The hottest leads in the pipeline that could never buy

A college sorority once scored as one of the hottest leads in a pipeline I built.

It was in good company. The engine had flagged 2,077 accounts as its highest-intent tier, the ones that skip the nurture sequence and go straight to someone who closes. When I sat down and actually read the list, it included a couple of urban-farming nonprofits, an industry news publisher, several divisions of large hospital systems, and a trade show that happens once a year and has no one on staff the other eleven months.

None of them were ever going to buy, and every one of them scored hot.

I'd built the thing, so I couldn't blame the vendor, which was inconvenient.

Every score is an ICP somebody forgot to write down

Most founders I work with can tell you who their ideal customer is. It's in a deck somewhere, usually with a persona name and a stock photo. What they can rarely tell you is who their lead score thinks the ideal customer is, and those two are often different people.

A scoring model is a definition of your buyer expressed as math. It decides who gets a call this week, who gets an email next month, and who gets ignored. That is positioning work, whether anyone signed off on it or not. The difference is that the version in the deck was argued over in a meeting, and the version in the scoring model was mostly decided by whatever data happened to be easy to collect.

So when I talk to a team about positioning, I ask to see the scoring rules alongside the messaging. The messaging tells me who they hope to sell to. The scoring tells me who their sales team is actually spending the afternoon with.

Scores read behavior and guess at intent

The engine I built looked for a few things that usually travel with a company that needs help: a website in rough shape, an ad pixel that says someone is spending money on traffic, and recent activity that suggests the timing is live. For a small business, that combination is a decent tell. Someone is paying to send people to a site that isn't converting them, and they're doing it right now.

The trouble is that a sorority with a dated website and a Facebook pixel throws off exactly the same signal.

Scoring models measure what they can see, and the intent to buy isn't something they can see. Budget, a person with authority to sign, a reason to act this quarter, none of that shows up in behavior. So the model reads behavior and infers the rest, and anything that behaves like a buyer gets treated like one. The math did its job perfectly on the sorority, it just answered a question nobody had asked it.

The higher the score, the worse it got

This is the part that made me put my coffee down.

Across the whole hot tier, about 31% of accounts couldn't buy. On the very top slice, the highest-scoring accounts in the system, the false-positive rate was close to half.

Once I looked at who was up there, it made sense. The biggest, most neglected websites on the internet tend to belong to hospital systems, national associations, and event brands. They're large, they spend on ads, and their sites are a museum of every redesign nobody finished. They throw off the loudest "someone should fix this" signal of anyone in the market, and they're also the least likely organizations on earth to hire a small shop to fix it. They have procurement departments, and procurement departments have opinions.

So the score was most confident exactly where it was most wrong. If the team had worked that queue top-down, which is what a score invites you to do, the first few weeks would have gone to the accounts with the worst odds in the pipeline.

That inversion is structural. Any signal that tracks "size of the visible problem" will over-rank organizations that are too big, too slow, or too far outside your lane to buy from you. The fit question and the signal question run in different directions, and the model only knows how to answer one of them.

The fix was a person reading the list

My first instinct was the one everyone has, which is to make the model smarter. Add a filter for nonprofits, a rule for hospital systems, a penalty for anything with "Greek life" in the description. You can do some of that, and some of it helps.

What actually fixed the queue was simpler and a little humbling: a human judgment pass on top of the deterministic score. Read the list, decide who could plausibly buy, and be opinionated about it. Out of 2,077 hot accounts, 1,428 survived as real, workable prospects. The rest moved to nurture, hold, or reject.

The second half of that decision mattered as much as the first. We didn't throw the cut accounts away. Every one of them stayed claimed in the system, because a sorority isn't buying today and a hospital division might reorganize next spring and spin out a team that is. The way I put it at the time was "be opinionated for higher quality. we still call dibs on the junk in case it comes back later."

Cut what you work, keep a claim on what you cut. The working queue gets smaller and sharper, and you still own the option on everything else.

Why this got worse when generation got cheap

Ten years ago, a scored list like that would have been worked by hand, slowly, by someone who would have noticed the sorority by Tuesday. Now you can point AI at the same list and generate personalized outreach for all 2,077 accounts before lunch.

That speed is real and I use it every day. It also means a queue that looks precise and is quietly a third garbage can go out the door before anyone reads it. Every email is well written, every one references something true about the account, and a meaningful share of them land on organizations that were never in your market.

Generating the work costs almost nothing now. The expensive part moved to whoever's willing to read the list and say no to a high score.

How to tell if your score has drifted from your buyer

You don't need to rebuild anything to check this. A few questions usually surface it.

Read your top 50 scored accounts by name. Out loud, if it helps. Count how many you'd be genuinely surprised to close. If the number makes you wince, your score has an opinion about your market that you don't share.

Compare your highest scores to your last ten closed deals. If the accounts the model loves most look nothing like the customers who actually paid you, the score is describing a buyer you've never had.

Check whether conversion rises with score. It should. If your top tier converts no better than the middle, or worse, you probably have the same inversion I had, and the signal is tracking problem size rather than fit.

Look at what you cut. If nothing ever leaves the hot tier except by closing or dying, nobody is applying judgment. If things leave and vanish, you're throwing away options you'll want later.

None of these needs a new tool. They need someone who knows who the business is for to sit down with the list.

The score is a draft, you're the editor

I still believe in scoring. A good score saves a team from reading 10,000 accounts to find the 500 worth a call, and the engine that ranked a sorority near the top also found most of the 1,428 real prospects I kept.

The mistake is treating the score as the final word on who you're for. It's a first draft of your ideal customer, written by a system that can only see behavior, and it reads the internet very literally.

Your positioning decides who you're for. Your scoring decides who you call. If those disagree, the scoring wins by default, because it's the one making decisions at 9 a.m.

So read the list. Somewhere near the top there's a sorority with a Facebook pixel, and it is not going to sign.