Randomization Method
Prior to the experiment, we will randomize firms into control and treatment groups. We begin with all firms subscribing to Shouqianba's service and apply the following eligibility filters before randomization. First, we require that a firm open the Shouqianba app at least once during the week immediately preceding the experiment to ensure the merchant’s engagement with the platform. The condition is analogous to McKenzie and Woodruff (2017), who restrict their sample to actively operating firms at baseline, since their treatment (i.e., management consulting) requires an active firm to receive it. Our treatment, which is peer information deliver, similarly requires an engaged merchant to be exposed to it. Second, we require firms to have at least 5 peers within a 3 km radius, ensuring a stable and reliable peer benchmark. Finally, we exclude firms with extreme performance, those earning more than 10 times or less than 1/10th of their peer median. All filters are based on the week immediately prior to the experiment.
We conduct block randomization within each stratum to assign 100,000 merchants into equal-sized treatment and control groups. Strata are defined based on the level of competition. Randomization is conducted at the merchant level within each stratum.
We define competition strata based on the number of firms operating in the same secondary industry category within a 3-kilometer radius. We use 3 kilometers based on evidence on short-distance urban mobility in China. Survey-based studies of retail catchment areas find that Chinese shoppers at neighborhood-scale retail centers are drawn overwhelmingly from within a 2–3km radius (Wang, Zhang, and Wang 2006). Consistent with this, e-bikes (the dominant mode for short local trips in Chinese cities) average round-trip distances of roughly 5–6km, implying a comparable one-way local travel radius (Weinert, Ma, Yang, and Cherry 2007). We classify markets into high-, medium-, and low-competition based on the distribution of the strata. This stratified randomization serves two purposes. First, it guarantees that the treatment and control groups are comparable in their competitive environment, preventing competition intensity from confounding our estimates of the benchmark information effect. Second, it provides the variation we need to test our hypothesis that the strategic response to benchmark information differs across market competition levels.