Today’s column is Part I of a two-part column written by Mike Afergan, CTO & SVP of Advertising Decision Solutions at Akamai.
Targeting is only as good as the data that you start with — and the idea of using shopping data as a powerful tool for advertiser’s campaigns is beginning to gain traction. Not only is this data powerful but it is also flexible, allowing an advertiser to address both upper and lower purchase funnel goals as I will explore in a two-part series of articles. Today I’ll explain how to use shopping data to accurately predict those lower-funnel users considered in-market and how to find them. I’ll also share with you some findings on the value of these users you while they’re in-market, and how predictive models go about identifying this quick-changing segment of online users.
Part 1: Using Shopping Data to Find In-Market Buyers
There are approximately 150 million online shoppers in the United States. But for obvious reasons, most advertisers aren’t interested in marketing to all of them. Rather, advertisers tend to look at their audiences in terms of target market and in-market consumers. Many focus their display advertising only on their target market which for most products is in the range of one to ten million shoppers. This is an audience online advertisers typically try to reach, using a combination of tools in display advertising such as demographics, geography, or psychographics.
A simple example of this distinction is the market for high-definition TVs. At any point in time, there is a rather large discrepancy between the size of the target market and the number considered in-market today to buy an HDTV. For HDTVs, the target market may include all consumers of a certain income and demographic, as well as home theatre aficionados. These are people who represent a higher likelihood of purchase at some point over the next couple of years. Marketers have typically spent A LOT of money directing their message to this fairly open-ended group – which is a reasonable approach for companies and agencies focused on stimulating long-term consumer interest. But, for advertisers such as retailers who seek to drive near-term ROI, there is a considerable ability to drive incremental transactions given the sort predictive targeting options available to marketers today.
In-Market versus Target Market
The ability to differentiate between in-market shoppers and their rest of the target market offers tremendous possibilities for boosting sales and maximizing return on ad spend. For example, a camera retailer could direct a branding campaign at a broad audience of photography buffs, but target relevant promotions to the subset who are actively engaged in upgrading their gear. By reaching each consumer with a message appropriate to his stage of the purchase funnel, advertisers can boost incremental sales as well as efficiency of ad spend.
But how different are these users really? The answer is very. First of all, there are roughly 30 million in-market consumers at any given time in the US for all products, although this number spikes considerably higher during the Q4 holiday season. As expected, these consumers display a much higher frequency of online shopping activity than normal. I recently led some in-house research and we found that when a shopper is in-market, he or she:
- Performs 5 times more shopping events,
- Initiates 8 times more shopping carts, and
- Is 6 times more likely to make a purchase than when not in-market.
But timing is critical, as the in-market consumer is a moving target. Every three weeks, roughly 80% of the group of in-market shoppers turns over, with in-market duration varying directly with price point. Typically the lower the product price point, the less time the average shopper stays in-market. The fact that in-market status is so transient creates a serious challenge for advertisers who do not have the ability to identify this group of high-value consumers. Keep in mind that these shoppers are completing transactions—potentially with competitors who may be able to recognize their in-market status.
How to identify in-market consumers
Recognizing in-market shoppers is, however, not an easy task. How do you know when someone is primed to make a purchase? Can you tell by their age and gender? Income level or lifestyle? The articles they read?
Generally, the answer is no. Those properties don’t change as people come in and out of market. Keep in mind that you and I browse a lot of the same articles when we’re buying a television and when we’re not. And I have plenty of friends who read television reviews even when they are not in market at all – just because they like to know what is happening in the industry. Furthermore, purchase intent changes very quickly—often in a matter of days, making active intent even harder to identify. To be successful in identifying in-market consumers one needs three things: (1) a large, rich, and cross-site dataset, (2) predictive algorithms, and (3) a real-time platform capable of processing the data, making the decisions, and ultimately allowing for the targeted advertising.
I believe that good modeling starts with great data – great and rich in several ways. To accurately predict shopping behavior, a model ideally would have access to billions of shopping data — past and present — points across hundreds of sites. And then the data must be rich across a wide variety of sites and types of sites. In my company’s predictive models, we typically find tremendous advantage in incorporating data from hundreds of websites across many product categories. This is, of course, in stark contrast to simple remarketing which is limited to the power of only one site. Further, it has to be rich across behaviors on those sites. Simply placing pixels on a few pages of a site is like trying to understand user behaviors in a large department store by watching who goes in and out the front door. We see that pixeling strategies often capture less than 30% of the relevant users and information.
Of course, taking advantage of the data in any meaningful way requires the power of predictive analytics. This approach and the associated algorithms are used across a wide range of industries to solve difficult problems that range from detecting credit card fraud to determining patient risk levels for medical conditions. It combines data mining and machine-learning technologies to create statistical models based on historical data which in turn are used to predict future events. This is in contrast of course to basic segmentation (“everyone who went to my homepage”). It is also in contrast to basic look-alike modeling. I look a lot like some of my colleagues in the office and we spend a lot of time surfing the same websites. But if you look closely at my recent purchases, you’ll see the discriminating detail is that I have made a lot of purchases related to my newborn daughter – which will tell you that my purchases over the next few months will be very different than some of my peers even though we “look alike.”
Of course these models and data are not valuable unless they can then be leveraged at great scale and in real-time. Given the rich nature of a shopping data set and the need to react quickly to users coming in and out of market, it is critical that a system using shopping data be able to process the massive data set and then distribute that information the appropriate targeting systems quickly and efficiently. But when that happens, the richness of the shopping data coupled with the predictive analytics leveraged in real time can be an incredibly effective tool in identifying in-market consumers — and driving incremental transactions and a strong ROI to the advertiser.