Signal loss calls for the use of, well, other signals.
“The biggest trend for us right now is finding ways to be less reliant on cookie data,” said John Kirk, media director in charge of digital investment at 22Squared, an Atlanta-based media agency whose clients include Baskin-Robbins, Publix and Southeast Toyota.
One alternative approach, Kirk said, is to “home in on audiences where we do have the data.”
In that vein, 22Squared has been testing a solution released by AdTheorent on Wednesday that uses machine learning to score programmatic inventory based on the probability that an impression will lead to a desired outcome. Southeast Toyota is also a launch partner for the product.
“We’re not chasing individual IDs and targeting and retargeting them,” said AdTheorent CEO Jim Lawson. “We’re determining the statistical parameters of a target audience.”
Programmatic flipped
Rather than list-based targeting using a CRM file or cookies to identify people and serve impressions, AdTheorent analyzes a seed data set, usually supplied by the brand, to learn as much as possible about that brand’s target audience, including behaviors, location data pulled from in-app SDKs, demographics and vertical-specific data, such as CPG, auto, travel or retail.
AdTheorent then builds a predictive model that hunts for inventory where ad buyers are most likely to find people with those attributes.
It’s a method for expanding a target audience without exposing personal data or targeting individuals, Lawson said.
“I know ‘privacy-safe’ has become a buzzword, big time,” he said, “but I’d argue that this is privacy-safe because it’s not user-focused, it’s an aggregated data set, and the seed data isn’t used for targeting.”
Fresh data
Using seed data to extrapolate audience information is also helpful from a data freshness perspective.
Data freshness “is always a challenge,” Kirk said.
“Buying behavior is constantly changing, which is true in general but that became even more apparent during COVID,” he said. “And third-party audience segments that aren’t the freshest just won’t be good for performance.”
But a machine learning approach allows AdTheorent to score audiences and inventory in near real time, Lawson said.
Say a car brand is looking for in-market auto shoppers with a likelihood of buying a vehicle within a certain timeframe.
The brand would supply AdTheorent with information about people who had purchased a car in the past (think CRM data or a keyword search list), and AdTheorent would use that information to extract a series of predictive attributes for the desired behavior and find related impression opportunities with a good chance of converting on specific KPIs.
“The way I’d describe it, we’re not just looking for like-minded individuals, we’re looking for like-minded impressions,” Kirk said. “The predictive algorithm is looking for any place where our ads have the potential to reach someone who’s ready to make a purchase.”
Measurement is provided through a simple site pixel that flags whether an event occurred after an ad was served – but that’s it.
“It’s a ‘yes’ or ‘no’ question,” Lawson said. “We just want to know whether an impression that was served yielded an outcome, but we’re not taking IDs after a campaign runs and putting them into a list.”
It’s not magic
However, AdTheorent does give advertisers insight into how its model works, which isn’t typical for machine learning-powered ad products.
Google’s Performance Max and Meta’s Advantage Plus are both black boxes that require advertisers to trade in control in return for performance.
“We want to help brands learn more about their audiences,” Lawson said. “It’s like the opposite of a black box.”
And having the ability to control the data inputs and what gets fed into the model is an appealing prospect, Kirk said, particularly from a strategy perspective.
“We hear from Google all the time, ‘Hey, it just works,’ like it’s magic or something, but that shouldn’t be good enough,” Kirk said. “We also want transparent data inputs and we want a voice in what those inputs are.”
