Look-alike modeling has been an important part of the media toolkit over the past, allowing brands to increase their audience pool by taking a core group of top-performing individuals, grouping them and using data and technology to find other individuals like them.
Over the past several years, data management platforms (DMPs), third-party cookies and their associated data are becoming obsolete due to self-regulation by technology providers and like legislation CCPA and GDPR.
The movement away from third-party cookies and third-party data overlays on cookies is causing total audience pools to drop in size as individuals have fewer associated identifiers (cookies to connect to).
However, look-alike modeling can also help businesses leverage their first-party data to build robust large-scale segments for marketing and advertising purposes.
Tealium’s regional vice president of strategic partnerships for the Americas, Travis Cameron, explained that the value of being able to expand target populations based on data associated with a high-value segment will take on a different dimension.
These changes range from the identifiers used (hashed PII data to match & expand) to the utilization of different data types (contextual, interest-based, pathing) vs. demographic or psychographic inferred data previously used.
“The value remains — marketers need eyeballs, and the quest to find individuals like those who just converted to optimize spend will always have a high associated value as a tactic,” Cameron said. “It is just going to become harder to model and expand your audience.”
Related Article: Now You See Me, Now You Don’t: Navigating the Cookie-less World
Understanding the Machine Learning Model
“Given the complexity of using large amounts of first-party data to build segments, it’s important for businesses to understand how the modeling is reaching its conclusions by understanding which features are used to build the model and their relative importance,” explained Alex Holub, CEO of Vidora. “In addition, it’s important to understand the ultimate goal of the look-alike segment.”
For instance, if the goal is to maximize clicks on a marketing campaign, using a machine learning model that directly optimizes for clicks will yield far better performance than a semi-supervised look-alike model.
Holub explained machine learning offers a few key benefits when building look-alike segments:
First, machine learning is continually adapting segments based on the latest information available. By continually learning, machine learning can incorporate new user activities or actions, in addition to broader legal changes which impact behaviors across all users.
Second, machine learning implicitly learns the importance of user behaviors and attributes, which allows machine learning to leverage all available first-party data and implicitly upweight or downweight the importance of each data point.
“In other words, machine learning can incorporate all available first-party data and learn what information is the most important to the model,” Holub said.
Cameron noted that it’s important to have a clean and correlated data set of your customers that contain like data about each consumer ahead of starting the exercise. “Know the outcome that you are looking for. Any model you’re using needs to be developed and applied with an outcome in mind, so figure out what you trying to optimize with this audience,” he said.
He recommended using an agile approach to test, learning quickly and ensuring you have a testing plan in place as you begin to act on the audience response. “Know where you plan to activate the audience through and ensure you have the right identity points and integrations to activate on.”
Deploying Data Science to Target Groups
Data science platforms can also offer a fast and reliable solution for businesses to model their first-party data. For instance, Vidora’s product Cortex helps businesses build, understand and integrate look-alike models into their business within a few days.
“For smaller teams, and teams which need to move quickly, a data science platform can augment productivity and be a great value,” Holub said. He added that he currently sees a lot of leveraged businesses look-alike modeling as a black box — typically as a component of their DMP.
However, given the increasing organizational dependence on look-alike modeling using first-party data as a key revenue stream, he predicts businesses will begin to leverage more sophisticated data science techniques.
“Using data science techniques should result in both higher quality segments for brands to market to but also segments which align more directly with the brand’s goals,” he said.
For example, data science techniques outside of look-alike modeling can build segments that directly optimize for user engagement (eg, clicks), down-stream conversions and increase in brand sentiment (eg, uplift modeling).
“I think we will see an increasing reliance on real-time data for look-alike models as the number of first-time users and anonymous users increases,” Holub added.
Cameron said companies with smaller adtech and data teams need to focus their efforts on developing and optimizing models and understanding their customers and the key data points that drive them, not engineering and cleansing their data.
“Achieving a state of data automation within which they can work with their full audience data, leverage a few key partners to better deploy their audiences and have strict measurement theses that allow them to execute at the same pace as a much larger organization,” he explained.
Related Article: 4 Steps to Identify Your Best Prospects Using Technographic Data