Looking for a job
The message asks about a role, applying, or joining. Phrases like "I need a job" or "looking for a role" are direct and easy to match.
Case study · Lead quality
Filtering leads without throwing away buyers
Paid search brought in job seekers and suppliers alongside real buyers. I wrote rules to catch them, then tested the rules against six months of past leads before letting them touch anything.
6 monthsof past leads used as the test set
~1 in 5paid search leads were not buyers
0real buyers blocked in the backtest
3rules, all readable in one screen
Problem
Sales stopped trusting the paid search leads.
Roughly one in five form fills from paid search was someone looking for work, someone selling a service, or a message too short to mean anything. Sales worked through them anyway, found nothing, and started treating the whole source as junk. That is the real cost. Good leads from a distrusted source get slow replies.
The three rules
Selling, not buying
The message offers a team, a service, or data. These people are on the supply side. They are real, but they are not pipeline.
Too short to read
A three-character message carries no intent. It is not spam exactly, it is just nothing.
Method
Test the rules on leads you already know the answer to.
Writing filter rules is easy. Trusting them is not. So the rules ran against six months of past paid search leads, where I already knew which ones turned into real conversations.
The rules disqualified about one in five, and every buyer that had gone on to a real deal passed through untouched. Zero false positives on that set. That is the number that got it switched on.
One deliberate choice: the rules read the message only. Not the email domain. Filtering on free email addresses is tempting and it is wrong, because founders and buyers at small companies use personal addresses. Rules that read what a person wrote are harder to write and safer to run.
Running it
A filter you cannot see is a filter you cannot trust.
The gate runs on a schedule. It writes to a log every time. A dashboard shows a green dot when it ran in the last twenty minutes and a red one when it has gone quiet, next to how many leads it disqualified.
Anything that silently removes leads needs that. Otherwise the first time you find out it broke is a quarter later, when the numbers do not add up.
Lead names, company names, deal values, and the source data are withheld. Proportions are rounded.