User-to-user recommendations within a large online retailer’s recommendation system generated a very small share of sales. With keen business sense (see Netflix) and with praiseworthy regard for the common intellectual good, this retailer allowed some experts to analyze freely a large database of its users’ recommendations and to publish publicly their analysis. The database includes all recommendations that the online retailer’s users made for books, music, and movies from June, 2001 to May 2003.
The online retailer’s recommendation system worked as follows:
Each time a person purchases a book, music, or a movie [DVD or video] he or she is given the option of sending emails recommending the item to friends. The first person to purchase the same item through a referral link in the email gets a 10% discount. When this happens the sender of the recommendation receives a 10% credit on their purchase. [1]
The database includes persons who purchased and made a recommendation, and persons who received a recommendation. The database does not include persons who made a purchase but neither made a recommendation to another person nor received a recommendation from another person.
Purchases that generate recommendations generated few recommendations. Persons who purchased a book and recommended that book made on average 2.0 recommendations per purchased book.[2] Notice that this statistic by definition can be no less than 1: purchases that produced zero recommendations were not recorded in the dataset. To estimate the average number of recommendations per purchase, one needs an estimate of purchases that did not produce a recommendation.
Only a small share of purchases generated a recommendation. Persons who received a recommendation and then purchased the recommended book forwarded the recommendation to another person in about 24% of such purchases.[3] Imitation and concern for social norms have a pervasive and powerful effect on human behavior. The share of purchases in which a person who did not receive a recommendation for a book, but nonetheless purchased it and recommended it is surely less than 24%. That implies that the total number of books purchased per year was higher that 6.1 million, probably much higher. Two plausible figures are 100 million and 30 million.[4] These figures imply shares of purchases that generated a recommendation at 1.4% and 4.9%, respectively.
The overall ratio of recommendations to purchases is much lower than 2. Suppose that the retailer was selling 100 million books per year (the results are qualitatively the same if the retailer was selling only 30 million books per year). Given 2.0 recommendations per book purchase for purchases that produced a recommendation, the overall ratio of recommendations to purchases is 0.028. Making a recommendation created the possibility of a 10% credit for the recommender and a 10% discount for the receiver. Nonetheless, relatively few recommendations were made.
The share of purchases that resulted from user recommendations is miniscule. If the retailer was selling 100 million books per year, then less than a tenth of one percent (0.04%) of purchases followed from users’ recommendations. User social networking through a systematized recommendation system wasn’t a major driver of sales.
Nonetheless, the user recommendation system may have net positive value to the retailer. About 43,000 book purchases per year can plausibly be attributed to user recommendation of books through the retailers’ system.[5] Suppose that the average book price was $30 and the average gross margin was 60%. Then the user recommendation system generated a gross margin of about $800,000 per year. That might be sufficient to make it a profitable feature.
Recommendation systems have major effects on sales. One unsourced report indicated that “35 percent of [Amazon’s] product sales result from recommendations.” Greg Linden, who should know, stated (see comments), “Personalization was responsible for well more than 20% of sales when I left Amazon in 2002.” Automated recommendations probably account for most of the sales through recommendations.
A trade-off between communicative control and potential social effects is an important aspect of social networking. Commentary on the recommendation analysis has largely neglected this issue (for relevant discussion, see here, here, and here, for starters). Being personally responsibility for an online retailer sending a specific purchase offer to a social connection has some social meaning that a potential sender might prefer not to evaluate, and in any case the user cannot change the message sent. The social diffusion of given names, and business successes that arose through social networking, such as Hotmail, Google, MySpace, Youtube, and others, depended on more loosely structured forms of communication.
[1] See Leskovec, Jurij, Lada A. Adamic, and Bernardo A. Huberman, “The Dynamics of Viral Marketing,” p. 3.
[2] See Leskovec, Jure, Ajit Singh, and Jon Kleinberg, “Patterns of Influence in a Recommendation Network (pdf), Table 1. The total number of book purchases was 2,859,096 over the 711 day period.
[3] Leskovec et. al., “Dynamics of Viral Marketing,” p. 8, Table 3.
[4] See Brynjolfsson, Erik, Michael D. Smith, and Yu (Jeffrey) Hu, “”Consumer Surplus in the Digital Economy: Estimating the Value of Increased Product Variety at Online Booksellers” (pdf), Management Science, v. 49 n. 11 (Nov. 2003) p. 1587, and Sandoval, Greg, “Amazon Losing Ground in Core Area: Books,” CNet News.com (Nov. 5, 2001).
[5] Leskovec et. al., “Patterns of Influence,” Table 1 (figure annualized).
2 thoughts on “systematized personal recommendations”