Incident 251: Amazon Allegedly Tweaked Search Algorithm to Boost Its Own Products
Suggested citation format
The e-commerce giant overcame internal dissent from engineers and lawyers, people familiar with the move say.
Amazon.com Inc. AMZN 2.15% has adjusted its product-search system to more prominently feature listings that are more profitable for the company, said people who worked on the project—a move, contested internally, that could favor Amazon's own brands.
Late last year, these people said, Amazon optimized the secret algorithm that ranks listings so that instead of showing customers mainly the most-relevant and best-selling listings when they search—as it had for more than a decade—the site also gives a boost to items that are more profitable for the company.
The adjustment, which the world’s biggest online retailer hasn’t publicized, followed a yearslong battle between executives who run Amazon’s retail businesses in Seattle and the company’s search team, dubbed A9, in Palo Alto, Calif., which opposed the move, the people said.
Any tweak to Amazon’s search system has broad implications because the giant’s rankings can make or break a product. The site’s search bar is the most common way for U.S. shoppers to find items online, and most purchases stem from the first page of search results, according to marketing analytics firm Jumpshot.
The issue is particularly sensitive because the U.S. and the European Union are examining Amazon’s dual role—as marketplace operator and seller of its own branded products. An algorithm skewed toward profitability could steer customers toward thousands of Amazon’s in-house products that deliver higher profit margins than competing listings on the site.
Amazon’s lawyers rejected an initial proposal for how to add profit directly into the algorithm, saying it represented a change that could create trouble with antitrust regulators, one of the people familiar with the project said.
The Amazon search team’s view was that the profitability push violated the company’s principle of doing what is best for the customer, the people familiar with the project said. “This was definitely not a popular project,” said one. “The search engine should look for relevant items, not for more profitable items.”
Amazon said it has for many years considered long-term profitability and does look at the impact of it when deploying an algorithm. “We have not changed the criteria we use to rank search results to include profitability,” said Amazon spokeswoman Angie Newman in an emailed statement.
Amazon declined to say why A9 engineers considered the profitability emphasis to be a significant change to the algorithm, and it declined to discuss the inner workings of its algorithm or the internal discussions involving the algorithm, including the qualms of the company’s lawyers.
The change could also boost brand-name products or third-party listings on the site that might be more profitable than Amazon’s products. And the algorithm still also stresses longstanding metrics such as unit sales. The people who worked on the project said they didn’t know how much the change has helped Amazon’s own brands.
Amazon’s Ms. Newman said: “Amazon designs its shopping and discovery experience to feature the products customers will want, regardless of whether they are our own brands or products offered by our selling partners.”
Antitrust regulators for decades have focused on whether companies use market power to squeeze out competition. Amazon avoided scrutiny partly because its competitive marketplace of merchants drives down prices.
Now, some lawmakers are calling for Washington to rethink antitrust law to account for big technology companies’ clout. In Amazon’s case, they say it can bend its dominant platform to favor its own products. Sen. Elizabeth Warren (D., Mass.) has argued Amazon stifles small businesses by unfairly promoting its private-label products and underpricing competitors. Amazon has disputed this claim.
During a House antitrust hearing in July, lawmakers pressed Amazon on whether it used data gleaned from other sellers to favor its own products. “The best purchase to you is an Amazon product,” said Rep. David Cicilline (D., R.I.). “No that’s not true,” replied Nate Sutton, an Amazon associate general counsel, saying Amazon’s “algorithms are optimized to predict what customers want to buy regardless of the seller.” House Judiciary Committee leaders recently asked Amazon to provide executive communications related to product searches on the site as part of a probe on anticompetitive behavior at technology companies.
Amazon says it operates in fiercely competitive markets, it represents less than 1% of global retail and its private-label business represents about 1% of its retail sales.
Amazon executives have sought to boost profitability in its retail business after years of focusing on growth. A majority of its $12.4 billion in operating income last year came from its growing cloud business.
Pressure on engineers
An account of Amazon’s search-system adjustment emerges from interviews with people familiar with the internal discussions, including some who worked on the project, as well as former executives familiar with Amazon’s private-label business.
The A9 team—named for the “A” in “Algorithms” plus its nine other letters—controls the all-important search and ranking functions on Amazon’s site. Like other technology giants, Amazon keeps its algorithm a closely guarded secret, even internally, for competitive reasons and to prevent sellers from gaming the system.
Customers often believe that search algorithms are neutral and objective, and that results from their queries are the most relevant listings.
Executives from Amazon’s retail divisions have frequently pressured the engineers at A9 to surface their products higher in search results, people familiar with the discussions said. Amazon’s retail teams not only oversee its own branded products but also its wholesale vendors and vast marketplace of third-party sellers.
Amazon’s private-label team in particular had for several years asked A9 to juice sales of Amazon’s in-house products, some of these people said. The company sells over 10,000 products under its own brands, according to research firm Marketplace Pulse, ranging from everyday goods such as AmazonBasics batteries and Presto paper towels, to clothing such as Lark & Ro dresses.
Amazon’s private-label business, at about 1% of retail sales, would represent less than $2 billion in 2018. Investment firm SunTrust Robinson Humphrey estimates the private-label business will post $31 billion in sales by 2022, more than Macy’s Inc.’s annual revenue last year.
The private-label executives argued Amazon should promote its own items in search results, these people said. They pointed to grocery-store chains and drugstores that showcase their private-label products alongside national brands and promote them in-store.
A9 executives pushed back and said such a change would conflict with Chief Executive Jeff Bezos’ “customer obsession” mantra, these people said. The first of Amazon’s longstanding list of 14 leadership principles requires managers to focus on earning and keeping customer trust above all. Amazon often repeats a line from that principle: “Leaders start with the customer and work backwards.”
One former Amazon search executive said: “We fought tooth and nail with those guys, because of course they wanted preferential treatment in search.”
For years, A9 had operated independently from the retail operations, reporting to its own CEO. But the search team, in Silicon Valley about a two-hour flight from Seattle, now reports to retail chief Doug Herrington and his boss Jeff Wilke —effectively leaving search to answer to retail.
After the Journal’s inquiries, Amazon took down its A9 website, which had stood for about a decade and a half. The site included the statement: “One of A9’s tenets is that relevance is in the eye of the customer and we strive to get the best results for our users.”
Mr. Herrington’s retail team lobbied for the adjustment to Amazon’s search algorithm that led to emphasizing profitability, some of the people familiar with the discussions said.
When a customer enters a search query for a product on Amazon, the system scours all listings for such an item and considers more than 100 variables—some Amazon engineers call them “features.” These variables might include shipping speed, how highly buyers have ranked product listings and recent sales volumes of specific listings. The algorithm weighs those variables while calculating which listings to present the customer and in which order.
The algorithm had long placed a priority on variables such as unit sales—a proxy for popularity—and search-term relevance, because they tend to predict customer satisfaction. A listing’s profitability to Amazon has never been one of these variables.
Amazon retail executives, especially those in its private-label business, wanted to add a new variable for what the company calls “contribution profit,” considered a better measure of a product’s profitability because it factors in non-fixed expenses such as shipping and advertising, leaving the amount left over to cover Amazon’s fixed costs, said people familiar with the discussion.
Amazon’s private-label products are designed to be more profitable than competing items, said people familiar with the business, because the company controls the manufacturing and distribution and cuts out intermediaries and marketing costs.
Amazon’s lawyers rejected the overt addition of contribution profit into the algorithm, pointing to a €2.42 billion fine ($2.7 billion at the time) that Alphabet Inc.’s Google received in 2017 from European regulators who found it used its search engine to stack the deck in favor of its comparison-shopping service, said one of the people familiar with the discussions. Google has appealed the fine and has made changes to Google Shopping in response to the European Commission’s order.
To assuage the lawyers’ concerns, Amazon executives looked at ways to account for profitability without adding it directly to the algorithm. They turned to the metrics Amazon uses to test the algorithm’s success in reaching certain business objectives, said the people who worked on the project.
When engineers test new variables in the algorithm, Amazon gauges the results against a handful of metrics. Among these metrics: unit sales of listings and the dollar value of orders for listings. Positive results for the metrics correlated with high customer satisfaction and helped determine the ranking of listings a search presented to the customer.
Now, engineers would need to consider another metric—improving profitability—said the people who worked on the project. Variables added to the algorithm would essentially become what one of these people called “proxies” for profit: The variables would correlate with improved profitability for Amazon, but an outside observer might not be able to tell that. The variables could also inherently be good for the customer.
For the algorithm to understand what was most profitable for Amazon, the engineers had to import data on contribution profit for all items sold, these people said. The laborious process meant extracting shipping information from Amazon warehouses to calculate contribution profit.
In an internal system called Weblab, A9 engineers tested proposed variables for the algorithm for weeks on a subset of Amazon shoppers and compared the impact on contribution profit, unit sales and a few other metrics against a control group, these people said. When comparing the results of the groups, profitability now appeared alongside other metrics on a display called the “dashboard.”
Amazon’s A9 team has since added new variables that have resulted in search results that scored higher on the profitability metric during testing, said a person involved in the effort, who declined to say what those new variables were. New variables would also have to improve Amazon’s other metrics, such as unit sales.
A review committee that approves all additions to the algorithm has sent engineers back if their proposed variable produces search results with a lower score on the profitability metric, this person said. “You are making an incentive system for engineers to build features that directly or indirectly improve profitability,” the person said. “And that’s not a good thing.”
Amazon said it doesn’t automatically shelve improvements that aren’t profitable. It said, as an example, that it recently improved the discoverability of items that could be delivered the same day even though it hurt profitability.
Amazon’s Ms. Newman said: “When we test any new features, including search features, we look at a number of metrics, including long term profitability, to see how these new features impact the customer experience and our business as any rational store would, but we do not make decisions based on that one metric.”
In some ways, Amazon’s broader shift from showing relevant search results is noticeable on the site. Last summer, it changed the default sorting option—without publicizing the move—to “featured” after ranking the search results for years by “relevance,” according to a Journal analysis for this article of screenshots and postings by users online. Relevance is no longer an option in the small “sort by” drop-down button on the top right of the page.