Cluster Bombing

Does this sound familiar? You work for a retailer with a large catalog and Internet business. Your bosses figure out that not all customers are the same when it comes to brand loyalty, and that their purchase behaviors reflect this disparity.

They decide to divide the buyers into manageable clusters. Then they hire a market research firm to survey customers on their needs and perceptions.

And guess what? They put you in charge of the project.

Don't quit (yet). Instead, imagine this scenario:

Under your direction, the research firm develops a questionnaire and shows it to focus groups before mailing it. The purpose? To determine whether the database is made up of relatively homogeneous segments that differ in how they perceive and use different sales channels.

The research will also try to find:

• Other stores your customers shop in.

• The share of a customer's business done in your stores.

• Which departments are or aren't shopped.

• What customers like about your stores, compared with those of your competition.

• Why some items are purchased through one channel and not others.

• How you measure up to your rivals in terms of price, service, appearance, parking, depth of product lines and returns policy.

The research firm runs the survey through focus groups around the country. Assuming they respond well, you send questionnaires to 3,000 randomly selected customers.

Here's where the fun — and terror — begins.

The first question you have to ask is: Which data should be used to build the clusters, and which should be employed to profile these segments?

Here are the options. You can:

• Rely only on the survey answers, and profile the resulting segments based on demographic and behavior data.

• Use demographic data and profile based on the answers to the survey and purchase data.

• Restrict yourself to behavioral data and profile the segments based on demographics and the answers to the survey.

• Use both the demographic and behavioral data and profile based on the answers to the survey.

Got it? And if that weren't bad enough, your higher-ups might expect you to segment the entire database. That's not easy to do.

Remember, the task is to develop a segmentation scheme that can be applied to the entire file. Not everyone in the database has answered these questions — only 3,000 have.

So what do you do? If you take the “classical” approach, building your segmentation around the answers to the survey, you could then create an assignment model. This would use demographic and customer purchase behavior to predict segment membership.

By going this route, you're betting that demographics and behavior will enable you to predict attitudes. If that happens, you can apply the model to the entire database.

Unfortunately, these models don't often work well. That's because there's no logical reason why demographics and behavior should predict attitudes.

Your best option in that case is to return to the surveys and build a segmentation model using demographic data only, and then work up another model based on what you find. Then apply that to the entire database.

You've done it. Now you can profile the demographic segments in terms of behavior and the answers to the questionnaire.

The alternative is to build a model based on behavior, develop an equation to predict segment membership, and then profile these groups based on demographics and the answers to the survey questions.

Or build your clusters based on demographic and behavioral data, then create your predictive model and then profile these segments using the answers to the questionnaire.

Which method is the right one?

I favor starting with either behavioral and/or demographic segmentation, profiling the groups, and then finally developing a research program targeted at them. For example, Sears might discover a large number of younger, wealthier customers who buy tools but not clothes. That might be a cluster worth researching!

---

DAVID SHEPARD is president of David Shepard Associates Inc., a direct marketing and database consulting firm in Dix Hills, NY.

CRM/DATABASE

Does this sound familiar? You work for a retailer with a large catalog and Web business that decides to divide its customers into manageable clusters. A market researcher is hired to get the buyers' feedback. And then they put you in charge of the project. But don't quit (yet).

Who Responds to What

	General population	Emerging Hispanics
Cash back	54.7%	45.4%
Points for free travel	39.1	35.3
Points for in-store redemption	32.6	36.7
Certificates	28.2	40.1
Points for Web/catalog redemption	25.2	30.0
Average hard benefit	36.0	37.5
Discounts	47.9	41.5
Upgrades	20.8	17.4
Members-only information	19.7	23.2
Members-only access	13.7	17.9
Average soft benefit	25.5	25.0
Source: Colloquy's ‘A Comparison of Loyalty Marketing Perceptions Among Specific U.S. Consumer Segments’