Lead scoring
A score is only useful if it changes who you contact
Lead scoring is assigning a number to a prospect that predicts how likely they are to buy, so a limited amount of human attention goes to the people most likely to reward it.
Most scoring models fail for one of two reasons. Either the score is never used to change behaviour, in which case it is decoration, or fit and intent are collapsed into a single number, in which case a perfect customer who is not shopping outranks a mediocre customer who is buying this week. Both are fixable and both are extremely common.
Fit and intent are different questions
Fit asks whether this person could ever be a good customer: their role, their company, their situation. Intent asks whether something is happening right now. They move on different timescales and they need different responses, so collapsing them into one number destroys the information you needed most. Keep two numbers and decide with both.
Weights should come from closed deals, not from a meeting
The usual way a model is built is a room of people arguing about how many points a job title is worth. The better way is to look at who actually bought, find the two or three attributes those accounts shared, and weight those. Everything else is intuition wearing a number.
A score with no threshold is not a decision
The point of a score is to draw a line: above this we reply today, below it we do not reply at all. Without a threshold you have ranked a list you were going to work through anyway, and nothing about your day changed.
Recency decays intent, not fit
A buying signal from six weeks ago is close to worthless, while a good fit is still a good fit. If your model does not decay intent it will keep recommending people whose moment has passed, which is how a scored list slowly becomes a list of stale leads with high numbers.
How to do it
Write the fit definition first
Three to five attributes that describe an account you would be glad to win. Keep it short enough that you could check it by hand on a single profile.
List the intent signals you can actually observe
Only signals you can see. Asking publicly for a recommendation, complaining about an incumbent, announcing a project, hiring for a role. If you cannot observe it, it cannot be in the model.
Score the two separately
Two numbers, not one. Then decide what each quadrant means: high fit plus high intent gets a reply today, high fit plus low intent gets a follow, low fit gets nothing regardless of how loud it is.
Calibrate against thirty real records
Score thirty people you already know the outcome for. If the model ranks a lost deal above a won one, the weights are wrong, and adjusting them now costs an hour rather than a quarter.
Review the threshold monthly
Track how many above the line converted and how many below it you later wished you had contacted. The second number is the one nobody measures and it is the one that tells you the line is too high.
Common mistakes
- One number for fit and intent, which buries the difference between a good customer and a buying customer.
- Weights invented in a meeting rather than derived from accounts that actually closed.
- No threshold, so the score ranks a list without ever removing anyone from it.
- Never decaying intent, so a signal from two months ago keeps somebody at the top of the list.
How Quillen handles lead scoring
Quillen runs an AI ranker over every candidate post it discovers and scores it for buyer intent against the ideal customer profile you describe, which is the fit half and the intent half kept as one pipeline rather than one number you maintain by hand. The ranker is charged separately from draft generation, at a higher AI markup, because scoring every discovered post is the expensive part of discovery rather than a free extra. Scores are visible on each candidate along with why it was surfaced, so a score you disagree with is a score you can inspect instead of one you have to trust.
Common questions
- How many attributes should a scoring model have?
- Fewer than ten in total across both fit and intent. Beyond that nobody can explain why a given lead scored what it did, and a model no one can explain is a model no one overrides correctly.
- Should I score people or companies?
- Both, for different things. Fit is usually a company question and intent is usually a person question, and the person showing intent at a badly fitting company is genuinely worth less than the same signal at a good one.
- What if the score disagrees with a rep?
- Record the disagreement and check it in a month. A model that is overridden constantly is wrong, and a model that is never overridden is not being read. Both are signals about the model, not about the rep.