Lead scoring: how to build a B2B lead scoring model that works

A lead score answers two questions that move at different speeds. Is this the kind of company and person who buys from us? And are they showing interest right now? The first is a fit score. The second is an engagement score, and when the activity happens off your own site (a funding round, a new VP, a job post) it is usually called intent or a buying signal.

Keep them as two fields. A combined view is fine for reporting, and HubSpot shows one as a grid (more on that below), but the two numbers should never be summed into one. A student who downloads five ebooks would otherwise outrank a VP at a perfect-fit account who visited the pricing page once.

Here is what a fit scorecard looks like for a hypothetical B2B software company selling to finance teams at mid-sized companies. Your attributes and points will differ; the shape is what to copy. Each row has a reason, and the reason should come from your deal data.

Cap the fit score at 100 and keep the engagement scorecard separate: a demo request might be worth 40, a pricing page visit 15, an email click 3. Low-value actions like email opens are best left out entirely, since privacy features and security scanners trigger opens nobody made.

A score is only useful once it triggers something. Set two or three bands and decide what each one does:

Pick the threshold from your backtest: the score above which leads converted to meetings at a rate worth a rep's time. Then agree it with sales in writing. An MQL definition marketing set alone is the fastest way to get scores ignored.

Score decay keeps engagement honest. Halve or remove engagement points after 30 to 90 days so that a lead who went quiet drops back to nurture on its own. Fit points should not decay, though they should be recalculated when someone changes job or a company crosses a size band.

The handoff itself is a routing problem: who owns the account, round robin or territory, and how fast the first touch goes out. That part lives in your CRM workflows and a router; see the comparison of lead routing tools and how we wire it in HubSpot automation.

HubSpot replaced its old score properties with a dedicated lead scoring tool. Per HubSpot's knowledge base, you could no longer create the old properties from May 1, 2025, and existing ones stopped updating on August 31, 2025. If a workflow or list still reads an old score, it is reading a frozen number. The details below are from HubSpot's lead scoring documentation (updated September 2, 2026) and its note on the old score properties, checked November 8, 2026.

The A1 to C3 grid is the useful part, because it keeps fit and engagement apart while still giving reps one label to sort by. A practical HubSpot setup: build the fit score from enriched company and contact properties, build an engagement score with decay on, use the combined label in rep views, and trigger a workflow when a lead reaches A1 or A2. If your best fit signal comes from outside HubSpot (an AI classification, a product usage count, a funding date), write it to a custom property first and score that property.

Rules break on messy data. Job titles are the obvious case: "Head of Money", "Finance & Ops Lead" and "Controller (Interim)" all need to land in the same bucket, and a list of keywords will miss half of them. Industry is the other. A company's LinkedIn industry often says "Software Development" when what matters to you is whether it sells to hospitals.

A short prompt handles both well. Give the model the title, or the company's website text, and ask it to pick one label from a fixed list, with a one-line reason. Then score the label in your scorecard like any other field. We built a title-fit AI prompt for Patlytics, and SuperDial's TAM was scored in Clay with AI industry classification, alongside HubSpot rep views. The best LLM for GTM page covers which models fit this job and what they cost.

Where AI scoring goes wrong

Predictive scoring trains a model on past conversions and lets it set the weights. It can find patterns a person would miss, and it saves the manual work of maintaining points. The catch is volume. A model needs a few hundred conversions to learn anything stable, and early-stage B2B teams with 30 deals a year don't have them. For most teams under that line, a hand-built scorecard from closed-won data, checked every quarter, will beat a predictive model and is easier to explain to sales.

The most common failure is that everyone scores the same. Half the database sits between 40 and 50, so the score cannot tell a rep who to call. Run two checks before launch and again every quarter.

Distribution. Plot how many leads land in each score band. You want a long middle and a small, clear top tier. If most records share a score, either the attributes don't vary across your data (everyone is in the US, everyone gets the same industry points) or blank fields are scoring zero across the board. For one vertical SaaS client with a TAM of about 45,000 companies, a rebuild moved the median score from 0.25 to 0.50 and left 2,001 accounts above 0.75, which gave reps a clear first list. The GTM engineering guide walks through that rebuild.

Conversion by band. For each band, compare the rate of meetings booked and of closed-won. The rates should climb with the score. If the top band converts no better than the middle, the weights are rewarding the wrong things, and the fix is usually to go back to the won and lost comparison with fresher deals.

Watch one more thing: whether reps use it. If they sort by something else, ask why. They are often right about an attribute the model is missing.

We start from the client's own closed-won and closed-lost deals, enrich the fields the scorecard needs, and write fit and signal scores to properties in HubSpot or Attio so routing, rep views and reporting can use them. Where a field needs judgement, like title fit or industry, we classify it with a tested prompt. Then we check the distribution and conversion by band before reps see it, and again after the first quarter. Scoring is usually one part of a wider build that includes signals, outbound and routing; the RevOps guide covers where it sits. Not sure where your gaps are? The free GTM scorecard grades your whole setup in twelve questions.