Time To Expand The Behavioral Lexicon
The underlying premise -- as much philosophical as technological -- of behaviorally based marketing is that one size does not fit all. Ironically, however, the way BT is most commonly spoken about, understood, and sold, belies that premise, as Joshua Koran, vice president, targeting and optimization, ValueClick, Inc., explains in the discussion below.
Behavioral Insider: You've voiced some dissatisfaction with the current usage and understanding of the term behavioral targeting. What's wrong with the way we conventionally use the term?
Joshua Koran: Behavioral targeting is in many ways a misnomer because it so drastically over-simplifies what's going on with the use of behavioral data.
It suggests it's strictly about serving ads. In fact what the technology is doing is building segments. These segments may be used for reporting, for personalizing Web site content and for analysis as well.
So when selecting among the approaches to BT, you need to specify what you're building segments for. All BT is not the same. People call very different approaches 'BT,' when in fact we need a larger lexicon. It's like calling targeting geo, demographics and contextual 'legacy targeting' -- it is too broad a name for what people are selling.
BI: If all behavioral targeting is not the same, how does a client go about distinguishing different flavors and matching methodology to their real marketing goals?
Koran: Within behavioral targeting you can distinguish multiple tiers based on the number of events they encompass. The first category would be behavioral segmentation based on a single event or action. The purest example of this is retargeting a customer who's added an item to a cart but never completed the transaction. There you have someone who's at the bottom of the funnel. This approach has proven to be very effective but because it's so easy to do, there's very little investment required and very little barrier to entry. So you have hundreds of different networks offering some form of retargeting with varying degrees of expertise in execution. Many, for instance, are very lax about analyzing recency and frequency into the way they retarget.
BI: What are the other tiers as you define them?
Koran: Approaches that segment using multiple actions require more sophisticated technology. The most well-known type of multi-event approach uses the principle of clustering. Tacoda worked hard to popularize this approach. What this entails is essentially grouping together visitors based on a lot of historical data about browsing, search and other behaviors to identify dominant interests. So by uniting certain types of behaviors sharing a single attribute, you generate categories like 'Soccer Moms' or 'Gadget Geeks.' This approach is quite popular and in some ways easiest to sell to marketers, because it's very familiar conceptually to them.
Marketers have generations of experience clustering consumers by zip codes, so it makes intuitive sense to them. It's also the way many large enterprises already organize their customer relationship management database. The drawback is that you are constrained to just one attribute per person, which does not fully reflect people's multi-faceted interests.
Another approach, epitomized by Revenue Science, is to apply business rules to generate each segment. Essentially you define a segment by saying, 'If consumer X behaves in a particular way, or performs a specific action Y number of times, then that makes them a member of a certain segment calling for a specific marketing response.' The drawback with this approach is that it takes a lot of trial, error and testing to determine which rules work. So testing is labor-intensive and thus can become very expensive.
BI: What is predictive targeting, and how is it different from what you've just described?
Koran: The difference between these approaches and the predictive approach of the Precision Profiles product we announced recently is that multiple attributes can be assigned to each consumer, and segments are dynamically generated based both on behaviors they've engaged in historically and the likelihood they will respond to a particular message in the future. The system constantly compares each consumer's past activity to what similar consumers have done in order to predict future actions for each advertiser.
BI: Does this replace retargeting? If not, how can the two be integrated? What kinds of behaviors lend themselves to predictive approach and which don't?
Koran: If you visualize it as a continuum, you can see retargeting works best at the very bottom of the marketing funnel. You know someone intended to purchase a particular product and your job is simply to remind them. Predictive behavioral modeling is focused somewhat higher up in the funnel. You know someone is interested in a product, and surmise intent based on their past behavior, but want to predict their likelihood of response.
There also remains a role for demographic, geo and contextual targeting at the higher end of the funnel. After all, many consumers on a given network don't have a behavioral profile, or have too sparse a history to predict anything. But you can target them because they look like other customers or have some similarities. You need to reach them at the top of the funnel as a first step in mining their interests to begin to establish intent.
BI: Is this something that's easily replicable?
Koran: There's a very high barrier of entry to do effective behavioral segmentation in general and predictive modeling in particular. The reason is that not all companies, even large Web sites and networks, have the right volume and breadth of actions. On the most basic level, you need a significant mass of page views in order to generate scale to provide really rich behavioral segmentation. Volume is important, but you also need the right kinds of transactional data. One of the things that is really distinctive about ValueClick is our access to a wide range of non-personally identifiable consumer behavior from our e-commerce and comparison shopping properties, making Precision Profiles one of the few behavioral targeting products to integrate conversion data into a predictive model.
BI: What are your goals for the learning and education curve?
Koran: The next six to 12 months will be a period where a lot of education still needs to be done about the different approaches to behavioral segmentation and the varied use of these segments within the catchall term behavioral targeting. Once that's understood, then behavioral segmentation can be really thought of within the context of a wider marketing strategy.