When the Match Rate Is Zero, Check the Column

A low match rate is a tuning problem. A match rate of zero is a different problem: the two lists you are comparing were never going to overlap.

Watermelon AI · Field Notes ·

You export a month of deposits from the bank. You export the open invoices from QuickBooks. You ask the software to pair them so you can see which customers paid. It comes back with nothing. Not a few misses. Zero.

The natural move is to assume the matching is too strict. So you loosen it. You ignore capital letters. You strip spaces and punctuation. You allow a few days of slack on the dates. Still zero. Now you start doubting the tool.

Few and zero are different results

When a match finds some pairs and misses others, the method is close. It is picking up real signal and needs adjusting. That is normal work, and you know how it goes.

Zero is not a weaker version of that. A clean zero usually means the two things you are comparing never had anything in common. Loosening the rules can't fix it, because there is nothing there to loosen toward.

The label on the column

In bookkeeping this happens most often with a field that is named one thing and holds another. A bank export has a column called Reference. You assume it carries the invoice number your customer wrote on the payment. It doesn't. It carries the batch number the bank assigned when it bundled the day's card deposits. Or a processor export has a column called Customer that actually holds the name of the account that sent the payout.

Every comparison after that pits invoice numbers against batch numbers. The software is doing exactly what you asked. You asked the wrong question.

The fix is rarely clever. Stop trusting the header and open a few rows. Look at what is actually in the field. Then link the records through something that truly points at the customer: the amount and date on the processor's payout report, or the remittance detail that came with the payment.

What we check first

Before we tune any match, we read five rows from each side and confirm the fields hold what their headers say. It takes two minutes. It is the cheapest test in the whole process, and it saves the most time when it fails.

A field name is a note somebody wrote when the system was set up. It describes what they meant at the time, not what the field holds today. Exports get reconfigured, software gets updated, and nobody goes back to rename the column.

So treat a perfect zero as information. Ordinary bad luck leaves a mess behind: a handful of matches where you expected dozens. A result with not one single pair is too neat for that. That is the data signaling, as plainly as it can, that the column holds something other than its label.

The takeaway: When a match comes back empty, open five rows and read the actual values before you change a single setting.

Want these checks running on your books?

We set this up on your books and your rules, and nothing posts without your approval.

Book a call

More field notes

All field notes →