Associates Inc., a consultancy specializing in marketing technology and analytics. He can be reached at draab@raabassociates.com.