Lead scoring for B2B SaaS

Here is the test for whether your lead scoring model is real. Pull last quarter’s closed won deals, look up the score each lead had when it was passed to sales, and plot win rate by score band. If the line is flat, you have been routing leads at random with extra steps for however long the model has been live.

Scoring is a prediction problem. Build it like one.

Why workshop derived scoring models fail

Because the point values come from opinions and opinions are not evidence. Somebody says a demo request is worth 25 points, somebody else argues webinar attendance deserves 15, and the number that ships is the average of the room’s confidence rather than anything about your customers.

The failure is not that the weights are wrong. It is that nobody can tell whether they are wrong, so the model never improves. Sales tries it for six weeks, hits three enthusiastic 90 point leads that go nowhere, and reverts to working the list by company logo. After that the score field is furniture.

The two axis model, and why one number is not enough

Fit and behaviour answer different questions. Fit asks whether this account could ever be a good customer. Behaviour asks whether they are in market right now. A single composite score lets a highly engaged bad fit score the same as a quiet perfect fit, and those two leads need completely different treatment.

Score each on a 0 to 100 scale, then route by quadrant.

Quadrant	Meaning	Routing rule	Realistic expectation
High fit, high behaviour	In market and worth winning	Route to AE within 15 minutes, Chili Piper style instant booking	Highest win rate band, should be 3x baseline
High fit, low behaviour	Right account, wrong time	Nurture plus outbound sequence, no SDR call blitz	Converts on a 2 to 6 month lag, do not judge weekly
Low fit, high behaviour	Enthusiastic and unbuyable	Self serve path only, no human time	Occasionally becomes a champion at their next company
Low fit, low behaviour	Noise	Newsletter, nothing else	Do not spend SDR minutes here at all

The low fit, high behaviour quadrant is where most single threshold models leak money. A student researching a dissertation can trip every behavioural trigger you own. Under a composite score they become a 78 and land in an SDR queue. Under a two axis model they land in a self serve nurture and cost nothing.

Derive weights from closed won history, with a worked example

Export 12 months of opportunities, won and lost, with the lead attributes present at the time of creation. You need the lost ones. A model built only on wins tells you what your customers look like, not what distinguishes buyers from non buyers, and those are different questions.

Compute win rate per attribute value against your baseline. Here is a real shaped example from a mid market workflow product with a 9.2% baseline win rate on qualified opportunities.

Two things usually surprise people in this table. The ebook and the webinar, which the team had been counting as strong MQL signals, carry essentially no predictive power. And enrichment data about tech stack outperforms most self reported form fields, because people lie on forms and Clearbit does not.

Scale the weights so each axis tops out near 100. Do not add attributes that show lift below about 1.2x; they add noise and make the model harder to explain.

Negative scoring and decay, the two things most models lack

Negative scoring removes the structurally unbuyable. Free email domains where your product sells to companies, competitor domains, job seeker and student titles, countries outside your sales coverage, and careers page visits, which correlate strongly with people looking for work rather than software.

In most B2B SaaS databases somewhere between a fifth and a third of inbound leads fall into this group. Removing them from the queue is the single largest immediate improvement to lead quality available, and it requires no modelling at all.

Decay handles the other half of the problem. Behaviour is a claim about now, so it has to expire. A pricing page visit from March tells you nothing in September.

Validate by cohort before sales sees a single score

This is the step that separates a model from a decoration, and it takes an afternoon.

Hold out the most recent quarter. Score those leads with your new weights as if the model had been live, then plot actual win rate by score decile. You are looking for a monotonic curve where each decile converts better than the one below it, and a top quartile that closes at two to three times the bottom quartile.

If the curve is flat, the attributes you chose are not predictive and adding more will not save it. That usually means one of three things: your ICP is too broad to differentiate, your form data is too thin, or the real signal lives in product usage rather than marketing behaviour. The third case is common in product led companies, and PQL scoring and routing is the right model for those teams rather than this one.

If the curve is steep but only in the top decile, you have a threshold model rather than a scoring model. That is fine, and simpler. Route the top decile and stop pretending the other nine bands mean anything.

Wire it into routing and the SLA

A score with no routing consequence changes nothing. Each quadrant needs a named destination, a response time and a rejection path, and all three belong in writing with sales. The sales and marketing SLA template covers the contract language; the part specific to scoring is the rejection loop.

Reps must be able to reject a routed lead with a reason code, and those reason codes feed the next recalibration. When 40% of high fit high behaviour leads come back marked “no budget”, you have learned something about a segment that no amount of attribute weighting would have surfaced.

Volume planning changes too, because a validated model means fewer MQLs of higher quality, and the lead goal arithmetic has to be redone. Run the new conversion rates through the lead goal calculator before you commit to a number, and check the economics against the cost per lead calculator and lead value calculator, since a better model usually raises cost per MQL while lowering cost per opportunity.

Recalibration and the signs of drift

Quarterly, one hour. Plot win rate by score decile for the quarter that just closed and compare the curve to last quarter’s.

Four signals that the model has drifted: the curve flattens, the top decile’s win rate falls while the baseline holds, rep rejection rate on high scores climbs above roughly 20%, or a single attribute starts appearing in most high scores because a campaign flooded one behaviour. That last one is the sneakiest. Run a webinar to 2,000 people and suddenly a third of your database looks engaged.

Rederive weights fully once a year, or immediately after a pricing change, a new product line, or an ICP shift. Those events invalidate historical win rates by attribute, and a model trained on the old ICP will confidently misroute the new one.

The honest tradeoff

Scoring costs maintenance forever and buys you prioritisation, not more leads. A validated model with quarterly recalibration takes perhaps 15 hours a quarter of somebody’s time, and its value shows up as SDR hours redirected rather than as pipeline created. If your team is three people and everyone works every lead anyway, build the negative scoring rules and skip the rest until volume justifies it.

Where it pays is at the point where inbound exceeds capacity. Then the question stops being how many leads and starts being which ones, which is the same question lead quality for B2B SaaS works through from the demand side, and where the broader SaaS lead generation hub picks up.

Start with the backtest on your existing model. If the curve is flat, you have your answer, and the lead scoring template gives you the structure to rebuild it properly.