Lead scoring, and when to skip it
Lead scoring assigns point values to a lead's attributes and behaviors, adds them up, and calls anything over a threshold an MQL. It works at a company with enough conversion history to fit those points on. At seed you have none, which changes what to build. Here is the anatomy, the precondition every explainer skips, and what to run instead.
By Kshitij Maheshwari, co-founder · Updated August 2026 · 11 min read
The four parts of a lead score
Every points model in every CRM is made of these four pieces. The last column is the one the explainers leave out: what each piece quietly assumes before it can mean anything.
| Part | What it is | Who sets the number | What it silently assumes |
|---|---|---|---|
| Attributes | Who they are: title, headcount, industry, country, stack. Fit. | You, from your ICP | That you already know which facts separate a buyer from a stranger. |
| Behaviors | What they did to you: visits, opens, clicks, form fills. Interest. | Your tool, from what it can track | That they have done something. On a cold list this half reads zero. |
| Points | A fixed number added or subtracted per attribute and per behavior. | An admin, by hand, in a settings screen | That someone measured how much lift each attribute is actually worth. |
| Threshold | The line above which a record becomes an MQL and goes to sales. | An admin, by hand, again | That the line is a probability. Usually it is a capacity decision. |
Sources: the parts, the settings screen and the worked point values are HubSpot's own, from its lead scoring documentation (last updated 14 August 2026); the scorecard definition is from Brendan Duncan's 2016 patent filing of the Duncan and Elkan research. The last column is ours.
Where the MQL came from, and where it went
Your CRM asks for a score because of a diagram published in 2006.
An MQL is a lead whose score has crossed the threshold an admin set, at which point marketing hands the record to sales. It is a stage definition, not a measurement of intent.
Forrester dates the SiriusDecisions Demand Waterfall to 2006, and says countless B2B organizations adopted it.
Forrester now owns SiriusDecisions, and its current guide tells readers to move beyond MQLs toward connected buying groups. The stage is twenty years old and its authors have moved past it.
What a scoring model is actually doing
Under the settings screen, a points model is a hand-fitted classifier. It reads two kinds of evidence about a lead and weighs one against the other.
- •Title, headcount, industry, country, stack
- •Knowable before anyone has replied to anything
- •Stable for months at a time
- •The half that works on a cold list
- •Opens, clicks, page views, form fills, meetings
- •Exists only after they have engaged you
- •The half a CRM tracks best
- •Reads zero on every outbound row
The research that replaced hand-tuned scoring leaned the opposite way from the startup advice. Duncan and Elkan's 2015 models weight fit features over behavioral ones, so good leads surface earlier, before anyone has clicked anything.
A point value is a claim that leads with that attribute convert better than average. Somebody has to have measured that.
Why it breaks before you have conversion history
There is an honest way to set a point value, and it is four steps long. Walk it at seed and you run out of data at step three.
-
1
Work out your baseline conversion rate
Every lead you have ever created, and the share of them that became a customer. One number, for the whole business.
-
2
Pick the attributes you think matter
Industry, headcount, title, country, whatever you can actually read on a record before you send anything.
-
3
Measure the close rate for each attribute
The share of leads carrying that attribute that converted. This is the step that needs recorded losses as well as recorded wins.
-
4
Award points where the attribute beats the baseline
The size of the gap is the size of the weight. No gap, no points. That is the whole method.
At three or four customers, step three has nothing to divide. There are no recorded losses, because nobody logs the accounts that never replied, and every close rate lands on 0/1 or 1/1.
CRM onboarding puts lead scoring in the setup checklist, right beside importing your contacts. So the model gets built on day three, with zero outcomes in it, and its real first job is making the CRM look configured.
What predictive scoring actually requires
HubSpot publishes a minimum for one of its two scoring products and none at all for the other. The number it does publish already sits above what a seed team has.
Fifty contacts, half of them converted
HubSpot's lead scoring documentation, last updated 14 August 2026: "A minimum sample size of 50 contacts, containing 25 converted and 25 non-converted, is required to generate a score."
No stated minimum at all
HubSpot's predictive scoring page, last updated 11 January 2026, lists every input and names no data floor. It is Enterprise only, and HubSpot calls it "blackbox machine learning".
Thousands of labeled outcomes
Duncan and Elkan trained their model on labeled leads from two companies, real wins and real losses recorded for each, before it beat hand-tuned scoring.
The one floor HubSpot does publish asks for 25 converted and 25 non-converted contacts. Count yours before you build anything.
Want your signals scored, and re-scored on what actually converts?
Book a Fit CheckA score is a number about a moment that has passed
The products concede this in their own settings. HubSpot ships a decay control that makes old points count for less, which is an admission about what a score is.
"This account is an 82. Work the 82s first."
- ✕An accumulation with no date attached to it
- ✕Nothing in it says what changed this week
- ✕You cannot write an opener from 82
"Eighty-two points of history, earned on dates nobody stored next to the number."
- ✓HubSpot ships decay at 1, 3, 6 or 12 months
- ✓Decay is the product agreeing that points age
- ✓Date the evidence, do not tune the sum
A lead score is a stored number about a person's accumulated history, hand-fitted once. A signal score ranks live dated events on an account and gets re-fitted from what actually converted.
Same word, two different objects. Which events can carry a first line on their own is the two kinds of signals, in our signal-based selling guide.
The model nobody re-fits
A one-off score is a reasonable thing to build once. It is an unreasonable thing to leave alone, because the market it was fitted to keeps moving underneath it.
What changes underneath it
The segment that started closing this quarter was not in last year's data. A price change, a new competitor, a category that got crowded: none of it reaches a weight that was set once and saved.
Who re-fits it, and how often
Re-fitting means recomputing those per-attribute close rates on new outcomes and moving the weights. That is a standing job at a company with a revenue operations function, and nobody's job at a company without one.
What it is actually ranking
Last quarter's business. A score nobody revisits keeps ranking by what used to convert, so today's leads get ranked wrong, or missed entirely.
Scoring is not the problem. A score nobody re-fits is. We run the signal map and the campaigns as one job, so what actually converted feeds back into the ranking, which is the only reason a score stays worth reading six months in.
Grouping signals, scoring them, and re-ranking on what converted is the whole job of our signal mapping add-on.
What to run instead at seed
Three objects, none of them a points model: a gate that answers in or out, an ordering of live events, and one short list you rebuild every week.
Rank on a stored score
Account 412: 62 points. VP Engineering, 120 staff, opened two emails. Over 60, so call them.
- ✕Nothing in the row is dated
- ✕The points came from a hunch
- ✕No first line lives inside 62
Gate, then order
Account 412: in the ICP. Their VP Engineering started three weeks ago. Top of this week's list.
- ✓In or out, with a re-check date
- ✓The event carries its own date
- ✓The first line writes itself
Both records above are illustrative examples, not client data.
The gate is your ICP written down as conditions, which is exactly what the ideal customer profile guide builds.
Where scoring quietly goes wrong
Three of these are written down in the patent filing behind the research that replaced hand-tuned scoring. The fourth is HubSpot's own documentation, and the fifth is the one that costs you a reply.
Five points a webinar means twenty webinars scores four times as high as five. Duncan's 2016 filing notes the best leads attend two to four, and that heavy attendance can mean a student.
If your team has worked Florida hardest, a model trained on that history learns Florida is a good signal. The same filing warns that machine learning relearns the scorecard, bias included.
Traditional scores are unbounded positive or negative values that do not map to a probability of conversion. An 82 is not an 82% chance of anything. It is 82 of your own points.
HubSpot's predictive tiers hold 25% of your contacts each, so the top quarter of a bad list still reads Very High, and the ranges shift as the list changes.
Acting on a high score as if it were news
A high score is accumulated history and it can be months old. Opening on it produces a note with no date in it, which reads to the buyer exactly like a scraped list, and that is the reply you do not get back.
When scoring genuinely earns its place
Scoring is a compression of history, and it pays as soon as there is enough history to compress. Four conditions, all of them true at the same time.
Before you build one
4 checks
-
More inbound than a person can read
If you can read every new lead yourself on a Monday morning, ranking them is ceremony.
-
Wins and losses both recorded
The accounts that said no, logged with a reason, are half of the training set.
-
Somebody owns re-fitting the model
A named person, on a schedule, moving the weights when the outcomes move.
-
The tier that carries the feature
HubSpot's manual scoring is Professional and above; its predictive scoring is Enterprise only.
That is a company two stages later than the one reading this. Build the gate and the list now, and the score becomes worth building the quarter those four go true.
- 1 A point is a claim about lift that somebody had to measure.
- 2 With no recorded losses there is nothing to fit the weights on.
- 3 HubSpot's published floor is 50 contacts, half of them converted.
- 4 Gate on fit, order on live events, rebuild the list weekly.
Questions founders ask
What is lead scoring?
Should a seed-stage startup set up lead scoring?
How much data does predictive lead scoring need?
What is the difference between a lead score and a buying signal?
Does a lead score go stale?
Is an MQL the same thing as a lead score?
Co-founder of Real Good GTM. He has been the first business hire and Chief of Staff at seed-stage B2B startups, building outbound pipeline before any playbook existed. This post comes from being handed a CRM setup checklist with lead scoring on it, three customers in, and having to work out what those points could honestly be worth.
Connect on LinkedInThe layer that replaces the score
The worldview behind ranking live events, the done-for-you version, and the one signal that arrives with a score already attached.
Signal-based selling
The whole motion: which events are worth acting on, which are only context, and how fast each one ages.
Read the guideSignal mapping
The few signals that predict a buyer for you, grouped, scored, and sharpened by what actually converted.
See the add-onThe intent data signal
The one signal that arrives pre-scored by a vendor, and the honest read on how much that score is worth.
Read the signalWant the ranking done on live events instead?
Book a fit check. We'll look at which signals your market actually emits, how we would rank them for your ICP, and tell you straight if outbound is not the right motion for you yet.
Book a Fit CheckNo hard sell. No fake numbers. Real good work speaks for itself.