How to acquire customers on the Web.

Most retailers on the Web spend more to acquire customers than they will ever get back in revenue from them. Many think that sky-high spending on marketing is necessary to stake out their share of Internet space. But is it really? How do retailers know how much to pay? Consider CDnow, which has developed a multifaceted customer-acquisition strategy that reflects a clear understanding of the economics of an on-line business. At the heart of its strategy is affiliate marketing, a concept the company pioneered. Under its BuyWeb program, anyone can put a link to CDnow on his or her Web site, and if a customer uses that link to arrive at CDnow and make a purchase, the referring site owner gets a percentage of the sale. CDnow pays no money if no sale is made, which makes the marketing program completely efficient. But CDnow didn't stop there. Being a Web store, it had complete data on the number of visitors to its site and what they bought, which it used to work out the lifetime value of an average customer. CDnow used that figure to determine how much to wager on the expensive and risky world of traditional advertising to reach a wider audience that wasn't already on-line. CDnow's experience, still a work in progress, contradicts John Wanamaker's oft-quoted lament: "I know half the money I spend on advertising is wasted, but I can never find out which half." As the CDnow example demonstrates, there is a way to find out which half really works.

Residual Scaling: An Alternative to Correspondence Analysis for the Graphical Representation of Residuals from Log-Linear Models.

The method of correspondence analysis, applied to a contingency table, provides a graphical representation of departures from the independence model. Generalized correspondence analysis has been proposed as a way of graphically representing departures from models other than independence. However, generalized correspondence analysis does not necessarily decompose a chi-square statistic for departures from non-independence models. We propose a method, called residual scaling (RESCAL), that does decompose chi-square and can also be used for decomposing the difference between any two log-linear models. The decomposition is represented graphically for ease of interpretation. RESCAL analysis results differ in interpretation from correspondence analysis results, although the differences have some advantages over correspondence analysis. RESCAL is demonstrated empirically with data concerning earned doctorates conferred in the United States from 1960 through 1982.