4.2 Localización del proyecto
4.2.2 Micro localización
We conducted a large-scale randomized field experiment in collaboration with a major European e-commerce company to investigate the effectiveness of different levels of personalization specificity in social retargeting. Our partner company sells a wide range of products with a focus on consumer electronics. For our study, we focus on the product categories of laptops, cameras, tablet computers, smart phones, and televisions. For the experiment, we solely advertise to consumers in the newsfeed area of Facebook as the newsfeed is generally the focal area for consumers and captures most of their attention (Wishpond, 2014).
Consumers that browsed the partner company’s website, viewed at least a category- level page, and were active users of Facebook, were eligible to participate in our experiment. Using their browsing behavior, we randomly assigned either category- or product-specific personalized social retargeting ads to these consumers. The random assignment to the two personalization treatments took place on our partner company’s website by assigning one of two conditions to consumers’ Facebook pixels (a cookie stored on consumers’ computers that can be read by Facebook). Consumers that then visited Facebook were treated with ads matching this assignment. This assignment method offers an advantage over conventional cookie targeting. Once consumers reach Facebook’s website without deleting their cookie, they are allocated to their assigned treatment group. Facebook stores this assignment linked to a consumer’s user account. This way, consumers remain in a treatment group even if they delete their cookies after reaching Facebook. If consumers delete their cookie before reaching Facebook, they are not addressed with advertising and remain eligible to participate in the experiment in case they re-visit our partner company’s website and receive a new, independent assignment to a treatment group. Additionally, we address the hypothetical case that
2.3 Field Experiment 25
consumers are assigned to both treatment groups, i.e. because of technical issues, by excluding consumers with a double assignment from the experiment. This way, we can guarantee a clean between-subject design for our personalization treatments.2
Throughout the experiment, consumers remained in their respective treatment groups.3
We operationalized personalization specificity by displaying ads that were related to either the last visited product category (category-specific) or the last visited product (product-specific). We made sure that the two types of ads were exactly the same besides the product and category attributes as shown in Figure 2.1. In contrast to former studies that, based on the chosen personalization algorithm, displayed several products to consumers simultaneously, we only advertised a single category or product per ad. This way, we aim to isolate the effect originating from category- and product- specific personalization and rule out alternative explanations originating from the difference in visual appeal or confounding factors originating from the composition of choice sets that are presented to consumers in ads with several products. Category ads displayed the three most popular products within a product category (in terms of sales) in a single ad image.
By default, ads were socially targeted when consumers were via one or several friends connected to the Facebook page of our partner company. Socially targeted ads displayed the name(s) of the friend(s) that liked our partner company’s Facebook page by stating “[Friend Name] likes [Company Name]” at the top of the ad (see Figure 2.1a). Generally, Facebook’s advertising algorithm displays friend connections to the advertising firm whenever possible. This means that friend connections need to be present and the friend that is supposed to appear as an endorser in the ad has not withdrawn Facebook’s right to use her name for advertising purposes in her account settings (Tucker, 2016). Notably, our social targeting operationalization does not represent an experimental treatment variable but rather a consumer characteristic, i.e. being connected to the advertiser’s Facebook page, that is used by the advertiser to target consumers and which is made explicit in the ad text.
We ran our field experiment for 28 consecutive days in May 2015. Overall, our exper- iment generated 3,476,626 impressions for 198,234 individual consumers. Consumers
2We further discuss the limited potential of contamination through social interactions with ads
in Appendix A2.1.
3We run all ad campaigns with the same budget restrictions, the firm’s willingness to pay per
1,000 impressions (CPM). This way the ad platform, Facebook, has no incentive to select different types of consumers into the ad treatment groups when being paid per impression. This differs from campaigns that are optimized based on consumers’ propensity to respond as common in cost per click (CPC) or cost per acquisition (CPA) optimized campaigns. We do not detect a systematic difference in the costs per impression for consumers in the two ad treatment groups.
Figure 2.1: Experiment Conditions
were shown a maximum of two ads on a daily basis. We measure the ad effectiveness using both clicks and purchases. Clicks measure how many times consumers have clicked on a social retargeting ad. Purchases indicate how many times consumers have purchased from our partner company within 28 days after clicking an ad. The Facebook ad reporting tool does attribute purchases to a consumer’s last clicked ad impression before the purchase. The ads generated 25,577 clicks, leading to an overall average click-through rate of 0.736%, and 1,070 purchases (within 28 days after having clicked on an ad), resulting in an average click-to-conversion rate of 4.183%. Consumers were excluded from the experiment after conducting a purchase with our partner firm to avoid serving consumers ads of products they had already purchased.