• by Phil Leggiere , March 21, 2007

"Know thy customer" is or should be the mantra of any advertiser committed to longevity in the digital age. Unfortunately, all too often what advertisers THINK they know is based on static, quickly obsolescent models both of who customers are, and what their behavior says about who they are. In the conversation below, Phil Kaplan, co-founder of AdBrite ad network, outlines how behavioral and demographic profiles can be dynamically related, to add currency and vital context to both.

BI: You've said that behavioral and demographic targeting work symbiotically together at AdBrite. Could you explain how your approach to both areas has evolved?

Kaplan: Like most ad networks, we started as just a list of Web sites. That was manageable when we were dealing with 500 or 1,000 sites. But as you move up to 40,000 sites, it ceases to be manageable in the sense that you can't just pick out sites by hand. So the natural progression was to work on providing detail and transparency about which sites advertisers could advertise on, But what we started to realize is that what they were interested in wasn't really so much what sites their ads were on -- but who it was they were actually reaching. When advertisers started asking us whether we could reach 18- to 24-year-old men who lived in San Francisco and were interested in automobiles -- that's what made us pursue the question of how you identify where, across your thousands of sites, the paydirt really was, particularly the long-tail sites.

BI: What data points did you use?

Kaplan: The first thing we did was not original in and of itself, certainly. We ran ComScore data on all the sites on our network for which it was available. So we found. to take an easy example. that Maxim.com was 80% male. Then we used U.S. census data on users. You can get a good idea from IP addresses what zip code someone is from -- and if you merge that with census data, then in terms of users you can say. with pretty solid accuracy, that if you live in zip code X you probably fit a certain income, ethnic and other demographic fairly closely.

BI: So at that point it seems like you were still working within the conventional parameters of what most people would consider content or demographic targeting.

Kaplan: Right. Where it got interesting is when you start relating both sets of data dynamically. Say we have someone visiting Maxim. We'll say OK, this visitor is from X zip code so we'll make some educated assumptions about their income demographic and because they're visiting Maxim we'll say there's an 80% probability they're male. Of course if the next 3 sites they go to are Women's Wear Daily, Vogue and another heavily female site, our assumption about gender begins to have to get readjusted. So once you begin doing this for every user for every site they visit , you've soon got a very good working demographic profile based on behavior. Essentially every time one of our users visits a site, a recalibration of our original profile assumptions gets made.

BI: What happens when you get to so-called long-tail sites that Comscore or the others don't track yet?

Kaplan: There are tens of thousands of sites that Comscore doesn't have data on - blogs, for instance. The way we build a profile on those sites is by associating them with the profiles of the users who visit them. So again we begin by making educated assumptions or hypotheses and then keep testing them against emerging behavior patterns. If the first visitor is rated female, [with] a certain ethnic and income segment, those are the characteristics we ascribe to that site, which are then adjusted for each subsequent visitor.

BI: How did or do you position your approach to clients?

Kaplan: At first we didn't describe what we were doing as behavioral targeting. We didn't really even tell anyone about it. But we actually found advertisers -- particularly brand advertisers -- would push us. We'd have an auto brand and they'd say, we'd love to find males 35-44 in a certain income bracket who live in this area and are in the market for pick-up trucks, and we'd say, we can do that. Then they'd say. we don't believe you. Tell us how you think you can do that. And so we'd tell them. We do periodic tests to see if our calculations are really accurate. We take a larger site which has a very well delineated ComScore database and then compare that to the data we've integrated for that site, by doing a blind targeted ad test for each set of data. We've found that most of the time our profiles are spot-on in being more predictive than targeting based solely on the original demographic profile information.

BI: What do you see as the priority areas going forward for advertisers, publishers and the learning curve?

Kaplan: We see behavioral targeting as just beginning to come into its own for brand advertisers. The biggest gap or inefficiency in the market right now is the belief on the part of brand advertisers that that the only way of insuring they reach the right customers is to go to the top-tier branded content sites -- and to do that, they'll gladly pay 10 or 20 or 30 times more than they would to reach the same size target audience by finding them on a spread of smaller sites. That's a knowledge and perception gap that for many is hard to overcome, but we expect that to change going forward.