Programmatic Video #WotW: What Are Lookalike Models?

Programmatic video is the fastest-growing category of programmatic buying, but a lot of marketers are still struggling to get to grips with the basics.

According to a recent study by Unruly, which polled 1,000 senior marketers in the UK and US, around half rate their programmatic video knowledge as ‘Average’, ‘Poor’ or ‘Very Poor’.

There’s clearly a knowledge gap between the programmatic front lines and everyday marketers, making 2015 a key year for programmatic education. As with any burgeoning new trend, one of the main causes of confusion is vocabulary.

Any discussion of the topic produces a web of jargon and acronyms that’s enough to send anyone running to Google. But fear not, this is easily solved.

Each week we’ll be serving up handy guides to the most important programmatic video terms, courtesy of Unruly co-founder and CTO, Matthew Cooke. Last week we looked at what Online Behavioural Advertising is. This week, Matthew explains what lookalike models are.

Stick around and see how quickly you can master the art of programmatic.

What are lookalike models?

The easiest way to understand lookalike modelling  is to compare re-targeting with and without lookalike modelling.

Normal re-targeting involves targeting ads at people that performed a particular action or viewed a particular web page. For example, a tracking pixel is placed on a car review site that allows  users visiting the site to be identified again later. Advertisers can then target car ads at those users when they are spotted browsing other sites. This is done because a user who visited a car review site is more likely to buy a car than a random user who sees a car ad.

One of the challenges with re-targeting is that it is often difficult to get the volume of data on users that you would like to target. Perhaps 5,000 people looked at a particular car review, but you wish to run 500 thousands ads. Targeting just those 5,000 people won’t have a huge overall impact.

This is where lookalike modelling comes in. You take the 5,000 users that did view the car site, along with all the other information you have about their behaviours and site visits, and then compare them to millions of other users that you have data for. You then pick from the larger data set the users that are most similar to the users that looked at the car review. When we say two users are “similar” we mean the users visited many of the same sites and performed the same sort of actions.

Although the pool of lookalike users will not generally be as likely to buy a car as the 5,000 users who were looking at the car review, they are still more likely to buy a car than a random user.

To give a feel for what is possible, lookalike models might increase the volume of people to target from 0.1% of a population to between 1 and 10% – depending on the quantity of original data, the size of the larger pool of user data and the number of actions recorded about all the users.