Dynamic CRM

Improving customer dynamics for a better bottom line

The Soft Sell

The company readily approved some test promotions, as it was keen on finding anything that might have a positive impact.

It approved a number of ideas including discount and dollar off e-coupons, but the management was a strong believer that they were an image-based company: Its customers' loyalty and interest were a result of the quality, style, and the intangibles of settings, places, and feel of the images it projected. This instinct turned out to very important.

But we asked, "How can we promote intangibles?" We came upon the idea of promoting products that the customer was not likely to buy. The idea was that the image a high-end product projected could set a mood, flatter the customer, highlight the company, keep the customer's attention, and might interest the customers in looking further. E-commerce is far behind other commerce in sophisticated soft sell.

So we ended up researching three promotion types:

• Percent discounts
• Fixed dollar amounts off a purchase price
• High-end image advertisements.

One of the many nice things about e-commerce market research is how easy basic studies are to implement. You can readily offer a test sample a promotion, either online or via email, to look for effects. The cost is comparatively low and result indicators are readily available.

The goal was to identify whom each promotion affected and pick out the appropriate target promotions to keep the customer on, or moving up, the ladder.

After selecting the test samples, we offered each of the promotion types to people of each category type. In short fashion, we could start looking at results.

We could see how the presence of any of the promotions influenced each transition (including the transition back to the same category). We now could calculate three added versions of each of the two matrices - each representing the transitions conditional on the presence of each of the promotions. In total, we had eight matrices.

As a visualization tool, we diagrammed the transition matrices with nodes, each representing a customer category with links between each giving the transition probability. We labeled the links in red, green, blue, and black to indicate the transition probability for promotion 1, 2, 3, or no promotion at all, respectively. (Some readers might recognize that we basically constructed context-contingent Markov models.)

The Right Promotion

Right away, we saw that we could improve customer interest with some form of contact. But it was also clear that differing promotions worked better with people of differing status. Although a lot of variability between models and link values existed, a pattern could be seen - the dollar coupon worked best for the lightest customers, the discount for the mid-level customers, and the mood promotions reminded the best customers that they do indeed like shopping at this virtual store.

We realized we could combine all these results into a super model that chose the most promising promotion for each person. We jointly optimized the two models with a simple rule: Each person, although part of at least two categories, would receive the promotion that had the largest impact among any of the groups they belonged. Each person had at least eight contingencies, and they received the promotion that topped the list.

We quickly initiated another market test to quantify the value of the super model. The summary of the results of these studies for the last two test months is as follows:

• The control group without promotions increased by only 3 percent the first month and 2 percent the second month.
• When we randomly assigned three promotions to all members of a test group, we saw an approximately 6 percent increase in each month.
• When we used the models singly to pick the best promotion for a customer group, we saw about a 14 percent increase in each month for each model.
• However, the combined optimization of the two models, where each customer received the promotion identified as the best out of all variations, saw a revenue increase of 20 percent and 24 percent in the first two months, respectively.

Rate This Article Rating: 1 (best) 2 3 4 5 (worst) Comments: Optional e-mail address:

This research clearly shows that dynamic-based customer segmentation is a powerful marketing tool. It also shows that state and transition models can formalize and implement dynamic-based solutions.

Personalization takes a major leap forward when you understand the multiple ways to look at the same customer and that customers' relations to products, brands, or companies can dynamically evolve, just as relations between people. I am sure this is only the beginning.

Barry Grushkin (blg23@cornell.edu) is director of consulting at the Machine Intelligence Co.

RESOURCES

Business Intelligence Info Center: www.intelligententerprise.com/biframe.jhtml

www.IntelligentCRM.com

http://www.intelligententerprise.com