Webrooming: how to amplify online ads

Delighted to present yesterday at HBS on the Conference on Artificial Intelligence, Machine Learning, and Business Analytics, I am sharing the key insights with you.

Beyond their role as distribution channels for purchasing products, ecommerce websites offer a large collection of product information, reviews, prices, and similar information which is freely available 24/7. For instance in the US, more consumers start their product search on Amazon.com than on any other website.??However, many advertisers evaluate returns on their eCommerce ads only by on-site purchases; ignoring web rooming; i.e. customers who are researching on-site for subsequent off-site purchases. Quantifying the magnitude of the off-site purchase behavior allows advertisers to account for the impact of their brand presence on the retail website in terms of on- and off-site purchases and use these insights for media allocation and planning decisions.

In this research, German Zenetti, Oded Netzer, Ruitao Cheng and I combine internal data from customers’ browsing behavior on Amazon.com matched with survey responses for the same customers about their offsite behavior to map webrooming across twelve product categories. We find that webrooming is economically substantial: among the consumers who researched a product on Amazon, more end up buying it off-Amazon than on-Amazon - in every analyzed category.?Moreover, consumers who visit more product pages (Detail Page Views), click more and spend a longer time on the product's Amazon page, are more likely to buy the product, both on and off Amazon.

We create two brand-level measures to quantify the magnitude of off-site purchases among on-site researchers: Research on-site purchases elsewhere share (ROPES) and Amplifier Ratio (AR).?

ROPES is the number of off-site brand buyers divided by the number of on-site category researchers.?For instance, if there are overall 5,406 customers that researched on-site for laptops in a certain time period but did not buy on-site, and 135 of them ended up buying a laptop from the brand A1C off-site, the share of off-site buyers (or ROPES) of brand A1C is 135/5,406 = 2.5%. This gives an estimate of how many customers were or could have been influenced by brand A1C via product and brand information and on-site advertising. In other words, brand A1C should consider taking these customers into account when making on-site advertising decisions. A brand may wish to maximize the size of their ROPES metric, because it represents how much of the category potential the brand captures. This is similar to a ‘share-of-the-aisle’ metric in physical supermarkets: of all the consumers going down the aisle, how many looked at or bought the brand??

The Amplifier Ratio (AR) captures the ratio of the brand’s off-site to on-site buyers. If a brand or an advertiser is using a return on advertising spend (RoAS) measure that is based on on-site purchases, AR can be used as a proxy to amplify the RoAS metric. For a brand in a product category, AR shows - for every customer who bought the brand on-site - how many on-site customers researched in the product category (i.e., viewed at least one product in the product category on-site) and bought the brand off-site. For example, if 76 on-site customers bought a brand A1C laptop on-site in a certain time period and 220 customers researched on-site for laptops but bought a brand A1C laptop off-site in the same time period, the Amplifier Ratio is 220/76 = 2.9.

For our analysis, we collected survey responses of 41,946 Amazon customers across ten categories in 2021 (Audio Speakers, Computer Printers, Doll Toys, Laptop Computers, Microwaves, Power Drills, Running Shoes, Smartphones, Tablet Computers, and Televisions). We conducted a first wave of surveys for the categories Computer Printers, Laptop Computers, Smartphones, and Televisions from 2/8/21 to 3/2/21. We ran a second wave of surveys for the categories Audio Speakers, Doll Toys, Laptop Computers, Microwaves, Power Drills, Running Shoes, Smartphones, and Tablet Computers from 7/7 to 7/21 and from 10/07-10/17. We find that the Amplifier Ratio exceeds 1 for each product category: after researching a product on Amazon, more consumers buy the product off Amazon than on Amazon. However, this number varies substantially between categories and brands. ROPES, the number of on Amazon category researchers who buy off-Amazon, varies from 0.5% to 10%, with product categories of dolls and toys, printers, and running shoes showing higher average ROPES than laptops and smartphones:

We ran a Random Forest model to predict the share of off-site purchases at the brand level given the set of predictors described above. We calibrate 300 decision trees based on random samples after obtaining hyper-parameters from a random grid search via cross-validation. In over 92% of the cases, the predictions do not significantly different from the actual values:

Because our Random Forest model is a non-linear model, we use the attributes’ importance calculated from the Random Forest model to interpret which features best predict ROPES. We find that brand-specific information (e.g., the share of detail page views or the number of on-Amazon sales and buyers of the brand in the category) are particularly good predictors of off-Amazon purchase. We also observe that advertising-related features, such as the brand’s share of Sponsored Brands and Sponsored Products, as well as the share of products that are advertised tend to have a high importance as predictors

First, as expected, the effect of all ad products on on-site sales is larger?than their effect on off-site sales. That is, all off-site contribution shares are smaller than 50%. Second, we find that upper funnel advertising like Display Ads, Video Ads (non Streaming TV, which was not included in the analysis) or Sponsored Display ads tends to have a relatively larger share of off-site sales contribution with on average 39% relative to lower funnel advertising like Sponsored Products or Brands advertising that affects mainly on-site sales, with an average share of off-site sales contribution of 19%. Third, we find that always-on related ad products have a larger relative share of off-site contribution (29%) relative to non-always-on related ad products with an average of contribution of 25%. Indeed, always-on ad products are likely to have stronger effects on off-site behavior.

Overall, we find that webrooming on e-commerce sites such as Amazon is substantial. Brands and advertisers on these e-commerce sites should consider not only the effect of their presence and advertising on these sites’ purchases but also as a major billboard for purchases occurring off-site. What do you think?