A couple comments/clarifications to add to the thread:

@ Jeff Braddock: The approach is described is absolutely not retargeting. As I explicitly pointed out, individual users are not retargeted. A behavioral profile is constructed based on a small core of purchase data and this profile is used to identify target users. My point is not that look-alike modeling isn't a valid approach. My point is that BT isn't the only way to "deliver target audiences". "moms aged 25-54" is a target audience, it's just not a precise a target as BT sales forces like to claim. But as the example shows, the claims are far beyond the reality and based on the data that I have seen BT doesn't outperform other ways of delivering target audiences if the goal is driving offline sales.

@ Alex English: Good question. Fortunately, Nielsen did release normative data on a 10/6/09 Citi Research conference call. John Burbank, CEO of Nielsen Online, quoted the average ROI across the >200 offline impact studies they have run at 157%. That's a useful benchmark. I do want to reiterate though that this article is in no way intended as an attack on Nielsen. As I say in the article itself, I think the work they are doing with their major media partners is in many ways the gold standard of BT. That's why it provides such a useful case study in the limitations of the approach.

Keep 'em coming!

Having said that I do this this example is unfair. Let's say that the look-a-like models aren't delivering 3X ROI but the real question that isn't addressed in the example is, does the look-a-like modeled campaign deliver more ROI than a non-targeted campaign? That's the unanswered question. Nielsen needs to release some normative data so we can really understand the value of their targets.

What ad networks are selling is site retargeting, not behavioral targeted audiences. To sell audiences, it requires companies like Andy's and his competitors to start investing in data. That means not taking shortcuts using cheap data resources like Quova and Digital Element to get the incorrect ZIP+4 information and applying the wrong demographic and psychographic data points.

In Andy's example, the point he misses is that Nielsen data will only work well if it can be matched properly. Look-a-like models work in the offline, so there is not reason they won't work in the online world. When the ad network refuses to invest in data that improves the matching, the campaign fails before it starts.

Bennett Zucker is correct that progress is being made in multi-variate targeting, and the limitations start and end with Ad Networks that simply display no interest in finding audiences. CBS Interactive's recent decision to ban ad networks evidence enough that ad networks need to move towards audience targeting if they plan to survive past 2010.

BT was the "it" solution for a while, but over time it's become a tool of choice only for pure remarketing and special situations such as auto and travel. You're right that no one has produced credible evidence that BT has done a whole lot for most categories, least of all CPG, which requires far more predictable reach and frequency than targeting, by definition, can deliver.

I applaud your effort to shed some light here, especially given your experience at Yahoo, which is generally regarded as among the best practitioners of the art and science of targeting.

I do think that the recent trend to multivariate targeting of audiences is a step in the right direction, but reaching television-like scale online is always going to be a big challenge for many ad categories.