The problem with the word "segmentation" is that it has lost its meaning. Like "analytics", a marketer saying "segmentation" can mean virtually anything. This isn't necessarily a problem, if we are simply using "segmentation" to mean "the parsing of customers in some way", but often people mean something very specific when they say the word, while the people listening interpret something very specific as well, but something specifically totally different.

So, here we go. In my view, any segmentation starts out with business goals. What is the company itself trying to do with its customers? Are we trying to grow profit? Grow revenue? Are we trying to do it in a specific area? Do we have multiple goals? What are their priorities? Ignoring these questions and just going down the path of segmentation without context is a big problem. So start by understanding business goals. The two Bobs would be proud of you, even just doing this.

• Product. What customers want and need from a product or service. Sometimes called "needs-based segmentation"
  
• Creative. How customers respond to different visuals, animation, etc. Some of the cool things going on in biometric research are very useful here.


• Promotion. How customers respond to different types of incentives... price, points, free shipping, etc.


• Channel. Which customers use which channels, and which they prefer.


• Targeting / Finding. Where to actually find customers wherever they are
  

This is represented in my little "segmentation tree" as a continuum, and that's deliberate. This is because the left side of the tree--needs-based segmentation--is really hard to combine with the right side of the tree. Also, making a segmentation scheme do all five things at once is almost impossible. There are two reasons for this:

1. A statistical problem called "assignment" which means that you can't make one segmentation based on needs, generated from a survey, fully predict or "find" individual customers


2. A scope problem that in any one study, you can't possibly combine all elements of "segmentation" into one uber-model. It's just impossible.
   

One way around this is with behavioral segmentation, or "deterministic" segmentation. This means you just start out segmenting your customers into groups 100% defined by their behavior, and then define the six attributes in the tree post-facto. This way, you're always certain you can target a customer. It might not be perfectively descriptive, but it's guaranteed to be perfectly actionable.

Later everyone.

The hottest thing in direct marketing today is targeting individuals online vs. targeting via publishers or content. The idea that we can "know" a user and target them frees up the 90% of online inventory currently referred to as "remnant" space-- e.g. the space left over after GM, Microsoft and Coke buy the top page banners at Washington Post and NY Times. The remnant space can be just as valuable as the prime space if we only knew who the users were. A user interested in gourmet cooking is still a user interested in gourmet cooking when she leaves the NY Times food page and goes to her hotmail account or to her son's preschool website.

Verticalizing the internet is a potential solution. Or rather, verticalizing internet users. The internet is already as verticalized as it's ever going to be. So how would one verticalize internet users? First of all, it's only going to happen on specific networks of sites. You can't, for privacy and chinese wall reasons, put a universal cookie a browser that decodes them perfectly everywhere. I guess you could, and maybe that's an interesting business model. But for now, you need to work within the context of a network or a publisher. Let's use Yahoo! as an example.

So, say Yahoo! decides that they really want to make their inventory appealing for technology marketers, specifically B2B technology marketers. One option would be to put up pages that are really appealing to B2B technology buyers and influencers. These pages, assuming they attract a lot of volume (another key issue for networks / publishers), would get high CPM / CPC from technology marketers. However, once the user leaves this site, they're still valuable, right? So here's the rub, and how you can actually verticalize the internet. You've got the site; you've got the interested users. So you're missing two steps: (1) follow them around, and, (2) identifying the same types of users regardless of where they are on the property. I'd call this "200-level" and "300-level" verticalizing.

200-level verticalizing. Following them around is pretty easy. This basically means finding the really good vertical targeting-type pages, tagging engaged customers, and making sure we follow them around. This is pure behavioral targeting. We can get quite fancy doing this, by adjusting content on the verticalized page to correlate to segmentation dimensions and then adjusting people's segments based on what they click on. The trick seems to be in the details on this. For as long as I've been a member of Microsoft "Live", for example, I still seem to only get display ads for Hannah Montana movies. I view this as low-hanging fruit.

300-level verticalizing. The real nugget will be identifying all traffic according to vertical. There are two ways to do this. The first way would be doing traditional market research, such as a quantitative instrument, that would be distributed to a broad cross-section of a network or publisher's traffic. All respondents would be pre-linked to as much "targetable" behavioral data as possible. For example, Microsoft could use all the cookie data it aggregates for Windows Live ID users. Then, an analyst does a latent class analysis to derive segments relevant to advertisers; does a multinomial logit model to assign; and hands it off to the ad sales team. A/B testing to determine incrementality, and you're off to the races.

The second way would be to use only behavioral data to create "mock survey data", specifically from the vertical-relevant sites. So, we'd define our orthogonal dimensions that would be relevant to advertisers, such as "self integrators vs. needs help" and "windows vs. open source" etc. for a B2B technology scenario. Then we'd create content on our verticalized site that reflected these dimensions, and start measuring who clicked on and was otherwise engaged with what. We then use the same exact technique I outlined above to define and target segments.

The promise of these approaches is, I think, pretty huge, specifically for B2B where it's much more important to be much more targeted online.