A Model of Individual Keyword Per...
A Model of Individual Keyword Performance in Paid Search Advertising Oliver J. Rutz Randolph E. Bucklin* November 2007 Oliver J. Rutz is Assistant Professor of Marketing, Yale School of Management, 135 Prospect Street, New Haven, CT 06520, email@example.com. Randolph E. Bucklin is the Peter W. Mullin Professor, UCLA Anderson School, 110 Westwood Plaza, Los Angeles, CA 90095, firstname.lastname@example.org. The authors wish to thank a collaborating firm for providing the data used in this study. 1
A Model of Individual Keyword Performance in Paid Search Advertising Abstract In paid search advertising on Internet search engines, advertisers bid for specific keywords, e.g. ���Rental Cars LAX,��� to display a text ad in the sponsored section of the search results page. The advertiser is charged when a user clicks on the ad. Many of the keywords in paid search campaigns generate few, if any, sales conversions ��� even over several months. This sparseness makes it difficult to assess the profit performance of individual keywords and has led to the practice of managing large groups of keywords together or relying on easy-to-calculate heuristics such as click-through rate (CTR). The authors develop a model of individual keyword conversion that addresses the sparseness problem. Conversion rates are estimated using a hierarchical Bayes binary choice model. This enables conversion to be based on both word-level covariates and shrinkage across keywords. The model is applied to keyword-level paid search data containing daily information on impressions, clicks and reservations for a major lodging chain. The results show that including keyword-level covariates and heterogeneity significantly improves conversion estimates. A holdout comparison suggests that campaign management based on the model, i.e., estimated cost- per-sale on a keyword level, would outperform existing managerial strategies. Keywords: Internet, Advertising, Paid Search, Bayesian Methods 2
Introduction Paid search advertising1 allows companies to address consumers directly during their electronic search for products or services. When a consumer searches for products or services with the help of an Internet search engine, he uses keywords (a keyword can consist of multiple words) such as ���Hotels Los Angeles��� to start the search. A company in the lodging business can address this consumer directly via paid search. It ���buys��� specific keywords, e.g., ���Hotels Los Angeles��� and creates a text ad that will appear when a consumer searches for these keywords. The serving of a text ad is called an impression. Paid search ads are displayed in the clearly marked sponsored section of the search results page (see Figure 1). In most cases, multiple advertisers are interested in the same keywords. For example, a keyword such as ���Hotels Los Angeles��� is attractive to almost any lodging company in the Los Angeles area. From an advertiser���s perspective, paid search ads are not like traditional ads. Advertisers cannot buy their place into the listings for a fixed dollar amount. Unlike banner ads, for example, advertisers cannot buy ���web real estate.��� In paid search, advertisers bid what they are willing to pay for a click on a paid search ad. An automated auction algorithm then determines placement and position, e.g., 1st or 3rd, of the ad in the sponsored listings. Paid search is pay-for-performance advertising. Companies do not pay for impressions served, but for actual clicks on their paid search ads. This may be one of the reasons for the rapid growth of paid search over the past several years. Currently, the two major forms of online advertising are paid search and display (or banner) advertising. In display advertising consumers are exposed to ads that appear as banners on a web-page. Display advertising had been the preeminent form of online advertising until 2005 when paid search overtook it. The largest search engine, Google, generated more than $4 billion 1 We will use paid search instead of paid search advertising for the remainder of the paper. 1
in net ad revenue in 2005 with their paid search operation called AdWords (Business Week 2006). Despite the rapid growth and scale of paid search, there has been little academic study of this advertising service, especially in the marketing literature. In paid search an advertiser faces four different decisions (or levers): (1) which keywords to select, (2) how much to bid for each keyword, (3) how to design the text ad and (4) how to design the landing page. Some existing research has focused on decision two: bidding. Game-theoretic in nature, this research does include limited empirical applications (Edelman et al. 2007, Chen and He 2006). To the best of our knowledge, no research has specifically investigated the design of the text ad and the landing page. On the other hand, designing ads and landing pages might not differ materially from traditional ad and web design, suggesting that findings from previous studies might apply. This leaves the first question: which keywords to select? Much of current management practice is to use heuristics or to assess campaign performance on an aggregate level, i.e., group keywords together and manage the resulting groups. Because the mechanism of paid search (including data reporting by the search engines) operates at the keyword level, this raises the question how decisions should be made at that level. Measuring keyword performance is an often difficult problem in paid search because the data are sparse, i.e., for most keywords few or no conversions to purchase (or other desired online action) are observed. Often this is the case even over extended periods of time. Our dataset contains daily information on paid search advertising from a major lodging chain on a keyword-level. A traditional approach to performance evaluation would be to calculate the marginal benefit of spending on each keyword, comparing advertising-related revenue-per- reservation with advertising-related cost-per-reservation2 (CPR). If that difference is positive, a keyword is attractive and the company should retain it. If not, the company should drop this 2 For the remainder of the paper we will use cost-per-reservation instead of advertising-related cost-per- reservation for the convenience of the reader. 2
keyword from the campaign. Alternatively, the company could also consider adjusting its bidding strategy. Search engines, however, do not provide their advertisers with data on the actual auction or competitive bids. We therefore base this work on the data that are made available to advertisers by search engines and used by campaign managers, leaving competitive behavior and auction issues as topics for future research. At a first glance, assessing individual keyword performance on a CPR basis seems straightforward. CPR is given by ConvRate Rate Conversion CPR = . Most keywords have impressions and clicks (costs) associated with them on a daily basis. But while searches lead to impressions, and sometimes clicks, they do not always lead to sales. Indeed, average conversion rates in paid search are very low. In April 2004, for example, only 84 out of 301 keywords (27%) in our data set led to reservations (see Figure 2). Thus, on a daily, or even weekly or monthly, basis most keywords simply did not generate any sales. This observed conversion rate of zero precludes us from calculating CPR. Of course, the true, long-term conversion rate may not be equal to zero for these keywords. But because the rate might be very small, it could take quite a while for a sale to occur. In the meantime, the firm continues to pay for the clicks on its ad without any accompanying revenue to show for it. The purpose of this paper is to develop a method to evaluate the performance of individual keywords based on cost-per-sale or, in our case, cost-per-reservation (CPR). We condition our analysis on a click. This gives us cost-per-click (CPC) information for each keyword. To calculate CPR we need the keyword���s conversion rate. To obtain a model-based estimate for this, we conceptualize conversion as a binary choice decision conditional on a click (i.e., a user has clicked on the paid search ad and has been taken to the landing page on the company���s website). 3