Cloning Your Best Customers (1,173 words)
For many years, companies like Claritas, MicroVision, Trans Union and the Polk Company have offered cluster systems, which define customer populations by demographically-defined lifestyle or lifestage descriptions. Developed using statistical analysis of census data and other sources, these systems became quickly popular and widespread in marketing research circles. Terms such as "dinks," "yuppies" and "empty-nesters" became popular marketing lingo.
The problem with this method of characterizing a customer base is it ignores the fact that households are not cliches but have their own individual lifestyles. This ZIP+4 clustering, while interesting, does not allow for individual distinctions and therefore households can not be counted on to behave to specific marketing strategies with any reliability.
If finding clones of your best customer is the goal, it is essential to understand who the best customers are and what characteristics those customers share which make them unique to your database. As we have found over the years, customer databases are not homogeneous. They are made up of many unique clusters, or groups, of various sizes within the database.
For those trying to get a handle on their customer profiles for the first time, it's important to start with simple segmentation techniques. Begin by identifying what critical data elements are essential to do this fundamental segmentation and profiling. Here's what is needed.
• Clean customer addresses which have been brought up-to-date by NCOA (National Change Of Address). This is essential for the appending process.
• Transaction data that has come from the General Ledger system. This data must be associated with a specific customer record. Transaction data allows companies to segment their customer base into best, worst and average customers. Often times a customer record may not be "householded." In other words, few contacts may be associated with the transaction, leading to unclear information about who the real purchaser is. Unassociated records should be set aside and not used in your profiling.
• Promotion, source, customer satisfaction or feedback. Information from customers as to what their attitudes are toward the company, where they came from or what promotion they responded to are important database elements.
• Household characteristic data. This will be used in the prediction of future behavior. Household data by itself is not going to predict future purchase behavior but in conjunction with historical purchase behavior adds a strong methodology for predicting future purchase behavior.
This method of combining data to the single record element does very well when you know a customer over a period of time. But what happens to those new customers or prospects that you have acquired from other sources or those who have made only a single purchase?
Note that one of the most predictive elements, "purchase history," is not there. These records are put aside until after customer profiling has taken place.
SEGMENT YOUR HOUSEFILE
The next step is to segment the customer records that are the most complete. Start with a sample of the records across various product, transaction and promotion groups. Include current and past customers, which are profitable, and those which are not. An Nth number will do; send these names out to an organization such as Acxiom Corp., Metromail/Experian or Polk for appending all manner of household elements that they may have to help profile the customer and prospect base. This sample record set keeps the cost down while allowing the analyst to work with a relatively small and manageable record set for the next phase of statistical analysis.
When purchasing external data to overlay on your file, an immediate question is which data elements to use. Household data is generally expensive, and characteristic data elements, which are not predictive, are valueless for profiling or predicting behavior. In some cases, certain data elements may not be readily available on most of the file.
It's also important to remember that data validity is sometimes questionable. Inferred or enhanced data can produce inaccuracies. Data entry errors or self-reported data from a customer increase the chance that the segmentation process will not be as accurate as possible.
Using tested statistical techniques will result in the sample set segmenting themselves into "like" groups with similar household characteristics. These groups may be used for further profiling. New suspects entering the database may be measured by "key" household characteristics. The characteristic sets can be used to go back to the main database for profiling; appending those key predictive data characteristics to the entire customer database. From this set of predictive characteristics, a modeling score can be created to score customers or prospects even though there is an absence of any significant purchase history.
SCORE THE DATABASE
The next step is to rank the database according to household characteristics by matching the most similar characteristics to those prospects. Take what was discovered about the key characteristics within the various groups and create a scoring method and algorithm that will be used to score all the records in the database.
After identifying those with the highest score, it's advisable to test the hypothesis by employing a number of staged and balanced marketing efforts. Direct your marketing efforts to each segment, including a control group which receives nothing but which will be monitored and measured and compared to the overall effect of the campaign(s).
From the tests conducted, and by comparing the results to the segmentation and profiling strategies, you can determine which group responded best. Revisit that profile within the responding customer segment and validate all the key characteristics that proved to be potential predictors.
FIND "LIKE" NAMES
After the house database has been exhaustively mined for those "best" customers, it is time to try conquest strategies from outside databases. Remember, a house list will always be more predictive and responsive than a compiled list.
Reverse the strategy with purchased lists. Select the vertical lists that claim to have the same purchases or interests; it is unlikely that these people will respond in the same way. Ask for a sample of each proposed list. Test its deliverability by running it through NCOA processing; a 10-percent failure rate is acceptable. Then append those "key" household characteristics discovered in the profiling phase to the list. Examine it again: Does this list compare favorably to the findings in the profiling?
If the list appears to have similar key characteristics, it's time to test the list with a marketing communications strategy that worked with the profiled house database. If the list does not respond with an acceptable response level, find another list vendor or mine more deeply until the vendor finds a segment that delivers the closest response levels.
Looking ahead, the future of personalized marketing may come from the Internet, as filtering software, intelligent agents and personal shoppers offer new promise. Marketers will get instant feedback, near perfect responsive transactions and deeper insights into the habits, feelings, likes and dislikes of their most elusive, loyal customer.
Bob McKim is partner in MS Database Marketing, Los Angeles, CA. For more information, visit the MS web site at www.msdbm.com, or call (310) 208-2024; fax: (310) 208-5681.