.png)
Take every account your company has ever closed-won. Upload the list to an ad platform and ask it to find more companies that look like them.
There's an obvious problem with that approach if you think about it literally. Every company on your closed-won list has vowels in its name. So, by the logic of a model looking for patterns in your winners, vowels are a characteristic associated with success. You could, in theory, use that information to find more companies with vowels in their names.
Obviously, that would be ridiculous. But it illustrates a real problem with the way B2B lookalike modeling and target-account selection often work. Finding characteristics that your successful customers share doesn't tell you whether those characteristics have anything to do with why they became customers.
That's where Episode 6 of B2B Effectiveness: Evidence-Based Marketing Ideas for B2B Practitioners picks up. In this solo episode, Dale W. Harrison, Chief Strategy Officer at The Insight Collective, takes the ideas developed throughout the series and turns them into a practical framework for what a lead score should actually measure: the Next-Gen MQL.
The MQL isn't broken. The scoring underneath it is.
Dale starts with the argument that has run through the series from the beginning: the MQL serves a necessary function. Sales is a one-to-one activity, so a business needs some way of deciding which prospects are worth a salesperson's time. Marketing can reach hundreds of thousands of people with a campaign, but a sales rep can only work through a limited number of conversations. The handoff between the two functions therefore needs some form of qualification.
The problem is what happens when that qualification is treated as a prediction it can't reliably make.
Long-term B2B benchmark data cited throughout the episode suggests that only around 5-6% of MQLs progress to a sales-qualified opportunity, with the proportion that ultimately becomes closed-won smaller still. If the overwhelming majority of leads being passed to sales never become meaningful opportunities, the issue isn't simply that sales needs to work harder. The qualification process isn't doing a particularly good job of identifying where sales effort is likely to pay off.
The problem with finding patterns in your winners
This is where the vowels example becomes useful.
A typical lookalike model starts with a list of accounts that became customers and looks for characteristics they have in common. Those characteristics might include industry, company size, geography, technology usage or thousands of other data points.
But a characteristic shared by your winners is only useful if it helps distinguish them from your losses.
Vowels in a company name is a pattern. So is a huge share of whatever engagement or firmographic data any lookalike tool surfaces. The only patterns worth anything are the ones that show up in closed-won deals and don't show up in closed-lost ones. Everything else is decoration.
This is also the problem with conventional points-based scoring. A downloaded eBook might be assigned ten points and a pricing-page visit two, but those numbers are often based on assumptions rather than a demonstrated relationship between the behavior and the commercial outcome. Changing the points later can make the model look different without making it more predictive.
The important question isn't whether a pattern exists. It's whether the pattern separates outcomes.
Data is not information
The distinction becomes even more important as the amount of available data increases.
Dale compares it to a blood test. A doctor doesn't need to take every drop of blood in your body to diagnose a condition. A small sample contains enough information to perform the test. Taking ten times as much blood wouldn't make the diagnosis ten times more accurate.
The same principle applies to lead scoring. Ten additional engagement signals aren't automatically ten additional pieces of useful information. A model can collect an enormous amount of behavioral data and still have little insight into the thing marketers most want to know: whether someone is actually going to buy.
That is particularly important with AI-driven scoring, which can process vastly more signals than a traditional scoring model. More processing power and more inputs can make a model more sophisticated, but they don't solve the underlying problem if the signals themselves don't reliably distinguish buyers from non-buyers.
What a lead score can actually predict
The alternative Dale proposes is to focus on factors that can be estimated from historical evidence rather than trying to measure an internal state that marketers cannot directly observe.
The Next-Gen MQL brings three of those factors together:
- Category expansion – Is the underlying market growing or shrinking in a particular geography? Market growth changes the proportion of buyers entering the category, which means the likelihood of a prospect being in-market isn't independent of what is happening in the market around them.
- Buyer-seller fit – Is the account likely to buy from a company like yours, and is your sales organization equipped to win that type of business? A company can be an excellent fit for the category while still being a poor fit for a particular seller.
- Firmographic customer lifetime value – If the account becomes a customer, what is it likely to be worth? Two prospects can have similar probabilities of buying but very different expected value because their likely retention, purchasing behavior and account economics differ.
Together, these factors change the question qualification is trying to answer. Instead of asking whether a lead is “ready,” the model estimates how attractive the opportunity is given what can actually be known.
None of that produces certainty. The point is to produce a more useful probability.
Thinking in bets, not certainties
This connects to another idea that has run throughout the series: marketing decisions are made with incomplete information.
You don't know exactly what a buyer is thinking, what conversations are happening inside the account or which competitor they may ultimately choose. There is uncertainty in every opportunity, and that means a good decision can still produce a bad outcome.
The practical response is to think in terms of expected value. If an opportunity has a 20% probability of closing and would be worth $250,000 in lifetime value if it does, its expected value is $50,000. That doesn't mean the company will make $50,000. It means that, given the information available, that is the economic value of the opportunity in probability terms.
That is a much more useful basis for deciding where sales effort should go than an arbitrary 100-point threshold. The score becomes a way of expressing the strength of the opportunity rather than pretending to provide a definitive answer.
A glass box, not a black box
The final piece is transparency.
A useful model should allow you to understand why an account received its score. If the score changes, you should be able to see which factors changed it and how much they contributed. That makes it possible to interrogate the model, test its assumptions and understand where it is working and where it isn't.
It also means the model can work in both directions. The same evidence used to score an individual lead can help define the characteristics of customers you are most likely to win, giving the business a more evidence-based view of its actual ICP. Targeting and qualification become connected parts of the same process rather than separate exercises.
That's the thinking behind the Next-Gen MQL: a qualification model built around probability, commercial value and evidence, with enough transparency to understand how it reaches its conclusions.
