Does the online lead-gen industry need a third-party scoring service?

A growing number of events suggest the answer is yes. Numerous reports of fraudulent and bad-quality leads, deceptive lead-generation tactics, FTC investigations and fines have rattled the confidence of performance advertisers and turned off consumers. When it comes to purchasing new consumer sales leads from online lead generators, some marketers still blindly assume that all leads are of the same quality and should command the same price, or at best they are able to discriminate by source alone -- not by the individual lead. Such an individual quality score would open up whole new pricing models between lead buyers and sellers and allow advertisers to dial up or shut down lead sources much more quickly. In this way, predictive scoring holds the potential to upend the entire business model of online lead generation, with marketer expenditures approaching $2 billion annually.

So why is the lead generation space behind the curve in using predictive analytics?

For one thing, it's hard to do, especially if you classify yourself as a math-challenged marketer. Traditional, in-house methods for developing predictive models can often turn into extremely complex -- not to mention costly -- endeavors. For even the biggest companies, it can border on rocket science, even when there are powerful computing resources at hand. Hence, many turn to third-party service bureaus and so-called modeling shops.

Over the past few years, these firms have been grappling with ways to reinvent and simplify this moon-shot technology so that each and every marketer doesn't have to reinvent the space shuttle for its individual lead generation efforts. So far, it's been shown that it's possible to predict what an online consumer sales lead will do before a marketer actually follows up on that lead. The result is that marketers can judge the quality of the leads presented to them before they actually buy or act on the leads.

Here's how it works.

1. Start with a sample file of past purchased leads, with known outcomes attached (which ones converted, which ones didn't, etc.)
2. Append external identity, demographic and behavioral consumer data to the file in order to generate a rich set of modeling variables.
3. Define what you are trying to predict -- that's your "dependent variable."
4. Using statistics software, generate a regression model around the data and modeling variables.
5. Score and segment all your leads using the new model. Does it predict well?
6. Create go-forward policies around different score ranges. For example: reject the very lowest scoring leads, send the highest scoring leads to your best call center agents, follow up on lower scoring leads with email only, etc.
7. Start scoring all incoming leads as you get them from lead sellers. If you see very low scores from a new source, or a source's scores trending sharply downward, you know you have a problem right away.

Say you are in the business of online higher education, one of the most robust segments of online lead generation. Lots of people find your offers online, click for more information and then register to attend, say, a required follow-up enrollment appointment, perhaps at one of your brick-and-mortar locations. Those familiar with online education won't be shocked to hear that a big portion of these so-called new student leads who register for an appointment probably won't ever show up. But what if you had a pretty good idea in advance who the no-shows would be so that you wouldn't waste any time with them?

In this example, from a sample of 30,000 online education leads, a predictive model was constructed based on the information of the actual people behind those leads. And, similar to a credit scoring process, each lead was assigned a score between 0-999. At the bottom of the range -- with scores of between 1 and 57 -- was a group of people tagged as very poor quality, who were given almost zero chance of showing up for the follow-up meeting. These represented more than 25% of the leads. At the top of the food chain -- with scores between 437 and 741 -- was a group of people who collectively had a 41.1% likelihood of showing up. With scores in hand, the online education provider whose 30,000 leads were scored could have kicked back to its lead-gen provider(s) the leads it didn't want to bother with and purchased only those most likely to convert.

A similar exercise was performed on a large sample of satellite TV leads. A custom score was developed to predict which leads were most likely to convert into actual new, paying satellite TV customers. If the company had gone through this process before it attempted to convert all of the leads, it would have been in a position to focus its budget and call center manpower on a refined group of prospects most likely to qualify and buy its services.

The techniques outlined here have their critics. Some lead-gen providers that are used to being paid for bogus or very low-quality leads would much prefer to do business as usual. But while most people would gladly accept, sight unseen, the prize "waiting behind door number two," few of us would buy merchandise we could not at least assess beforehand, unless we already were familiar with the merchandise and were buying it from a reputable vendor. Online leads should be treated no differently.

This commentary is insightful. I recommend it to others.