Statistical Modeling Meets B-to-B Marketing

Business-to-business marketers have been a bit slow in adopting the newest statistical modeling techniques to drive their marketing programs, with good reason. For one thing, B-to-B files tend to be relatively small. When your entire target universe is made up of a mere 10,000 prospective machine tool manufacturers, it’s easy to put off the expense and effort of building a regression model.

Another barrier has been the availability and quality of the data itself. Business marketers often use multiple channels, and coordinating data sources can be a strain. Worse, those marketing through distributors may be cut off from end-user data entirely. For example, a manufacturer may know where the product was shipped, but not how much was paid. Even if customer data is available, it can be hard to keep clean. According to D&B, a CEO changes every minute, and a new business files for bankruptcy every 8 minutes. If your data is inaccurate, why bother to model it?

But things are changing—rapidly. As B-to-B marketers gain sophistication and experience—and as they face competitive pressures—modeling is cropping up as a useful tool for a variety of applications. Models are being used for both acquisition and retention, with significant results. Just think: B-to-B accounts tend to represent much greater value than a consumer household, so it’s worth doing just about anything that will help you find more of them, or gain a larger share of their budgets.

Acquisition modeling
On the acquisition front, the most common application for statistical modeling is profiling current customers and using the profiles to identify high-potential “look-alikes” among prospect universes. This technique can work for companies of all types, but is most common among marketers whose products are of interest to a broad variety of business market segments. Sprint, for example, considers every business in the country a potential customer for its phone services. But targeting still makes their marketing dollars go farther, and improves the mood of their sales force. So when the Kansas City corporation pulls in prospect lists from compiled files such as D&B and InfoUSA, they apply a model to rank the names based on their similarity to Sprint’s best current customers and their dissimilarity to past non-responding accounts.

According to Tim Hand, group manager of CRM analytics at Sprint, the model’s value lies in refining the quality of the cold names coming in. “The model allows us to prioritize the leads,” he says. “In our experience, a rep will only make calls on about 5 out of a typical batch of leads we give them, so it’s critical that these names be as productive as possible. We used to select based just on SIC. With the model, we get much more detail, including credit worthiness, which is a key variable for us. Instead of giving each rep 100 leads in his or her territory, we’ll be sending 20.”

The reps have given Hand’s group positive feedback on the new leads—a gratifying if rare experience for a marketing team. But Hand also wisely set up a control group to confirm the power of the model beyond simply the anecdotal input from sales. “We took 3 or 4 months of leads and sales results, and compared the sales activity in the control group accounts against the accounts given to the sales teams. The model predicted the sales results successfully,” Hand observes.

Another example of modeling for acquisition comes from Abacus, whose B-to-B Alliance cooperative database contains 1.3 billion transactions from over 350 participating B-to-B mailers. Abacus clients are primarily cataloguers, in such categories as office supplies, advertising specialties (like logo merchandise), and seminars or training courses. These mail-order marketers live and die by new customer acquisition, and, unlike the industrial side of the B-to-B world, have a long history of cooperation with competitors to exchange names. The Abacus service allows mailers to select among millions of names based not only on the category purchased, but down to the individual transaction level.

The Abacus modeling approach is similar to that used in the Sprint example—but the volumes are simply gigantic. With its 2 full-time statisticians and 5 technicians, Abacus builds somewhere around 2,100 models a year for its clients, seeking to identify names with transactional behavior similar to that of the client’s best customers. The models consider 60 or 80 variables, and typically narrow them down to the most powerful 15 or 20. No surprise, the variables net out to the typical direct marketing drivers, like lifetime dollars spent, recency of the name on Abacus’s masterfile, and purchase frequency. After that, the variables are likely to reflect the product category (e.g., seminars) and then the specific offer.

Since Abacus updates its database weekly, and recency is often a key variable, most Abacus participants request a new model build for each new campaign, and they average 6 models a year. According to Steve Tinlin, senior vice president of B-to-B services at Abacus, the objective for most Abacus clients is to find additional names to mail profitably. Tinlin notes that many of his clients are able to select more than a million names for a campaign, compared to the 100,000 or 200,000 they might have been expecting from a database pull.

Modeling for acquisition is most effective when the results can be applied to large files. Another large source of names—one that does not require that you contribute your own files—is MeritBase, the cooperative database run by MeritDirect. With over 200 million gross records from 1,000 different lists, the best way to improve your results is with a model based the results from a previous campaign, versus a profile against your best customers.

The process works as follows: the client mails a likely sample from the MeritBase, pulled on the criteria typically used in B-to-B list selection, like list source, company size, and recency (e.g., hotlines). The responders and non-responders are then analyzed, a process that includes pass-along attribution of orders from a company where one person was mailed but the order was placed by someone else. One fourth of the names are held out for validation, and the rest are modeled. Each record in the MeritBase is then assigned a decile score from that client’s model, and when the client takes names for the new campaign, MeritDirect charges only for the names mailed.

The models allow mailers to select names in greater volume and with more efficiency. According to Blair D. Barondes, vice president of database marketing at MeritDirect, the results can be impressive. “We recently had the opportunity to review the results of actual prospecting mailings for several clients after the mailing based on their MeritBase models, so we were looking at a head-to-head comparison of in-the-mail performance. In all cases, the top two deciles performed at least 20% better than average, and some were as high as 60% better.”

Retention modeling
When it comes to retention, business marketers are applying modeling across the entire go-to-market process. Consider IBM’s experience with propensity modeling. IBM uses analytics to identify the target accounts most likely to be interested in certain product sets, such as hardware, software, and services. The worldwide small and medium business marketing teams have had particular success with this approach, says Theresa Kushner, director of market data and analytics.

Kushner’s team builds models for the inside sales teams, known as telecoverage reps. The models score current customers, based on past purchase, for propensity to buy by product category. The team then develops proposed phone dialogues, known as “reason to call” scripts, designed to suggest certain product offerings. The reps then contact the high-propensity accounts to pitch such offerings as a software product upgrade, or a set of new printers. “Our objective is to make the telecoverage reps more productive,” says Kushner. “We want them to be in regular contact with their top accounts, and have meaningful conversations with them. These scripts give them a good reason to call, and ensure that they are covering the market opportunity overall, and not simply calling the easiest and most obvious accounts every day. We now feed 26 models per year to the sales team in North America alone.”

How is the propensity modeling working? Kushner is pleased with the results. “We did worry about is the possibility that propensity models create self-fulfilling prophecies,” she says. “So we decided to test the process by replacing some of the top decile names with ‘placebo’ accounts from the bottom deciles to make sure we were on track. Glad to say, the results demonstrated a satisfying lift. Our conclusion is: don’t let the reps guess what accounts to call. Let the data tell you where the opportunity is.”

Propensity modeling qualifies as a classic example of data-driven cross-selling and up-selling. But IBM also uses statistical models for other retention purposes, among them:

• Channel management. IBM regularly models its customer base to support decisions about whether accounts should be covered by face-to-face sales, telecoverage, the e-channel or resellers.
• Account penetration. IBM models to identify opportunity to introduce an entirely new product line into a current account.
• Account reactivation. In the Asia Pacific region, Kushner’s team modeled dormant accounts against best accounts and sent the high-value names to the telecoverage reps to call. The revenue generated in the dormant accounts showed a 3 or 4 times lift from business as usual.
• Reseller channel support. IBM helps its business partners with marketing programs to help them promote IBM products to their installed bases of customers. Next step: develop a cooperative database with modeling capabilities to help resellers target more effectively.

One of the reasons retention modeling has become so varied is the availability of new tools to support current-customer analysis. For example, B-to-B has in the past few years seen the arrival of a PRIZM-like tool from Ruf Strategic Solutions that identifies 114 cluster profiles that can be applied to each company location, or site. The profiles attempt to define purchase propensity, based on such characteristics as sales per employee, company growth rate, consumption of raw materials, whether the site is a headquarters or a branch location, and the wealth, employment and crime levels of the geographic area.

Compared to regression modeling, clusters can be an inexpensive and fruitful way to begin statistical work on a customer file. Kurt Ruf, president, recommends that marketers begin with a “data audit,” appending profiles to their current customers. “These clusters describe the interdependencies of the business ecosystem,” says Ruf. “They help you tease out patterns and hidden dynamics. You find that the economic indicators like labor, finance and crime will impact a target company’s buying behavior.”

Future trends
What is the future of statistical modeling in B-to-B? A number of promising experiments suggest that the future is bright. Consider these examples:

• Loyalty Matrix, a data analytics firm in San Francisco, is helping a client identify which of its 17 e-newsletters is most effective in influencing sales results by matching actual customer sales levels with newsletter subscriber behavior. “We look at who opened and clicked,” says Steven Schwab, senior vice president of professional services. “The analysis then allows corporate marketing to make appropriate refinements in the messaging, by customer segment.”
• According to Anne Milley, director of analytical strategy at SAS, companies are beginning to use text mining software to identify otherwise hidden market trends and opportunities. Examples include a search among SIC filings, parts warranty claims, customer service complaints, or press releases—any kind of unstructured data that may prove profitable for parsing and analysis.
• SAS has helped HP get control over its myriad marketing campaigns with a tool called IMPAQ Express, a web-based database that gives campaign managers and analysts instant information on the right target audiences for the 30 or more campaigns HP runs a week.

B-to-B may have been late to the modeling game, but its payoff is likely to be huge.

Back to Articles and Columns »