Lookalike modeling breathing new life into old channels
Columnist Jordan Elkind explains how a mainstay of the ad tech industry is opening up new possibilities for online and offline targeting.
Lookalike modeling isn’t new. It’s been a mainstay of the ad tech industry for years, used to help advertisers expand digital audiences while maintaining relevancy of targeting. The principle is simple. Brands want to attract new visitors to their site. What better way to do this than to identify prospects who resemble existing visitors (or customers)?
What is new is the dazzling variety of ways in which digital marketers are deploying lookalike modeling techniques to enhance the return on investment across marketing channels — both online and offline.
With more data than ever before on user journeys and behaviors, increased adoption of platforms (like customer data platforms and data management platforms) to centralize and analyze that data, and growing ubiquity of machine learning tools and techniques, lookalike modeling is breathing new life into old channels.
What is lookalike modeling?
Customer-centric businesses have long recognized that the best way to acquire new visitors is to focus on users who resemble their existing visitors (or better yet, high-value customers). For digital marketers looking to drive traffic and conversions, this long meant identifying and purchasing media against audiences based on a small number of static demographic attributes. Your recent site visitors are statistically more likely to be females, aged 18-29? Perfect — serve display advertisements to similar audiences elsewhere on the web!
The problem is that demographic segment-based targeting, while enabling advertisers to reach audiences at scale, isn’t a great proxy for relevancy. Women aged 18-29 are a diverse demographic, only a subset of whom are likely to be interested in a brand’s offering. As a result, performance can tend to show a steep drop-off as audience size increases.
Enter lookalike modeling, a form of statistical analysis that uses machine learning to process vast amounts of data and seek out hidden patterns across pools of users. Lookalike modeling works by identifying the composition and characteristics of a “seed” audience (for example, a group of recent site visitors or high-value customers), and identifying other users who show similar attributes or behaviors.
By analyzing not just demographic but behavioral similarities — e.g., users who have demonstrated similar browsing patterns — lookalike modeling enables advertisers to leverage powerful and complex data signals to find the perfect audience.
Why does it matter?
Lookalike modeling is a trusty tool in the digital media arsenal — and it’s quickly becoming indispensable to other channels as well. The convergence of ad tech and CRM (customer relationship management) — powered by platforms that enable advertisers to go well beyond cookies and CRM professionals to gain visibility into the digital journeys of known users — has made it possible to build lookalike audiences of unprecedented sophistication.
A model is only as good as the data it’s given; the growth in technologies that help marketers understand their customers better has only increased the potency of lookalike modeling.
To see why this matters, imagine that a fashion retailer with brick-and-mortar store locations wanted to acquire more haute-couture trendsetters. How might it craft a relevant lookalike audience? Before the rise of technologies like data management platforms (DMPs) or customer data platforms (CDPs), it could have built a seed audience out of a pool of cookies from users who had browsed certain parts of their site or purchased certain items online.
But with a single source of customer data spanning online and offline engagement, it would be able to unify disparate signals of purchase intent from many customer touch points: onsite behavior, email engagement, offline purchases, app usage, call center contacts, product reviews and more. This would provide a much richer — and more accurate — view of customer behavior to power high-performing lookalike models.
So, what are some of the ways marketers are flexing the power of lookalike modeling in new channels?
• Social: Facebook Lookalike Audiences enables marketers to build a seed list based on pixel audiences (e.g., users who have recently visited the site or browsed a particular page) or a custom list of users.
For example, the fashion retailer described above could use a platform to identify all customers with a predicted affinity — based on dozens of behavioral data points — for haute couture, and simply transfer that audience directly to Facebook. Marketers can then indicate how targeted vs. broad they would like the lookalike targeting to be.
• Search: Getting in front of high-potential prospects when they’re in-market — searching or doing price comparison for a relevant category — is every marketer’s dream. The introduction of Similar Audiences through Google Customer Match enables marketers to automatically optimize bidding strategies around key lookalike audiences.
Advertisers can upload lists of customers to Google and then configure Similar Audiences to optimize for search, shopping (product listing ads), YouTube and Gmail ads, and more.
• Direct mail: Rented or borrowed lists have long been the preferred form of direct mail prospecting. But a number of vendors now offer the capability to run lookalike models based on a marketer’s known customers against the larger mailable population, to identify which addresses are most likely to respond to that catalog or magazine.
Ultimately, the ability to stitch together customer journeys across touch points, channels and platforms has provided marketers with unprecedented visibility into customer behavior — and made lookalike modeling a crucial capability for a growing number of marketing channels.
Marketers who are not yet taking advantage of these capabilities should consider putting lookalike modeling to work to win the kinds of new customers who will help them grow their business in the long run.
Opinions expressed in this article are those of the guest author and not necessarily MarTech. Staff authors are listed here.