Data misclassification has become too common during a time when technology like artificial intelligence (AI) and quantum computing can sort through triggers, signals and computational analysis in record time.
Data that buckets people into categories is often inaccurate, and the industry is stuck in an old-school paradigm of predictions.
Adlook conducted its study of 1,335 U.S. online respondents in September 2024. The methodology involved a two-step process that combined user polling with bid-request data analysis via Adlook to assess the reliability of socio-demographic targeting.
advertisement
advertisement
Validation and overlap analysis of the data came into question when in the survey data found in the segment of “Moms," 52% of the targeted users identified as men. Some 67% of users in the “Parents” segment declared they did not have children.
Data misclassification has become too common during a time when technology like artificial intelligence (AI) and quantum computing can sort through triggers, signals and computational analysis in record time.
It seems strange to think about bad or inaccurate data during a time when technology can do so much more.
“It’s very much still a challenge,” said Scott Sutton, CEO of Later, an influencer marketing platform, commenting on the study.
There’s a balance between privacy and targeting. These inaccuracies occur when marketers and brands use “outdated statistical modeling vs. more advanced AI tooling,” Sutton said. “It’s a limitation of the methodology and access to data.”
Sutton thinks his company has a better way to target audiences. He's trying to create a new target category called performance-driven social fueled by data and AI. The company has more than 300 million posts worth of data direct from users with connections directly to the social networks. Millions of transactions and $1.5 billion run rate of purchase goods that flow through the company’s data set. It helps him predict actual performance and audiences.
The findings from Adlook, a cookieless brand growth platform and part of the RTB House Group, shows how the old way to target digital ads failed to meet performance metrics.
The study highlighted challenges in the accuracy of socio-demographic targeting, with a key insight being the disconnect between targeting assumptions and actual audiences.
For women age 18 to 24, a commonly targeted segment, for example, preciseness became less than 20% in the data. Among those targeted, 43% were men, and 61% were older than 24 years, while 35% were above 55, and only 18% were women ages 18 to 24.
More than half of impressions were misclassified in conflicting age groups, meaning that the same users were incorrectly labeled across multiple, contradictory segments.
Some 40% of “homeowners” in targeting data were renters, and vice versa.
Only 18% of impressions intended for “women 18-24” reached that audience. Instead, 43% went to men, and 35% went to users older than 55.
Simple socio-demographic segments that should be mutually exclusive revealed significant overlaps, meaning that the same users were inaccurately classified into multiple, conflicting categories.
Some 35% of impressions were simultaneously eligible for both the “women” and “men” segments, while 55% fell into two or more age groups. And 28% of impressions were eligible for both the “Age < 34” and “Age > 55” segments.
Does this work for all types of categories? I told Sutton that I love dogs, but I do not have one at this time.
Sutton said in his company would use an ad-targeting model that creates “a rich context model for a brand, for a campaign, and context of influences and audiences” and matches it three ways.
This determines the relative propensity for me to buy something in a specific category of from a brand, although I do not have a dog.
“You might love dogs and have a need to buy a product, but not necessarily own a dog,” he said. “I care less whether you own a dog. I need you to want to engage with the content and have the purchasing power.”