David Raab Article Archive

Response Attribution
David M. Raab
DM Review
May-June, 2004
.

Response attribution is one of those pesky details that can get in the way of grand visions. The vision in this case is optimization of enterprise marketing programs, which means ensuring that marketing resources are allocated to the most productive possible uses. This implies a vast, all-knowing marketing management system that identifies all spending options, attaches a value to each, and chooses the most effective. It sounds great in theory, and it’s even possible to build process models that can simulate the results of different strategies. But how do you measure the actual results of different choices, so you can assess and refine your assumptions? That’s where response attribution comes in, and that’s when things get ugly.

The problem is that marketing systems are not in fact all-knowing. (You read it here first.) Response attribution is the process of determining which marketing activity caused a customer action. At the most basic philosophical level, causation is always problematic. But marketing adds its own wrinkles to this fundamental mystery of existence. When individuals are exposed to multiple offers for the same product, which gets credit for the purchase? How should attribution account for exposures that are not tied to a specific individual, such as broadcast television advertising? How can it include consider the impact of non-marketing interactions such as purchase, product usage and customer service experience? Should it take into account competitive and environmental influences as well?

Response attribution was originally applied to direct marketing promotions. In these situations, each response is tied directly a specific individual, and most promotions can be linked to individuals as well. The simplest case is direct mail, where marketers typically print a promotion identifier on the order form that will be returned by the customer. This identifier, called a source code or key code, is recorded when the order is processed and thus the connection is made. Slight variations are possible with media where there is no physical order vehicle, such as telephone orders. These include having the customer read the source code aloud when placing a telephone order and using different toll-free telephone numbers for responses to different television advertisements.

But complexity mars even the Edenic simplicity of direct marketing. When a customer receives several catalogs offering the same product, does it make sense to credit the response to just one? How to treat pass-along orders from people who were not sent the original promotion? What about orders that cannot be linked to any particular catalog? Should consideration be given to what might have happened if a promotion had not been sent–assuming, for example, that a fraction of magazine readers would have found a way to resubscribe even if renewal offers had not been sent?

In practice, direct marketers can generally come up satisfactory answers to these questions. The answer will often depend on the purpose at hand: deciding how many catalogs to send to one person needs a different analysis than determining whether one direct mail letter is better than another. The topic is more important than debating how many angels can dance on the head of pin, but the discussion can sometimes feel equally subtle.

Outside of direct marketing, things get complicated fast. Retailers may have customer lists and even use them to send catalogs, but connecting these to in-store purchases ranges from difficult to impossible. Just identifying the customer is a challenge: retail transactions involving cash or even credit cards (depending on privacy rules) can be effectively anonymous. Many merchants have loyalty programs largely to give customers an incentive to identify themselves at time of purchase. Even when the customer is known, what promotion they are responding to may not be clear. One common approach is to send coupons marked with both customer and promotion identifiers, which are presented at time of purchase. But this can be expensive and many customers will not redeem the coupon even if the promotion did in fact influence their behavior. A more universal solution relies on inferred response: if the customer received an advertisement for an item and later purchased it, the marketer assumes a connection. This is usually implemented through queries that compare lists of customers with purchases of specified products in specified time periods. The products are those included in the promotion; the time period is the range during which the promotion is assumed effective. Often marketers look at several different queries, incorporating different date ranges and groups of products, each implying a different definition of response.

Companies selling through distributors, retailers and other third-party channels have an even harder time collecting customer purchase data. Warranty cards are one common source of information for manufacturers, but coverage is far from complete. Other consumer goods manufacturers often rely on product movement data or third-party surveys to get some idea of sales changes and how these correlate with promotions. Of course, such data rarely identifies individual purchasers and is affected by many factors other than the company’s own promotions.

Next month’s article will describe new challenges in response attribution, and how today’s marketers are responding.

**********************************************************

Last month’s article discussed the basic challenge of response attribution, which is linking customer behavior to marketing activities. From an information management perspective, the basic questions are whether the system can identify (a) the people who receive the marketing message and (b) the people who respond through actions such as making purchases. Since each question can have a yes or no answer, there are four possible situations:

– known receiver, known responder: this is the most favorable situation. Since the receivers are known, their long-term behavior can be compared with non-receivers to measure the full impact of the promotion. Ideally the responses carry an identifier, such as a source code, that links them directly to the original promotion. Otherwise, so long as responder names and addresses are captured, it may be possible to match the responder list against the original promotion list. Common examples are direct mail, catalog and targeted emails with a response mechanism.

– known receiver, unknown responder: this happens when promotions are sent to specific individuals but the responders either cannot be identified at all or cannot be connected to the list of receivers. Common examples are mailings that drive traffic to retailers or dealers. These often use anonymous response devices, such as non-personalized coupons redeemed at the store, to measure results without linking them to individuals. In such cases, analysts can measure the volume of sales directly tied to the promotion but cannot include purchases by promotion recipients who did not redeem the coupons. Marketers may still be able to compare over-all behavior of recipients vs. non-recipients, but only if there is some other mechanism to tie sales back to individuals.

– unknown receiver, known responder: here the marketer doesn’t have the list of people who will see the promotion, but still gets the names of responders. Examples include any type of direct response advertising through mass media such as TV, radio, newspaper inserts, billboards and Web sites. It’s often possible to identify the promotion the responders are reacting to, through codes printed on advertisements or by using different telephone numbers or addresses for different advertisements. Such information measures immediate advertising results. But it does not capture effect that seeing the advertising had on subsequent behavior. This cannot be measured without knowing who received a promotion even if they didn’t react immediately.

– unknown receiver, unknown responder: this is the classic situation of marketers selling through retail channels. They advertise in mass media such as print or broadcast, which means they have no list of the receivers. Sales are then made to essentially anonymous consumers, either because their names are literally unknown or because the retailer or dealer will not share them. Marketing results can only be inferred from changes in total sales after the advertising is executed. Even this crude inference is possible only if marketers hold back advertising to another, roughly comparable group so differences in results–presumably attributable to the advertising–can be measured. Because mass media are inherently non-selective, such tests are frequently done by comparing different geographical markets–a very blunt approach, and one that is increasingly unreliable given today’s eroding barriers between local and national media.

Figure 1 summarizes these situations:

The more closely responses can be linked to recipients, the more precisely the impact of specific advertising efforts can be measured. Over time, such measurements enable marketers to learn how to improve their results. Thus, one strategic objective for marketers is to shift their efforts to techniques that allow better measurement. For example, adding a loyalty program lets an airline not just reward its most valuable customers, but also assess the results of different promotions on long-term behavior. These promotions may be alternative reward schemes or incremental revenue generators such as incentives to purchase seats on lightly booked flights. Even small improvements in these areas can have material impact on airline profit margins.

Other techniques to make responses more measurable include substituting personalized coupons for anonymous ones, using tracking devices such as cookies or tracking codes on Internet promotions, and capturing caller ID information from in-bound telephone contacts. New technologies such as RFID tags may provide still more identity information, although privacy concerns may limit their application.

Some of these techniques are new because the underlying technologies are new. Others have long been available but have become more practical as costs have fallen. So one new element in response attribution is the wider variety of measurement techniques available.

Another element is the greater mix of channels. Internet marketing places an increasing role in many businesses. Other channels, such as direct mail and telephone marketing, are also being added by many firms that originally sold elsewhere. Thus organizations must now measure responses across multiple channels to assess the full impact of a particular promotion. This makes it even more important to identify specific promotion recipients and responders.

Marketers are also increasingly concerned with long-term relationships, which can only be tracked by linking multiple interactions. This means that the gold standard of traditional direct marketing–capturing a response code that links each order back to the original promotion–is no longer adequate. Instead, marketers need to assemble a comprehensive history of all promotions and responses for specific individuals.

Building such histories is the role of a marketing database. Such databases started appearing twenty years ago, so the basic design principles and technical challenges have long been known. But even today, most marketing databases serve primarily to simplify creation of customer lists for targeted promotions. Response attribution is still mostly conducted with an eye to linking specific responses to specific campaigns.

This is starting to change as marketers come to grips with the fact that long-term relationships cannot be analyzed as a series of unconnected interactions. They must abandon the traditional goal of response attribution–assigning a specific cause to a specific event–and adopt a broader approach organized in terms of customer treatment policies. These policies govern large numbers of interactions over long periods of time. Their results are measured in terms of over-all customer behavior, not specific events. In fact, because the behavior of a single customer is not statistically significant, the scope of analysis actually shifts from single transactions all the way to customer groups.

Conducting effective analysis at the customer group level requires new techniques for data gathering and analysis. Data gathering must track customers over longer periods and must capture the policies in effect at specific points in time. Analysis must employ sophisticated statistical methods to infer the impact of different policy components. Both the data and the results are likely to be less precise than traditional response attribution techniques. But the business benefit–truly measuring and ultimately optimizing long-term customer results–makes the adjustment worth the pain.