As the relationship changes from casual browser to interested prospect and then to active buyer, your Web site can collect a whole host of personal preferences, plans, penchants, and peccadilloes. You need merely ask. Coupling that information with offline data from surveys and commercial databases, you now have a chance to build a fully realized portrait of each customer. Just how balanced is your portrait?

In the end, you have to decide what granularity of identification will suit you best. Is it enough to know the following about a certain type of visitor?:

• Interested in product A

• Works in industry B

• At a company that's size C

• Is moving through the qualification process at a rate of D

Or, in order to sell more, faster, at a higher margin, does it add significantly to the bottom line if you also know the following about her?:

• A shopping type E (adventurous explorer)

• Reviews the F section with a frequency rating of G

• Wears size H shoes

After all, just how valuable is it that you know which customers like green marsh-mallow moons in their Rice Chex?

While the sheer quantity of data elements is the most significant factor in your personal profile depth scoring, each element must be weighted according to its value. A customer identification number has no weight at all because one is indistinguishable from the next. The usual information collected about a customer (name, address, phone number) is critical but carries a low weight because it is not actionable.

Types of Customer Information

Actionable information pertains to a customer's predilections, purchasing history, and declared interests. Be sure you properly weigh implicit, explicit, and factual information:

• Implicit: He looked at those pages so he must be interested in these items.

• Explicit: He filled out a survey and told us he was interested in these items.

• Fact: He looked at those pages and bought these items. Web site visitors tell you explicit information, and you derive implicit information. For instance, a customer can say he likes reading biographies and wants Amazon.com to email notifications about famous figures in European history. But if he buys books about dogs, Amazon knows what to put on his recommendation list. Watching what customers actually do is far more revealing than reading what they say. And it's revealing in ways that don't necessarily make sense.

Suppose the database shows that visitors who are shopping for electric razors are also buying personal CD players, or that visitors who read the detailed specifications of the surface mining and construction equipment are seldom interested in extended warranty information. Are these the sorts of correlations marketing mavens are going to come up with in brainstorming meetings? No. They don't make any sense, but they're true. So now the marketing mavens have a new datapoint to work with, and the systems behind the scenes have the ability to act on the information in real time.

Besides weighing data elements based on whether they are declared or derived, their value must take into account freshness and results. Knowing the correlation between electric razors and personal CD players is the first step, using that information is the second, and measuring the results of that use is the most important.

Data Cleansing

Living up to customer relationship management means ensuring that the information used by the marketing systems and the customer service representatives is fresh, current, and accurate. That means bringing data together from many systems and that means figuring out how to get all that data to look alike.

Data normalization has typically applied to the format of information that is entered into a system. Does the middle initial carry a period? Does the phone number include parentheses or dashes? Is the zip code five digits or nine? Is there a hyphen in the middle? But in these days of CRM, data cleansing goes far beyond punctuation.

Let's assume you have a sales contact management system, an invoicing system, and a customer care database in each of four divisions. Let's say John Smith sends you an email from <JohnSmith@Yahoo.com>. Which John Smith is this?

You'll need multiple points of comparison. Maybe <JohnSmith@Yahoo.com> let slip that he was having trouble with your product while he was in California for the first time. You can then eliminate all of the John Smiths who live in California. Possibly he mentioned which product or service of yours he was using. Perhaps he includes his phone number in his email signature file. That could be the clue you need to identify this John Smith from the twentyseven others in your database.

Data cleansing focuses on the verification and validation of the information. If all your John Smiths are formatted the same, you're off to a good start. If none of your John Smith records have been verified for more than 6 months, their value deteriorates. I'm trying to clearly depict a set of problems that are neither easy nor quick to solve. If it's going to cost so much and create such pain, how do you go about measuring the value of all these possibilities? The question is whether the cost of collecting and processing the information is worth the value you derive from having the information, less the pain you cause your customers in its collection.

Personalization Quotient

Dr. Kamran Parsaye, president of Intelligence Ware, Inc. and author of Intelligent Database Tools and Applications (John Wiley & Sons, 1993) wrote a white paper called "PQ: The Personalization Quotient of a Website." At the moment, the paper can be found online (www.kellen.net/ect586/personalization_parsaye.pdf), even though the company Parsaye worked at when he wrote it (NovuWeb) cannot.

In his paper, Parsaye made a valiant attempt to create "a framework and a theory to measure how personalized a system is in terms of the Personalization Quotient (PQ) and illustrate how the theory can be used to improve e-service." The concept of the personalization quotient is then used to measure how personalized a system really is.

In this paper, Dr. Parsaye differentiates between an impersonal system, which treats everyone the same way, and a fully personalized system, which adjusts its behavior to specific users. An impersonal system has a PQ of zero, since it provides the same static response to all users regardless of their characteristics.

Personalization comes about as a reaction to individual information, and Dr. Parsaye divides personalization into three areas—customization, individualization and groupcharacterization. Customization is the oldest and at times the easiest to address. It allows you to set specific preferences, e.g., the stocks you want to track, the type of news you want to see, the colors you want set on your screen, etc. Individualization goes beyond this fixed setting and uses patterns of your own behavior (and not any other user's) to deliver specific content to you. [For instance,] if you have clicked a lot on finance-related items but not on sports, it will show you more financial news rather than sports news, without your asking for it. In group-characterization you receive a recommendation based on the preferences of people "like" you, e.g., books may be recommended to you based on books ordered by people with similar interests. Approaches based on collaborative filtering, case-based reasoning, etc. focus on the group-characterization measure.

• PQ: The Personalization Quotient takes all three of these issues into account.

It has three specific components, PQ1, PQ2 and PQ3, where:

PQ1 measures the system's ability to customize items.

PQ2 measures the system's ability to use individual preferences.

PQ3 measures the system's ability to deal with group-based preferences.

We then measure PQ as the average of these two elements, i.e.:

PQ = (PQ1 + PQ2 + PQ3)/3

Here each PQ1, PQ2 and PQ3 will be a number between 0 and 100. A system with a PQ of 100 is totally personalized, while a system with a PQ of zero is totally impersonal.

Dr. Parsaye then describes creating an ultimate profile of your site visitor:

One way to represent and measure similarity of users and customers is in terms of DNA strings or attribute vectors.

A DNA string for a Web user is a set string of integers between 0 and 9, e.g., the string 1309735183291. Each integer here shows the relative value of some trait, e.g., scoring an 8 or a 9 on the "sportspage" indicator means that you view many sports-related pages, while a 0 means that you never see such pages at all. Similarly, other integers on the string can tell us how you visit the site and how you click through on banner advertising—all in relative terms. Similarly, we can define a DNA string for a Web page by considering the components that comprise it. For instance, the number of banners and the type of banners.

He concludes by suggesting "An interesting direction for enhancements will be that of measuring the comparative PQ of two systems."

He then wanders off into a world where only mathematicians dare to tread by slipping into some serious formulae such as: PQ3(U, P) = 100 / maximum((ä U / ä P), (ä P / ä U)). But how do we factor in the pain caused to the site visitor who is followed around from page to page by a cookie and asked for an opinion about whether a woman's life is fulfilled only if she can provide a happy home for her family? That's where the personalization index comes in.

Personalization Index

The universe of profile elements is virtually unbounded, covering familiar items such as last name and business address, technical concepts such as IP address and connection speed, and domain-specific attributes from pore size (for cosmetics) to lifestyle risk profile (for insurance). By adding incremental profile information, e-business managers are able to move prospects and customers through the four stages of e-customer understanding, transforming category 1 anonymous users into the distinct, real-world category 4 individuals.

Collecting information is one thing. Using it in a judicious way is another. The personalization index (PI) distinguishes just how well you are using the data you are gathering. The index is a measure of how effectively an e-business is leveraging this customer data.

If your PI is above 0.75, then you are making the most of the information you are collecting. That means your efforts are not wasted, nor are those of your customers who are providing the raw material.

The preceding assumes that you are using a significant number of elements to make a personalized Web experience. If you are only collecting two elements and using them both, your PI score may be 1.00, but here it means you are only going so far as market segmentation rather than personalization—you're only grouping your prospects and customers into broad categories. While useful, broad categories aren't as powerful as true personalization based on dozens of attributes.

When you collect more and more elements, you can classify users into more and more clusters, and broad segmentation moves toward personalization. This is where you start to foster a customer relationship and turn it into a loyalty relationship, significantly raising the cost for your customer to switch to another vendor.

If your PI is less than 0.30, then you are collecting more information than you are using. The good news is that you have a huge untapped reservoir of actionable data about your customers. The bad news is that the data is lying fallow and probably getting stale fast. You need to either start using the data you have more effectively or cut down on how much explicit data you are trying to collect. Most likely, the correct answer is both. You are spinning your wheels collecting that information, but you are not using it to benefit your customers, which adversely affects your customers' experience.

That's the greatest downside to a low personalization quotient. All that time and effort that you force your customers to invest in giving you information is a waste. They get nothing out of it. Even when the process is simple, such as scanning a key chain fob at the grocery store, there's still no real value to the customer. Why bother? Why are they being bothered? At this point, we have finally attracted, navigated, persuaded, and converted that unknown prospect into a known customer. Can we get that customer to come back?