.png)
When a B2B marketing team wants to find more customers, the obvious place to start is the customers it has already won. What do they have in common? Which characteristics can be used to identify more companies like them?
It sounds sensible. It's also where a lot of marketing analysis goes wrong.
Lookalike audiences, revenue signals, lead scores and AI-generated insights often start with the same assumption: if you can identify what your successful accounts have in common, you can reproduce that success. But a shared characteristic only tells you so much. Unless you know whether it also appears among the accounts you lost, you cannot tell whether it has anything to do with winning.
That is the subject of Episode 13 of B2B Effectiveness: Evidence-Based Marketing Ideas for B2B Practitioners. Dale W. Harrison and Diego Sosa explore what happens when marketers mistake the characteristics of their winners for evidence of what makes them successful – and how that can lead to decisions that are no better than guesswork.
What your winners can't tell you
Consider the classic lookalike audience. You upload a list of closed-won accounts to an advertising platform, and it identifies other companies with similar characteristics. The assumption is that these companies are more likely to buy from you.
But what if your closed-lost accounts look just as similar?
If the two groups share the same characteristics, the platform has no way of distinguishing between them. You could upload your losses instead and get essentially the same audience. The list might accurately describe your customers, but that doesn't make it a useful way to find more of them.
Dale illustrates the problem with a deliberately ridiculous example. Every single closed-won deal has vowels in its name. You could call vowels a perfect revenue signal, with a 100% correlation to success. But every closed-lost deal has vowels in its name, too.
The problem is that the same pattern appears in both groups, so it tells you nothing about why one deal closed and another didn't.
The question isn't what your winners have in common. It's what distinguishes them from your losers.
When the numbers are right but the conclusion is wrong
This becomes particularly important when marketing teams decide where to spend their budgets.
Dale and Diego discuss a report from a marketing attribution vendor that found 17.9% of closed-won accounts had visited the seller's LinkedIn company page. The report treated those visits as an important revenue signal, suggesting that advertising investment should follow.
But 17.9% tells us only what happened among the winners. Without knowing how many closed-lost accounts also visited the page, there's no way to establish whether those visits help distinguish successful deals from unsuccessful ones.
In fact, Dale says that when he has examined the same comparison, the proportion of lost accounts engaging with the company page has often been higher.
As Diego points out, the number itself wasn't wrong. The conclusion drawn from it was. At best, the data adds nothing to your understanding. At worst, it directs budget toward activity that doesn't improve your chances of winning.
The problem with measuring only the wins
This is part of a broader distinction between data and information. Data is what you collect; information is what helps you reduce uncertainty and make a better decision.
A hundred spins of a roulette wheel give you a hundred data points, but they don't tell you what the next spin will produce. Collecting a million spins doesn't fundamentally change that. More data isn't necessarily more information.
The same applies to lead scoring. A model might assign points to job titles, content downloads, company size and other attributes, then produce a score that looks precise. But if those attributes don't help distinguish accounts that convert from those that don't, the score offers little practical value.
Dale cites a test run by the RevOps team at Zoho CRM. Alongside the MQLs generated by its lead-scoring system, the team sent sales a batch of randomly selected contacts. The random leads achieved the same acceptance and close rates as the scored leads.
In other words, the scoring process wasn't producing better results than chance.
That's a useful test for any qualification model. The question isn't whether it can assign scores or identify patterns. It's whether using those scores helps you make better decisions than you would without them.
Count what didn't work
The importance of looking at losses becomes clearer with a simple example.
Imagine two slot machines. Both cost a dollar per pull and pay $100 when they win. One pays out once every ten pulls; the other, once every thousand.
If you record only the wins, both machines look identical. Every recorded payout is $100. You can distinguish between them only by counting the losses as well.
Marketing faces the same problem when it measures success in isolation. Knowing the characteristics of your closed-won accounts tells you what success looks like, but not how likely those characteristics are to produce it.
The same principle applies to the metrics used to evaluate sales performance. Comparing the sales cycle of closed-lost deals with the overall average is less revealing than comparing closed-won and closed-lost deals directly, because the overall average already includes the losses.
It's also worth asking what a dashboard metric leaves out. Average revenue per deal, for instance, is generally calculated using deals that closed. It doesn't account for the opportunities that consumed sales time but generated no revenue. Those losses matter when you're trying to understand the return on sales effort.
Without them, the picture is incomplete.
Why the picture can be so flattering
If the limitations are this straightforward, why are they so often overlooked?
Dale points to the incentives built into marketing technology. A dashboard that attributes revenue to a campaign makes a stronger case for continued investment than one that shows the same activity among lost opportunities. There is an obvious commercial appeal to presenting the most favorable interpretation of the data.
LinkedIn-influenced pipeline is one example. Depending on how attribution is defined, a closed-won account might qualify if someone, at any point in the previous six months, interacted with something loosely related to the seller on LinkedIn. That could include a colleague commenting on an unrelated post that appeared in their feed.
The attribution may be technically valid under the vendor's definition. But without applying the same criteria to closed-lost accounts, it doesn't establish that the activity contributed to the sale.
If engagement is just as common – or more common – among lost accounts, the attribution tells you little about what drives revenue.
The consequences extend beyond wasted budget. As Diego warns, presenting misleading metrics to a CFO or board can damage confidence in the numbers themselves. Once decision-makers question how a figure was produced, even sound analysis can become harder to trust.
What useful analysis looks like
The alternative isn't to abandon data or stop looking for patterns. It's to test those patterns against the outcomes that matter.
Dale comes back to two questions:
- Compared to what? A figure means little without a relevant baseline.
- What distinguishes winners from losers? A characteristic shared by both groups isn't enough to explain why one succeeds and the other doesn't.
This becomes especially important when using AI to analyze CRM data. AI can identify common characteristics and combinations of attributes, but commonality alone doesn't establish value. Every company in your pipeline, for example, employs people. That fact won't help you identify which accounts are most likely to buy.
More useful findings may be less obvious: characteristics that appear in some closed-won accounts but are much rarer among closed-lost ones, particularly when several differences occur together. The goal is to identify patterns that genuinely separate outcomes, not simply produce a convincing description of your existing customers.
The same caution applies to engagement scoring. One content download may tell you very little. Ten downloads might indicate sustained interest, but a hundred won't necessarily make someone ten times more likely to buy. Dale illustrates this with a change in certainty from 92% to 94%: a substantial increase in activity can produce only a marginal change in what you actually know.
If a score keeps rising as someone consumes more content, that doesn't automatically mean their likelihood of buying is increasing at the same rate. The score needs to reflect evidence of a meaningful difference in buying likelihood, not just the volume of activity.
Otherwise, it's like the person who notices that ice cream sales and drownings rise together and concludes that opening more ice cream stands will somehow solve the problem.
Start with the question, not the data
Being data-driven isn't the end goal. Collecting more data, building more elaborate models and tracking more engagement can all create the appearance of progress without improving a single decision.
The real test is whether your analysis reduces uncertainty about what to do next. That means looking beyond what your successful accounts have in common and examining what separates them from the ones you didn't win.
Your losses aren't just deals that failed to close. They're the comparison that gives your wins meaning.
