Online advertisers and government agencies use algorithmic tools to tailor and target their campaigns to reach as many people as possible.
Sometimes they miss their mark, and in the case of local, state and federal programs designed to help people, that can unintentionally amount to inequity in the distribution of information or resources.
Cornell researchers have developed a method that can help online advertisers ensure they're reaching their intended audience - including "unknown users," often those with low socioeconomic status or nonbinary gender identities, who are often overlooked in precision ad targeting. The researchers' multiple-campaign approach effectively reduces "skew" - under-delivery to certain demographic groups when trying to reach a balanced audience - at a reasonable cost.
"We feel this intervention has the potential to improve equity in terms of who's being shown ads for public resources and who has access to them," said Isabel Corpus, a doctoral student in the field of information science. "And if an agency is not measuring and not mitigating this unintended skew, it could potentially undermine initiatives that are actually intended to reduce social disparities."
Corpus is lead author of "Into the Unknown: Accounting for Missing Demographic Data when Mitigating Ad Delivery Skew," which she will present at the ACM Conference on Fairness, Accountability, and Transparency (FAccT '26), June 25-28 in Montreal.
The senior author is Allison Koenecke, assistant professor of information science at Cornell Tech and the Cornell Ann S. Bowers College of Computing and Information Science.
The researchers' approach involves an advance in a concept called budget-splitting, in which an advertiser or organization will run a single ad campaign with no demographic targeting, then evaluate it for skew - as measured by the number of impressions, or views, the ad garners. Gender-based skew, for example, is the ratio of impressions by inferred male (or female) users, relative to total impressions by inferred male and female users. They then replace that single campaign with multiple campaigns that take into account groups underrepresented in the original.
Other researchers had introduced the concept of budget-splitting at FAccT '25 as an alternative to Meta's Variance Reduction System (VRS). Meta has touted VRS as a way to ensure that delivery of certain types of ads - for housing, credit and employment - is consistent with the distribution of eligible users on its platform. However, VRS does not measure and reduce all skew, and is only applicable to Meta platforms.
Corpus and Koenecke took the idea of budget-splitting a step further by adding "unknown users" to the equation. For their study, they worked with a state-level government agency that provided grants, consulting and other services to entrepreneurs. The agency wanted its advertising to reach all potential entrepreneurs, including women, who have historically been underrepresented among business owners. A 2023 Census Bureau report found that just 39% of the 36.4 million U.S. businesses were owned by women.
The researchers noted that Google provides advertisers with three labels to indicate user gender: male, female and unknown. Using just the first two in targeted advertising leaves out a measurable chunk of potential targets; "unknown" users tend to be those with low socioeconomic status or nonbinary gender identities.
"As a government advertiser," Corpus said, "you're not looking to exclude anyone from potentially being shown these resources."
Corpus and Koenecke designed a campaign with four target audiences, separated by Google-inferred labels: male; female; male and unknown; and female and unknown. To reduce gender-based skew without excluding unknown users, they first targeted "female" and "male plus unknown" users; in the second wave, they targeted "male" and "female plus unknown."
Their approach resulted in both a measurable reduction in skew and heightened cost-effectiveness compared to a simple budget-split approach, which uses single-group (i.e., gender) targeting.
Koenecke said this work shows the value of academic research in tackling real-world problems such as equity and efficiency in targeted advertising.
"These algorithms have a long history of being biased in who sees the ads downstream - regardless of whether biases are intentional or unintentional," Koenecke said. "Increasingly, organizations are realizing the value in quantifying these potentially hidden biases, and taking steps to ameliorate them."
