Don't shrink the market to fit your data

A while back I pointed an AI at a list of small companies to find the ones that had a specific problem. It's the kind of job these tools are genuinely good at, read a lot of public data, look for a pattern, tell me who to call.

The signal came back usable on one company out of eight.

That's a bad hit rate, and the tool knew it. So it did what a helpful system does, it proposed a fix. The data I needed was thin at the company size I was targeting, it was much richer at larger companies, so the recommendation was to raise the size floor. Move the target from companies with a dozen or so employees up toward a hundred, and the numbers get healthy again.

It was a reasonable suggestion. It was supported by the data. It was also exactly backwards. If I hadn't been holding a separate opinion about who the buyer was, I would have taken it and felt smart about it. That's the most dangerous way to feel about a spreadsheet.

The quiet way an ICP drifts

Nobody sits in a meeting and decides to shrink their market to fit their tooling. That decision never gets made out loud, it accumulates.

It goes something like this. You pick a target. You build a way to find them, a list, a scoring model, a signal, a set of filters in whatever database you pay for. The instrument works great on some of the target and poorly on the rest. The part where it works produces leads, meetings, a clean dashboard. The part where it doesn't produces nothing, and nothing is very easy to stop paying attention to.

Six months later your ideal customer profile matches the fields your tools happen to fill in well. The deck says you sell to mid-market operations leaders, and when somebody asks why, the honest answer is that those are the people your enrichment vendor has good data on.

You end up with a cleaner pipeline report and a smaller business, and nobody can point to the day it happened.

I see this constantly, and it almost never looks like a strategy problem from the inside. It looks like focus. Focus is good, everyone agrees focus is good, so narrowing to where the data is strong feels like discipline and works more like a retreat with a nicer dashboard.

Two failures that look identical in a spreadsheet

Here's the distinction the whole thing turns on.

When a targeting effort comes back weak, one of two things happened. Either the measurement failed, or the customer is wrong. In a spreadsheet those look exactly the same, a row with no usable signal. A column of blanks doesn't tell you whether the buyer is absent or the camera is pointed at the wrong wall.

Only one of those is a reason to shrink.

If the customer is wrong, if the people you picked don't have the problem, can't pay for the fix, or don't feel it badly enough to act, then yes, narrow, move, rethink. That's the market telling you something, and you should listen.

If the measurement failed, the market hasn't told you anything yet. You asked it a question in a language it doesn't speak and took the silence as an answer.

In my case the signal I'd built leaned on two kinds of data, and both were sparse at small companies for boring structural reasons. Small companies don't publish much. They don't have the reporting layers that generate the public exhaust bigger companies do. My instrument was built for a kind of company that leaves footprints, and the companies I wanted to reach mostly don't.

That's a fact about the instrument. It says nothing about whether a twenty-person company has the problem, and I had good reasons to believe they did.

You need a thesis the data can't overrule

The reason I could push back at all is that I came in with an independent opinion about the buyer.

The note I wrote to myself at the time was roughly this: the 20 to 100 person company is a good target, one that wants to invest in and keep its people, so let's find a different way to open that door.

That sentence did the work. It held the target still and made the method the variable. Without it, the only thing in the room with an opinion was the data, and data will always vote for whatever it can see.

This is the part I'd push any founder on. Before you build a single list, can you say in plain language why this buyer is right, in terms that have nothing to do with how easy they are to find? Something like "they feel this pain in a specific moment, they can make the decision themselves, and the cost of the problem shows up somewhere they look every month."

If you can, you have something to defend when the instrument disappoints you. If you can't, your tools will write your positioning for you, and they'll write it around their own blind spots.

Change the door, keep the room

So I kept the target and threw out the signal.

I won't walk through the replacement in detail, partly because the specific signal is the kind of thing that stops working once everybody uses it. The shape of it is useful though. The first signal tried to infer the problem from data small companies don't produce. The second one looked for something small companies do produce, a decision they'd already made that told me they were the kind of company that would care.

The first approach needed the data to be rich. The second needed only one thing to be true, and it was true in public.

It returned roughly nine thousand qualified contacts. Same market, same company size the tool wanted me to abandon.

It also did something I didn't expect. It pointed at a different person inside those companies, someone whose actual job is the problem, which changed the story we told them. The old framing put us in a feature category, where we'd lose a side-by-side comparison to tools with bigger budgets. The new one put us in an outcome the buyer already owns and gets judged on.

So fixing the instrument didn't just recover the market, it improved the pitch. That happens more often than you'd think, because a bad instrument is usually a symptom of not understanding the buyer well enough, and the work of building a better one forces you to understand them better.

Why the helpful answer was wrong

I want to be careful here, because the tool did its job well, which is the uncomfortable part. It was being helpful. It saw a bad number, found the variable that most improved it, and recommended changing that variable. That's what a good optimizer does, the same way a good GPS will cheerfully route you through a lake if the lake is technically shorter.

The problem is what it was optimizing. It was solving for the hook, how do we make this specific signal produce more results. I was trying to solve for the business, how do we reach the people who need this.

Those two goals agree most of the time, which is why the drift is so hard to catch. They only split when the instrument has a structural blind spot, and that's exactly when you most need someone to notice.

Every team has a version of this, with or without AI. The lead scoring model that slowly redefines a good lead as whoever behaves like the last batch of good leads. The paid channel that keeps getting budget because its attribution is cleanest, not because its customers are best. The outbound list that keeps getting filtered down to the contacts with verified emails until the remaining list is mostly people who already ignore you, now with excellent deliverability.

In each case the instrument quietly becomes the strategy, and the strategy gets narrower every quarter.

What to actually do about it

When a targeting effort comes back weak, before you move the target, ask three things.

First, is the silence coming from the market or from the measurement? If the people you're targeting simply don't generate the data your signal needs, you've learned about your data, not about them.

Second, what's your reason for this buyer that has nothing to do with how findable they are? Write it down. If it doesn't survive being written down, the data might be right and you should move. If it does survive, defend it.

Third, what does this buyer leave behind in public that tells you they're a fit? Every market leaves some trail. Small companies leave different footprints than big ones, and the job is finding the footprint that exists, rather than demanding the one you already know how to read.

Then rebuild the instrument and run it again.

Sometimes you'll still be wrong about the market, and that's fine, at least you'll be wrong for a real reason. The expensive outcome is the one where you never find out, because the dashboard got cleaner and everyone took that as a yes.

The market was fine the whole time, the flashlight was pointed at the wrong wall, and the helpful suggestion was to move into a smaller room.