Display advertising’s competitive spillovers to consumer search

Lewis (2010) uses this variation to study the effect of frequency on clicks and new account sign-ups.

This is a burgeoning research area—since we released this work in 2011, others have begun to work on similar research.

Blake et al. (2013) found little short-term value of brand keyword advertising for eBay because those who clicked on eBay’s search ads and later made a purchase would have clicked on the organic links to the company’s website on the search result page anyway. Hence, customers use brand keywords to find brands’ websites.

Lewis (2010) discusses this natural experiment and its virtues and failings in depth.

Only 18 % of impressions were followed by another impression from the study within ten minutes.

This method incorrectly attributes some searches to the brand’s product such as queries for the advertiser’s financial statements or stock price. Additionally, ambiguous search terms such as ‘apple,’ ‘galaxy,’ and ‘progressive’ complicates the analysis of queries. After reviewing some of the categorized searches, we believe that most of the categorized searches are for the brands’ products. See Table 4 for examples of affected brand-related queries.

We do not analyze the Best Buy campaign for competitive spillovers because by 2011 Best Buy was the only nationwide electronics retailer remaining after the demise of Circuit City and many other stores.

We present this simple analysis rather than a linear regression with control variables because the exogenous variation of the natural experiment removes the need for control variables to identify the effect of the display ad and reduces the linear regression to a comparison of means between the test and control group. Our analysis is equivalent to estimating the following (saturated) linear probability model where S e a r c h _{i j t} is an indicator variable equal to one if user i searched for brand j during period t and zero otherwise; A D _{i t} is another indicator variable equal to one if the target ad was delivered to user i during period t; 𝜖 _{i j t} is the error term.

Robustness checks examining the effects of the number of prior target exposures during the control periods showed no statistically significant effect, with the point estimates generally negative, suggesting no misattribution. These results are available in the Online Appendix.

Any diminishing returns to additional impressions (i.e., wearing out) of the effects, which could complicate this simplified counterfactual computation, are found to be of modest importance in the Online Appendix. Similar to the findings in Lewis (2010) for clicks and new account sign-ups and in Johnson et al. (2015) for combined online and in-store sales, our model shows no significant effects of decreasing marginal returns, but the confidence intervals grow wider in the number of impressions.

Nearly all users who searched for either the advertiser or at least one of its competitors did not search for both.

The search queries were already “canonized” in the Yahoo! Search raw data: user-entered queries are transformed via many grammatical and data-driven rules such as correcting common misspellings, trimming pluralized words ending in ‘s’s, and removing articles and common words such as ‘the’ and ‘out.’ As a result, a query such as ‘FX Lights Out’ becomes ‘fx light’ in the data.

All statistically significant relevant outcomes show positive effects—we include potentially negatively affected outcomes under the control ads’ significant terms in Table 4 because, by design, any negative effects of the target ad are observationally equivalent to positive effects of the control ad.

An unpublished study by one of the authors in 2010 found that showing display ads on both mobile and desktop platforms to the same individual was very difficult due to the endogeneity of individuals’ platform usage.