How to Build a Lead Scoring Model

When figuring out how to build a lead scoring model, you choose the indicators that predict whether the customer is a good one, assign each indicator the appropriate weight, sum them up, and set the threshold beyond which the lead should trigger an action. That is the entire process. The difficulty lies in choosing the right indicators and identifying which ones actually predict revenue generation.

This article provides a detailed guide to creating a scoring model that your team will trust and act on, rather than a complex one that will be forgotten after you create it. If you first want to understand what lead scoring is, start with the definition; here, we assume that you already know the concept and are ready to build the model.

Start from the outcome, not the signals

The most frequent mistake that is made during the creation of the model is starting from a list of measurable indicators and giving them certain points. As a result, you will get a model based on the easy-to-track indicators but not on the actual prediction of the revenue generation.

Work in reverse order. Start from the most valuable customers and analyze what they had in common before turning into the customers. What did they do? How did they interact with the product and marketing of your company? What characteristics did they have?

These patterns are the indicators that predict the outcome and form the ground of your model. The model created using “what our valuable clients did before conversion” is realistic and the model created using “what our tools can track” is convenient but not predicting anything.

So, the first step in the model creation process is not technical. You just need to honestly evaluate the customers and the pre-conversion activity.

Separate the two types of signals

Once you define the signals that you will use for creating the model, divide them into two categories as they play different roles.

Fit signals characterize the lead: the company size, industry, position and geographic location. Such signals help to determine whether the lead could become your customer: do they match the profiles of accounts your company serves?

Behavioral signals describe the actions performed by the lead: started the trial period, completed the onboarding, performed some important actions, invited team members, visited product multiple times and reached usage limits. These signals help to understand whether the lead is engaging in the activity, what is a strong indicator of the lead becoming your customer.

The reason why you should distinguish between the two types is the following: the two types are answering different questions and thus should have different weights. The fit signals qualify while behavioral signals prioritize. So, the lead that perfectly matches the profile but does nothing is not ready for conversion yet, but the lead that shows the deep engagement despite the mismatch of some signals may still be interesting for your company.

The best models use the fit signals to determine the candidates for scoring and use the behavioral signals to determine whether the lead is ready to convert.

If you need to choose where to place your weight, pay attention to behavioral signals. They describe the actual intentions while the fit ones describe the potential.

Assign weights to signals

Now you should assign the appropriate weights to the indicators. This is the step that requires your evaluation skills and where models usually fail by treating the signals equally.

The weights of the signals should reflect the importance of each one of them: how much they predict the readiness of the customer. In other words, the weight is the probability of conversion.

Concretely, it usually means that behavioral signals of the deep engagement (onboard completion, usage limits) receive the most points, fit signals receive moderate number of points, and weak signals like email open receive the minimum of points. What matters here is not the exact value but the hierarchy of the weights: the most valuable signals should get the maximum of points while the least valuable should get the minimum.

One more warning: there are no universal point values. The right weights are the ones that make your model work: the leads with high score should convert in higher proportion compared to those who received the low score. To evaluate it you can start from the best guess but then adjust the model depending on the results.

Name the weak signals and resist them

You should name the indicators that seem to predict the readiness of a lead but actually do not, as they sneak into most of the models and corrupt them.

Such examples are email open and click events. They are easy to track and feel like the evidence of interest but opening emails is close to free action that does not indicate the willingness to purchase. So, weighting them highly, you push your passive readers to the top and hide the leads who are really using the product.

One-time page visits, positions, webinars signups – all of them feel meaningful and are not meaningful enough. The real engagement is in the patterns, not in the isolated events.

The discipline of creating the proper model is as much in rejecting certain signals as in scoring them. Every additional weak signal reduces the value of the strong ones and the final score becomes useless. In case of doubt, leave the signal out.

Set the threshold that you can defend

The score is just a number until you decide what it should trigger. The threshold is the line beyond which the lead becomes “ready” and the decision about it is quite serious.

The too low threshold would flood your sales team with the leads that are not ready at all and would waste their time, as well as teach them to ignore the scores. The too high one will lead you to sitting on the warm leads, waiting till they cool down and lose the people who were ready.

There is no universal formula that helps to identify the right threshold. It depends on the available capacity of your sales team and on how many leads you can afford to lose comparing to the conversations that are wasted. The practical approach is to set the threshold on the level when the leads with score above the line convert in significantly higher proportion than those with lower score. Then you can readjust the threshold according to the performance of flagged leads.

This step is where your model is connected with the stages of the qualification process. The crossing of the threshold means turning of the raw lead into the marketing qualified lead (MQL) or into the sales qualified one (SQL). In case these terms are strange for you, the MQL vs SQL article explains the stages.

Make the model small

There is a temptation to add more and more signals to the model. Do not let yourself do this. The small model is always better.

A model with several strong signals is easier to explain to the sales team and thus, they are more likely to trust it and use the scores. It is easier to inspect: once the bad leads appear, you can see why it happened. As it relies on the most valuable signals, the small model works better than the complex one and is less fragile and opaque.

The large complex model becomes the black box: high scores appear and nobody knows why, bad leads pass, and nobody knows the reason. The sales team returns back to the gut feeling which was the initial problem. You created the scoring model exactly to avoid it.

Fewer signals with the proper weights are the desirable state. Not the most sophisticated model, the one that your team will trust and use.

Wire it to action, then refine

A model that produces scores but makes nothing with them is just a decoration. The value of the scoring appears when the score triggers an action.

This means that the scoring process should be connected to the lead routing: once the lead crosses the threshold, the right person or workflow should immediately receive the lead and its context so that the follow-up would be timely and relevant. Also, you need to use the score for determining whom to route to the sales team yet: the leads with low score on the behavioral signals but fitting the profile are not failures, they should go through the lead nurturing process till they start engaging more actively and increase their score.

Then the refinement step should happen: products change, customers change, and the model that worked well in previous quarter does not work anymore. The periodic checking of the conversion rates of your model leads is necessary and when they do not convert, this means the information for you: some signal is overweighted, some threshold is wrong, weak signal slipped in. This helps you to tune the model constantly.

Scoring model is not the thing you create once and forget, but the thing you create and maintain.

Build it, or have it built

Building the model, the judgment about which signals predict your customers, is work only you can do, because only you know your customers. But operating it, the continuous scoring, the threshold-based routing, the handoff with context, the ongoing tuning, is repetitive and time-sensitive, which is why it is usually built as an automated system rather than maintained by hand.

That is the split worth understanding before you start. The strategy is yours. The tireless execution of it, running the model continuously so no ready lead sits unnoticed, is exactly the kind of thing teams either build once and maintain, or hand to a done-for-you lead scoring setup so it runs without someone babysitting it. Either way, and more broadly as part of marketing automation for SaaS, the point is that a model only earns its keep when it runs reliably, not just when it exists on paper.

Frequently asked questions