5 Questions with Anja Lambrecht: Digital Advertising, Targeting, and Apparent Bias

A periodic feature by Cornerstone Research, in which our affiliated experts, senior advisors, and professionals talk about their research and findings.

We interview Anja Lambrecht of London Business School. Professor Lambrecht is an authority on the digital economy and digital marketing. Using econometric methods and field experiments, she researches a variety of issues related to online advertising, promotion, pricing, and consumer behavior on the internet.

What are some of the most common forms of digital advertising?

Digital advertising falls broadly into two categories, which are based on where the ads are delivered: display ads and search ads.

Display ads comprise static images, videos, pop-ups, and/or text that can appear on any webpage that provides space for advertising. Display ads may be shown on a variety of content webpages or on social media. In addition, it is possible to target display ads to consumers based on a variety of factors, such as their prior browsing behavior.

Search ads, as their name indicates, appear on search engine results pages based on users’ search queries. Search ads typically show text, but they may also feature images.

How do advertisers target digital ads to consumers?

Advertisers and digital advertising platforms may have access to a variety of data on consumers. Such data include consumers’ demographics, browsing patterns, purchasing patterns, social network activity (i.e., likes, retweets, follows, pins), geographic location, and more. Advertisers may use these data to target existing customers with ads about new products that are complementary or similar to products they have previously purchased or browsed.

Algorithms can also help advertisers identify the consumers most likely to respond to a given ad, based on their interests, behaviors, locations, or other characteristics. For example, advertisers may use algorithms to identify and target ads for ice cream to consumers who are near an ice cream shop. They can even adjust advertising based on whether it is a hot or a cold day.

Not all targeting is equally effective for all consumers, however. Using field experiments conducted through Twitter, I have shown that the effectiveness of targeting strategies varies across consumers. Specifically, I have found that users who are particularly engaged in newly trending organic content are not highly responsive to ads.

How do you measure the effectiveness of digital advertising?

Different advertisers have different goals, so talking about the “effectiveness” of digital advertising is a necessarily nuanced discussion. That said, advertisers use several common metrics to evaluate the success of an advertising campaign. They often measure changes in clicks, the length of an average visit to a website, and the share of visits that convert into a purchase. However, these types of simple “before and after” comparisons only reveal correlation, not causality. For example, if an advertiser shows an ad to a consumer and the consumer then makes a purchase, the advertiser cannot necessarily attribute that purchase to the ad exposure, since the user might well have purchased anyway.

To establish causal relationships between advertising campaigns and outcomes, my research often relies on controlled field experiments, also referred to as A/B tests. Through such controlled experiments, I can ascertain whether differences between a treatment and control group of consumers are due to the treatment, or to external factors.

For example, I analyzed an experiment to assess the effectiveness of personalized ads on an online travel site. One randomly selected group of consumers was shown customized ads—ads for specific hotels—based on their browsing history. The other group was shown a generic ad for the online travel site. I then compared sales across the two groups to test whether ad customization was effective. Interestingly, in this specific situation, the generic ad format was more effective, on average, in converting consumers to purchase, though there was variation across consumers.

Marketers use vast amounts of data to target and customize their digital ads, which some argue raises privacy concerns. Does targeted advertising provide any benefit to consumers?

The discussion about targeted advertising and privacy is complex. Industry observers have raised concerns related to data security and consumer privacy, suggesting that firms should meet certain standards related to the collection, use, and protection of data. At the same time, the use of data to improve ads creates value for advertisers, consumers, and the online platforms hosting the ads in many ways.

Advertisers’ ability to target ads based on consumer characteristics can benefit consumers directly. Such data can allow marketers to create ads and shopping experiences that consumers are likely to find useful. For example, consumers looking for microwave ovens may appreciate seeing microwave ads rather than, say, clothing ads, as they are browsing websites.

Well-targeted ads can also be valuable more broadly. They benefit not only advertisers but also online content providers and, ultimately, consumers. Well-targeted ads may be more likely to increase the probability that a consumer will purchase. This will then be more valuable for the advertiser whose willingness to pay for an ad will increase, potentially raising revenues for the website or platform on which the ads appear. Most online content sites rely on advertising revenues. Maintaining high advertising revenues for a given space on a webpage increases their ability to stay in business and provide useful content to consumers.

You noted that advertisers may use algorithms to select which ads to show to which consumers. We have seen the mainstream press criticizing algorithms as perpetuating bias in some instances. What does that mean, and how do some of these instances of alleged bias play out?

At a macro level, much of the discussion of algorithmic bias has focused on algorithms being designed and trained using data that do not appropriately reflect all groups of the population. In such cases, algorithms may “replicate” unrepresentative results when used for decision-making.

However, my research shows that disparate outcomes can also arise because of the simple mechanics of how algorithms learn over time. Specifically, searches for uncommon names are more likely to produce search engine results pages featuring ads that are not only less relevant, but that also potentially have a negative connotation (e.g., ads for criminal background checks). This occurs simply because the algorithm has not processed enough data to determine that a given ad is not relevant.

I have also found that what may look like bias can in fact be a result of the economics of ad delivery. For example, my field experiments have shown that women are less likely to encounter STEM career ads on social platforms than men. According to my research, this is not because women are less likely to click on such ads, nor because of an underlying bias in the country in which the ad is shown. Rather, it is because women are more likely than men to purchase across a variety of industry sectors, which makes them more valuable to advertisers in the aggregate. Women’s “eyeballs,” in other words, are more expensive to purchase.

Since the cost to display an ad to women is greater than the cost to display the same ad to men, advertisers that do not target a specific demographic segment end up being more likely to show the ad to men—even if their targeting strategy is gender-neutral. What looks like gender bias in the display of career ads is in fact the unintentional result of a profit maximization strategy.