Targeted Promotions on an E-Book Platform: Crowding Out, Heterogeneity, and Opportunity Costs

Transcription

1 Targeted Promotions on an E-Book Platform: Crowding Out, Heterogeneity, and Opportunity Costs Nathan Fong, Yu ch i Zhang, Xue m ing Lu o, an d Xiaoyi Wang

2 2

3 What is the Opportunity Cost? Typically optim ize for direct response Exploit similarity to previously purchased/clicked item s 3

4 Targeting and Custom er Search Targeting decreases search activity Web content (Tam and Ho 2006) Product listings (De llae rt and Häubl 2012) Targeted promotions (Fong 2017) Generally seen as efficient (saves effort) Positive spillovers (Anderson and Sim este r 20 13, Sahni ) Targeting could lim it exposure Hypothesis: negative spillovers for dissimilar products 4

5 Field Experiments on a Mobile Platform E -book app Prom otes new titles using push notifications Tapstream data 5 m onths pre-test 3 weeks post-test 5

6 St udy 2 Design Sam ple of Eligible Cust om e rs Prefer Genre D Prefer Genre E Prefer Genre F Genre D Promo Genre E Promo Genre F Promo Holdout Preference defined by cutoff in share of reading 67% cutoff: 19,522 eligible use rs 6

7 Targeting Increases Direct Response (Prom oted Book)

8 Spillover Effects Same Genre: Positive Cross-Genre: Negative

9 Robustness to Targeting Definition Cutoff is arbitrary Does not account for volume Trying a lot of books increases genre share Are genres really categories? Test sensitivity to cutoffs Use continuous measures of fit Target based on completion Use broader categories in product hierarchy 9

10 Source of the Negative Spillovers? Pure substitution? Study 1: Differences in total purchasing Reactance to targeting? Positive same-genre spillovers 10

11 Targeting Affects Depth of Search Unique Titles Inspected Without Purchase Same Genre Cross-Genre 11

12 Targeting Affects Breadth of Search Unique Categories Searched Same Genre Cross-Genre 12

13 Search Depth and Breadth Mediate Crowding Out Effects 1.11** (Ind: 0.66 [0.43, 0.87]) Same Genre Books Searched Same Genre Purchases 0.60** -0.80** Cross Genre Books Searched In d : [-0.65, -0.31] Targeting Effect ** 0.89** Ind: 0.15 [0.09, 0.23] Cross Genre Purchases Same Genre Search Incidence -0.17** 13 Other Genres Searched 1.14 ** In d : [-0.33, -0.09]

14 Quantifying the Opportunity Cost Targeting for direct vs. total response Fit causal tree for each prom otion gg and outcome oo, p re dict δδ oo gg,ii for each user ii Targeted treatment: tt ii oo = argmax gg oo δδ gg,ii The opportunity cost of optimizing one outcome oo in terms of suboptimal result for other outcome pp: 14 ii pp δδ pp ttii,ii pp δδ oo ttii,ii

15 Targeting for Direct Response Exploits Genre Preferences Targeting on Dire ct Response Genre D Share Genre E Share Genre F Share Tot al Sale s Books Finished Search Dispe rsion N No Prom o 13.7% 12.3% 0.2% ,767 Genre D 18.2% 14.8% 0.9% ,14 8 Genre E 6.2% 26.2% 1.7% ,061 Genre F 5.4% 11.9 % 18.5 % ,540 Targeting on Total Sales Genre D Share Genre E Share Genre F Share Tot al Sale s Books Finished Search Dispe rsion N No Prom o 7.7% 53.6% 0.4% ,38 2 Genre D 8.6% 8.5% 1.4 % ,849 Genre E 14.8% 19.9% 2.2% ,0 71 Genre F 9.1% 18.1% 1.6 % ,214 Av e rag e 12.8 % 19.0 % 2.0% ,516 15

16 Targeting Schem es Differ Substantially Incremental Effect by Targeting Scheme % of optim al 22% of optim al Random Direct Response Total Sales % of optim al 29% of optim al Random Direct Response Total Sales Outcome: Direct Response Outcom e: Total Sales 16

17 Conclusions Targeted prom otions can generate negative spillovers for dissimilar products Active search redirects sales Direct response metrics can mislead Purchasing breadth important for: High variety-seeking markets (e.g. content) Retailers with wide assortments 17

18 Targeting on Completion Direct Response Spillove r Effe cts Prom oted Product Same Genre Cross-Genre 18 Note: Sample includes all those completing at least one full book during pre-test period (35,704 users). Targeting based on completing a book in the promoted genre (3,002 users)

19 Targeting on Broader Categories Direct Response Spillove r Effe cts Prom oted Product Same Genre Cross-Genre Note: Targeting based on 10 broader categories instead of 239 narrow genres 19