That came to mind the other day when we interviewed ValueClick VP of targeting and optimization Joshua Koran. He is one of the made members of the BT (that's "behavioral targeting") mafia, one of the minds behind Yahoo's behavioral engine for a number of years before coming to ValueClick in 2007.

The ad server and performance ad network is not always known for its BT initiatives, so Koran brought us up to speed on their offerings. But in the process, we asked him to reflect a bit on some of the basics of BT, the major models and challenges marketers face in unraveling the complexity. While obviously framed within his pitching the ValueClick proposition, we found Koran's breakdown a good refresher course that we thought readers might enjoy.

Behavioral Insider: Where does ValueClick fit into the range of behavioral offerings that seem to be proliferating in the field?

Joshua Koran: There is a lot of confusion in the marketplace over what exactly defines a user's interest. There are three main approaches to generating that interest that all go under the rubric of BT. One of the earliest ones was a cluster approach where each visitor is assigned to one and only one cluster. This grew out of offline database technologies to cluster data and figure out what was the best representative segment for personalizing offers. Tacoda took that approach online and created these beautiful audience maps and bubble charts to represent to marketers which audiences were engaging with their particular message.

The drawback of the cluster approach is that it is assigning each person to one and only one segment [in a session] and most people have more than one interest. But you could classify and reduce folks down to 10 or 100 different clusters. I think that is why it is still popular today. The other benefit of that cluster approach is that it is automated.

BI: And the rival approach?

Koran: The custom business rules approach, popularized by Audience Science. Depending on if/then logic, each visitor could be in multiple segments. Marketers love the power to basically make up any custom business rule they like to classify their segments. The challenge with that is that it is a manual process. And how do you know if the rule [to define an interest] should be three events in five days or five events in three days? No one knows. You have to test. But technology is constantly evolving. Look at the new Yahoo front page. What may have taken you 20 page views a couple of years ago to check email and stock quotes you now get on a single page. Even if you tested, is it still is the right setting of recency and frequency?

BI: Where does retargeting fit in?

Koran: You don't need any complicated processing for retargeting. The barrier to entry is so much lower. You can store all of the information in cookies. It is actually the best form of BT you can use. The visitor is interested in a product category and is interested in a retailer. The drawback is, of all the visitors you find on the Internet, probably a minority have been to a given Web site.

BI: With that landscape in mind, lay out how your "predictive modeling" in the ValueClick suite is different.

Koran: I designed the Yahoo BT platform, which is also a predictive approach. You start with the marketer's goal in mind. Are they trying to maximize click-through rate, or conversion rate, or engagement? You look backward to find what were the activities that people who eventually clicked or converted did prior to that good event. And then you score other people based on their doing these similar activities.

The benefit of the predictive approach is that it is automated and it gives the ability for each visitor to have multiple interests. And it actually is goal-oriented. It has the marketer's goal built into the segment definition. It is not that these people are a soccer mom or gadget geeks or have some activity in the past three events in five days. It is that they're predicted to be doing something in the future, making a purchase in a given category or clicking on a given type of ad.

The main difference is that we adapt the segment to each advertiser. So it is not a generic interest segment of your interest in buying a car or in buying a sports car. But it adapts that segment to the particular advertiser that is running. There is probably a big difference between people who buy Porsches and people who buy Mercedes, even though both may be classified as luxury automobiles.

BI: But there are a lot of ad networks purporting to sell BT and a lot of confusion in the market about how best to define interest. How can marketers really evaluate differences among the methods?

Koran: It is a challenge. The differences will be on the access to data. What are the inputs or the scale of those inputs? Second, maybe even more important, is what is the process they use to aggregate or compute on top of that data? If you have great data but a bad process, then probably the output is not going to be very useful, and if you have better algorithms than anyone else out there but no data to feed into it, then likewise your output won't be very useful.

Once you've got a large scale of inputs and a good process on top of that, the final differentiation is your ability to deliver that. That is the other misconception out there, that you should either be buying contextual or you should be buying audience segment. But really you should be buying both. Buy that auto intender in the right context, and sometimes it is while he is comparison-shopping and sometimes it's while he's reading an article. Let the optimization of the ad server figure out what is the right context to communicate with them.

BI: How do marketers really know which methods works best?

Koran: We believe you should not have to make that choice upfront. You should work with a provider who has access to all of these approaches and is optimizing your campaigns based on whichever approach happens to be working best for your particular campaign.

That is why we are not a closed environment where we only sell you our own BT. We do not see 100% of the visitor activity online, and we never will. We're also selling other people's behavioral segments. We have about 500 million of our own profiles and about 100 million of third-party profiles hosted in our database.

For your campaign, maybe a portion of it should be delivered through someone else's process. We work with the majority of the profile retailers, about 20 different partners. From my perspective this is where the future of the online ad network will go: the people who have the best ad servers that can take in other people's data and match it against current context to identify what is the right ad to show to any given visitor at the right time.