Merchandising As Targeting, Targeting As Merchandising

Merchandising As Targeting, Targeting As Merchandising by Phil Leggiere, Wednesday, June 6, 2007 THE METAPHOR of the auction place has captivated the online world of late. Thus far, however, applications of the model have been divided into two separate realms, e-commerce transactions (a la affiliate marketing) and conventional banner or text ad placement. The next iteration of the model, Pradeep Javangula, CTO of Tumri, predicts below, will involve a new synthesis of the ad network and e-commerce: a merchandising network, with product offer delivery optimized by behavioral and other forms of targeting. Behavioral Insider: Tumri doesn’t seem to fit into any conventional category of ad network. Could you describe what you do? Pradeep Javangula: We call what we do automatic personalization. Our background is in artificial intelligence and the semantic Wb and our focus is on classifying large volumes and content, specifically content related to product offers and services. So we see Tumri as a targeted merchandising network for delivering product offers, which may include ads but are more than just conventional banner or text ads. We think in terms of offers rather than ads. Our clients include many of the most successful online merchants, over 1500 merchants and 500 brands. We match up their offers to the right publishers and consumers. BI: How do you make your merchant catalogs more targetable? Javangula: Shopping catalogs have enormous amounts of very rich information. In many ways, the Web -- or very large parts of it -- can be looked at as one big catalog. Take any large e-tailer. The sheer quantity of information they have online is enormous. The opportunities for organizing and, by extension, targeting product offers based on that data are similarly enormous, but overwhelming for most marketers. Latent in all that data, though, are all sorts of demographic and psychographic attributes. Across all product categories for every brand, you can associate demographic and psychographic markers or tags. Take a Polo shirt, [for example], and its higher-end demographic and more trend-oriented buyer, [while] Lee Jeans are for the more price-conscious and practical-minded clothes shopper. What we do is then classify every piece of product information and give it a demographic and psychographic tag. BI: Where does the consumer come in? Javangula: Then the next step is to find the most relevant match between every user who visits a publisher Web site and the product content they’re most likely to be interested in based on information about who they are. When users visit a Web page, there’s an IP address associated with them which tells you the location they came from -- which in turn gives you a good idea of their demographics as culled from census data. We look at the time of day they visit the site and can infer what they’re interested in, based on profiles of items typically shopped for at that time of day and the demographic profile of people who populate the site that time of day. Very important, you can see the URL the user is coming from. We categorize every Web publication by content analysis in much the same way we extract demographic and psychographic attributes related to product catalogs. BI: Could you discuss the behavioral component? Javgula: Finally, if they’ve clicked through to the Web site using a keyword search, you then move from demographic information to purchase intent. So from the moment a user enters the site we have begun to frame a fairly detailed profile of who they are and what their interests are -- a profile that will be continually refined based on other publications they go to. From there, we begin to evolve a more behaviorally based profile of them based on product browsing and sales history. BI: How do you do that? Javangula: Within the publisher site itself we look at individual pages as proxies of a user’s intent. Say one million users have been to that page. Each one that comes brings a profile that validates, conflicts with or in some way adds to the brand, demographic, psychographic or behavioral attributes we associate with that page. So page by page the system is continually learning from experience. Based on everything we know, we do a real-time algorithmic decision based on our best hypotheses about the most relevant product offer we can introduce. BI: What kind of options do publishers have in selecting which offers will run on their sites? Javangula: Basically as much or as little as they want. We have them create a very easy-to-use template or wizard describing the kind of offers they want to run, which can be customized by categories, merchants, price range, keywords and other criteria based on their own user profile. Many give us a general set of criteria and then use our technology to drill more deeply. Others want more control. Either approach can be accommodated. BI: What kind of response is there from clients? Javangula: Once they see major improvements in click-through and conversions, most of them get the model very quickly. We think the boundaries between the advertisement, as it’s traditionally been known, and overall content, are becoming more and more blurred. Clients say to us ‘Wow. This is an application, not an advertisement’ -- and that, in many ways, is the best way to look at it.