If you mention the word "chocolate" here and there in your online blog and comment posts, the odds are pretty good you are also interested in Audis, lacrosse, Easthampton and weeds. Huh? One of the fun aspects of digital tracking, especially behavioral segmentation, is how the data reveals unexpected affinities among groups and interests. According to Peerset, an ad targeting venture that leverages social data in a unique way, our interests are reflected most accurately by our expressions online rather than our browsing history.
"We're like a Pandora of advertising," says Mike John-Baptiste, CEO. "We understand what people like from what they express about themselves. We create correlated interests." It turns out that people who mention "Starbucks" also show an interest in John Mayer at a much higher rate than they do for Beyonce.
Peerset gathers data from publicly available sources of user-generated content that are already fairly well structured. So social networking and dating sites, bookmarking sites, etc. are places where people tend to express their tastes and voice their interests explicitly. A patented technology scours this information without tracking users. Peerset simply is taking a picture of personal expression to chart the correlations in tastes. Chocolate mentioners for some reason also have an interest in weeds. Go figure.
But that is exactly what Peerset tries to do when it takes that one set of data on correlated interests and applies them to ad targeting among its publishing partners like Hi5 and Justin.tv. "If Joe on Hi5 says he likes John Mayer and romantic walks and basketball, we take those three interests and, based on those three things, we see how relevant he is to Starbucks," says John-Baptiste. Basically it is a predictive targeting technology that uses previous expressions and statistical analysis to anticipate the likelihood of that person's interest in the marketer's product. Among publishing partners Peerset addresses about 12 million users, but it is also partnering with demand-side platforms to reach a much larger potential audience. The DSP can take the technology and apply it against the data points they already have on a user.
John-Baptiste argues that looking for correlated interests based on explicit personal statements is a more reliable use of social media than some other methods. Leveraging the social graph may render the friends of a chocolate lover, for instance, but only so many of those friends are also chocolate lovers or favor a particular brand. Peerset is not actually interested in the social graph, per se. It simply uses social media content because it is the best place to find people stating the full range of their interests. "We take people independently of friendship," he says. "We just care about what they like."
The interest correlation approach also tries to address some of the scaling issues that traditional behavioral targeting has experienced, because a predictive model widens the segment. "In behavioral the big issue is this binary decision between reach and precision," John-Baptiste says. Because traditional behavioral targeting is segmented according to specific actions a user had already taken, that set can often be prohibitively small. A predictive model allows a marketer to define an audience based on what they are likely to want rather than strictly on what behaviors they have already demonstrated. "We can usually find a range roughly between 5% and 30% of our user base for a brand," he says.
Peerset tested the targeting approach in a Microsoft Windows 7 campaign a year ago. Microsoft was interested in scale and in buying non-guaranteed inventory. But company strategists wanted targeting applied to reach the right people. Peerset's approach was to find people who were logging into a site under a Windows operating system but who showed an affinity for Apple and might be tempted to switch to Macs -- likely customers for the more-Mac-like Windows 7. It turns out that simple expressions of "creative" and "video" in social expressions correlated well with an affinity for Apple. Peerset partnered with Cadreon on the campaign, and the click-through rates on this targeting approach were three to five times better than the average for the publishers' sites, the company claims. John-Baptiste says the company is now running its 17th campaign with Cadreon.
Exploring correlations in taste really became possible only when people started expressing themselves so relentlessly online. "You can target on psychographics," he says. "The introduction of Web 2.0 allowed people to express themselves to the world, and we can deliver advertising based on what they said and what they asked for." But predictive technologies go even beyond that. "We principally believe we understand people's tastes, the hard wiring, the things people don't even say themselves."
And if the model holds true, it leads to some puzzling correlations that may make sense statistically, but seem nonsensical otherwise. Do you like chocolate? Interested in a weed killer with that?