People have a mental model of shopping that is based on experiences from brick-and-mortar stores. We intuitively understand how this process works: all available products are displayed around the store and the prices are clearly marked.
Many stores offer deals via coupons, membership cards, or to special classes of people such as students. Typically, everyone is aware of these discounts and has an equal opportunity to use them.
Many people assume this same mental model of shopping applies just as well to e-commerce websites. However, as we are discovering, this is not the case.
In 2010, shoppers realised that Amazon was charging different users different prices for the same DVD, a practice known as price discrimination or price differentiation. In 2012, the Wall Street Journal revealed that Staples was charging users different prices based on their geographic location. The paper also reported that travel retailer Orbitz was showing more expensive hotels to users browsing from Mac computers, a practice known as “price steering”.
These reports of price discrimination and steering provoked a great deal of negative publicity for the companies involved. The lack of transparency also raises many disturbing questions. How widespread are the e-commerce practices of manipulating search results and customising prices? What customer information do companies use to do it? When e-commerce sites personalise prices or search results, by how much do prices change?
My colleagues and I at Northeastern University have taken an initial stab at answering these questions in a new study. We examined 10 major US e-retailers, including Walmart and Home Depot, along with six hotel/rental car sites, including Orbitz and Expedia, to determine if they implement price discrimination or steering, and if so, what user attributes trigger the personalisation.
We recruited 300 people from the crowdsourcing site Mechanical Turk to run product searches on the 16 sites. We paired each of these real users, who each had their own real, idiosyncratic browser history, with an automated browser that ran the same searches at the same time as the real users, but did not store any cookies.
By comparing the search results shown to these automated controls and to the real users, we identified several cases of personalisation. We saw price steering from Sears, with the order of search results varying from user to user. We saw price discrimination from Home Depot, Sears, Cheaptickets, Orbitz, Priceline, Expedia and Travelocity, with product prices varying from user to user.
So what user attributes trigger personalisation? The problem is that real users have a long history of browsed sites, searches, clicks and online purchases that we as researchers don’t know. Thus, when we observe personalised results in our experiments, we can’t tease out the underlying cause.
To figure out what user attributes drive e-commerce personalisation, we conducted another round of testing using fake accounts that we created. All the accounts were identical except for one specific attribute that we changed. In particular, we tested for personalisation based on Web browser, operating system, logging-in to a user account, and purchase history (we had one account book cheap hotels and rental cars for a week, while another account booked expensive hotel rooms and rental cars).
Our fake accounts uncovered many different personalisation strategies employed by e-commerce sites. For example, Travelocity reduced the prices on 5% of hotel rooms shown in search results by around US$15/night for smartphone users. Interestingly, Cheaptickets and Orbitz gave unadvertised “members only” discounts of about $12/night on 5% of hotels rooms to users who were logged in to their accounts on the site.
Expedia and Hotels.com conduct what marketers and engineers call A/B tests to steer a subset of their users toward more expensive hotels. By dividing visitors into different groups, companies are able to use A/B tests to see how users respond to new website features and algorithms. In this case, visitors to Expedia and Hotels.com were randomly assigned to groups A, B, or C based on the cookies stored on their computers. Users in groups A and B were shown hotels with an average price of $187/night, while users in group C were shown hotels with an average price of $170/night.
Home Depot served almost completely different products to users on desktops versus mobile devices. A desktop user searching Home Depot typically received 24 search results, with an average price per item of $120. In contrast, mobile users receive 48 search results, with an average price per item of $230. Bizarrely, products are also $0,41 more expensive on average for Android users.
Initially, we assumed that the sites would not personalise content, given the extremely negative PR that Amazon, Staples and Orbitz received when earlier cases were revealed. To our surprise, this was not the case.
Unfortunately, the business logic underlying much of this personalisation remains a mystery. None of the discounts we located in our experiments were advertised on sites’ home pages, so the deals do not appear to be part of marketing campaigns. When we spoke to representatives from Orbitz and Expedia, they confirmed our findings, but did not elaborate on the rationale for the design of their websites. Representatives from Travelocity confirmed that they do offer deals for mobile users, with the goal being to motivate them to use the site more and install the Travelocity app.
What is clear from our study is that price discrimination and steering on e-commerce sites is becoming more prevalent, and more sophisticated. As a user, it’s almost impossible to know if the prices you are being shown have been altered, or if cheaper products have been hidden from search results.
If you are looking for the best deals and are willing to work for it, we recommend searching for products in your normal desktop browser, an incognito or private browser window, and your mobile device. Of course, e-commerce companies are constantly experimenting with new personalisation techniques, so in the future, an entirely different attribute may trigger personalistion.
Ultimately, we hope that our study will encourage companies to be more transparent about how they personalise prices and search results. Rather than using opaque and creepy algorithms to alter content secretly, companies could stick to the kinds of real-world incentives that shoppers already know and love, like coupons and sales.
Christo Wilson is assistant professor of computer and information science at Northeastern University
- This article was originally published on The Conversation