Cornell ad targeting method aims to reduce gaps in public campaigns

Cornell researchers say they have developed a way to make online ad campaigns reach demographic groups that automated targeting systems can miss. The work matters for public agencies because uneven ad delivery can affect who learns about grants, services and other government resources.

The method, described by Cornell University researchers Isabel Corpus and Allison Koenecke, is aimed at reducing “skew,” or underdelivery to parts of an intended audience. Corpus, a doctoral student in information science, is the lead author of the paper, “Into the Unknown: Accounting for Missing Demographic Data when Mitigating Ad Delivery Skew,” which is scheduled for presentation at the ACM Conference on Fairness, Accountability, and Transparency in Montreal from June 25 to 28.

Koenecke, the senior author, is an assistant professor of information science at Cornell Tech and the Cornell Ann S. Bowers College of Computing and Information Science, according to Cornell.

How the method works

The approach builds on a technique known as budget splitting, Cornell said. In that process, an advertiser first runs a broad campaign without demographic targeting, then checks who received the ads by looking at impressions, or views. If one group is underrepresented, the advertiser replaces the original campaign with several campaigns meant to correct the imbalance.

Cornell said earlier researchers presented budget splitting at FAccT ’25 as an alternative to Meta’s Variance Reduction System. Meta has described that system as a way to align delivery of certain housing, credit and employment ads with the eligible user population on its platforms, but Cornell said the Meta system does not address every form of skew and applies only to Meta services.

Corpus and Koenecke extended budget splitting by accounting for users whose demographic labels are incomplete or unavailable. Cornell said Google gives advertisers three gender labels: male, female and unknown. If a campaign considers only male and female labels, it can leave out a significant group of potential viewers.

According to Cornell, users labeled “unknown” often include people with low socioeconomic status or nonbinary gender identities. Corpus said government advertisers seeking to promote public resources generally do not want to exclude people from seeing those ads.

Testing with a government advertiser

For the study, the researchers worked with a state-level government agency that offered entrepreneurs grants, consulting and other services, Cornell said. The agency wanted to reach potential entrepreneurs broadly, including women, who have been underrepresented among business owners.

Cornell cited a 2023 Census Bureau report finding that women owned 39% of the 36.4 million businesses in the United States. The researchers used that context to examine how an ad campaign could reduce gender-based delivery imbalance while still including users with unknown gender labels.

The team created four campaign audiences based on Google-inferred labels: male, female, male plus unknown, and female plus unknown, according to Cornell. In the first round, the campaign targeted female users and male-plus-unknown users; in the second, it targeted male users and female-plus-unknown users.

Cornell said the design reduced skew and was more cost-effective than a simpler budget-splitting approach that targeted single demographic groups. The university did not provide detailed cost figures in its summary.

Koenecke said the study points to the role academic work can play in measuring and reducing hidden bias in ad delivery. Cornell said the researchers view the method as a potential tool for organizations that use targeted ads to distribute information about public resources.

This story draws on original reporting from Phys.org.