Feature · Lead Scoring Best Practices

Lead scoring best practices for the model your team still trusts in six months.

Lead scoring models fail in predictable ways: too many criteria, static thresholds, no review cadence, no manager override, and a scoring surface the AEs do not trust. The best practices below are the ones Strkr customers running scoring in production have validated. Each one has a failure it prevents and a implementation shape.

Start free trial See pricing

The practices that separate working models

Six choices that decide whether the model survives contact with the team.

A lead scoring model is a set of design choices more than a formula. The choices below are the ones that distinguish a model the team uses from a model the team ignores. the frame: most of the choices are about restraint; the pressure in a scoring project is to add more signals, more weights, more buckets, and the discipline is to subtract.

Keep it under ten criteria

Simplicity in the model beats precision on paper.

Scoring models that run on more than ten criteria lose transparency; nobody can predict the score of a given Lead by looking at it. Keep the model to five to ten criteria chosen for predictive power. The accuracy loss against a 50-criteria model is small; the trust gain is enormous. Reps who understand the score use it.

Weight by actual close rate

Not by intuition, not by vendor template.

The weight on each criterion should come from your actual close rate when the criterion is present versus absent, trailing twelve months. A criterion that doubles the close rate is worth more than one that moves it 10%. The weighting conversation is a data pull, not a workshop opinion.

Three bands, not a spectrum

Hot, warm, cold is enough; eight tiers is too many.

A three-band split (hot, warm, cold) with clear thresholds produces actionable routing. Eight tiers produce confusion: reps cannot distinguish a Tier 3 from a Tier 4. The simplest band structure that drives different actions is the right structure. Add bands only when the routing requires them.

Explain every score

The surface should show why, not just what.

Every Lead's score should display with the criteria contributing to it: "+30 fit, +20 budget, +15 timeline, -10 bid shopping = 55." Scores that appear without the breakdown look magical and get mistrusted. Explanation is a design requirement, not a nice-to-have.

Allow manager override

The system is wrong sometimes; the owner knows.

A manager seeing a Lead the system rated cold but which personal context says is hot needs an override path that logs the reason. The override is not a failure of the system; it is the correction mechanism that keeps the system honest. The audit trail reads the overrides in aggregate for the next model tune.

Review on a schedule

Quarterly tuning is enough; monthly is overkill.

The model gets reviewed quarterly against close-rate data by band and by criterion. If the hot band is closing above 30%, the threshold is too low. If below 10%, too high. If a criterion's predictive power has drifted, re-weight. The quarterly cadence is a 30-minute meeting; shorter cadences invite overfitting.

What to score and what not to score

The criteria that produce signal versus the ones that produce noise.

The pressure in a scoring project is to add every available signal to the model. The practice is to add only the signals that predict close outcomes with meaningful lift. The sections below name common criteria and classify them by whether they predict or decorate.

Score · Fit signals

ICP match, segment, geography, vertical.

Fit signals predict close rates reliably because the sales motion is tuned for a profile. A Lead inside the ICP closes at multiples of a Lead outside. Weight fit heavily; a non-fit should be a near-automatic cold band regardless of other signals.

Score · Intent signals

Form fills, demo requests, pricing-page views.

Explicit intent (a demo request, a pricing-page view, a "contact sales" click) predicts near-term close better than passive engagement. Weight explicit intent heavily. Weight passive engagement (newsletter open, blog view) low or not at all.

Score · Buyer signals

Title, decision authority, team size.

A Lead whose title indicates buying authority (VP, Director, Head of) closes at a higher rate than an individual contributor. Weight title; weight implied authority (team size, budget mentioned) if the data is captured. The buyer signal distinguishes evaluation from exploration.

Do not score · Email opens

A noisy signal that correlates with image loading.

Email open rates include invisible pixel loads from mail-protection services; the signal is contaminated. Remove it from the scoring model or weight it at zero. The exception: a Lead who stops opening after a long run of opens is a disengagement signal.

Do not score · Social presence

LinkedIn connections do not predict close.

A Lead with a strong LinkedIn presence is not more likely to close than one without. Social-presence signals look sophisticated but do not predict outcomes. Weight at zero. The exception: a role or title captured from LinkedIn (which is a fit signal).

Do not score · Session depth

Time on site is a proxy for curiosity, not intent.

Time-on-site and page-view-depth signals correlate with research intent, not buying intent. A competitor researching the product reads many pages carefully. Weight at zero; carry the data for curiosity, not for routing.

Model drift and how to detect it

The signs the scoring model has stopped matching reality.

A scoring model works until it does not. The drift happens quietly: a channel's quality changes, the ICP shifts after a positioning change, a reorg changes who AEs are selling to. The sections below name the detection signals so the team catches the drift before it costs a quarter.

Close rate by band

The hot band should close significantly better than warm.

The core sanity check: trailing-twelve-month close rate by band. Hot should be 2x or more of warm; warm should be 2x or more of cold. If the ratios flatten, the model has stopped discriminating. The dashboard tile that watches this is the first drift signal.

Threshold drift

The hot band is closing at 50% or 5%.

A hot band closing above 30% means the threshold is too low (too few Leads in hot, missing opportunities to classify as hot). A hot band closing below 10% means the threshold is too high (too many Leads in hot, overloading the hot-routing team). Both are drift signals.

Criterion decay

A criterion that stopped predicting.

The quarterly review compares each criterion's lift to the baseline. A criterion whose lift dropped by half is a decay signal: the market or the ICP has changed. Re-weight or remove. The usual decay candidate: a channel source that was high-quality last year and has degraded.

Override frequency

A spike in manager overrides is a drift signal.

The weekly override report counts overrides per rule. A rule or criterion generating overrides at twice the baseline is drifting. The reason notes on the overrides point to the pattern; sales ops uses them to adjust the model.

Rep trust

A rep who ignores the score is a drift signal.

A rep working the pipeline without filtering by score is signaling they do not trust the score. Managers probing the reason get the pattern: a rep working their book knows the hot-rated leads are shallow. The rep's instinct is the drift signal; the model has lagged the ground truth.

Routing distribution

A sudden volume shift in a band.

A band's weekly volume shifts significantly from its baseline (a 2x hot volume, a half warm volume). Something changed upstream: a campaign went live, a channel broke, an ICP shift propagated. Investigate the upstream before adjusting the model.

Implementation checklist

The steps for shipping a scoring model that works.

The implementation checklist below is a walkable sequence. Each step produces a artifact and feeds the next. Teams that skip steps end up with the common failures (over-criteria, static thresholds, no review cadence). The checklist is also a readiness assessment for scoring-skeptical executives who want to understand what is being built.

Step 1

Baseline the close rate without scoring.

Pull the trailing-twelve-month close rate without any filter. The baseline is the number the scoring model has to beat. Teams skipping this step end up tuning a model they cannot evaluate; teams running this step have the comparison baked in.

Step 2

Pick five to ten candidate criteria.

Pick the criteria for predictive power, not completeness. Fit signals (ICP, segment), intent signals (demo request, pricing view), and buyer signals (title, team size) are the usual starting set. Hold the list to under ten; add criteria only when the quarterly review justifies.

Step 3

Measure each criterion's lift against baseline.

For each candidate criterion, pull the close rate for Leads with the criterion present versus absent. A criterion with 2x lift is worth including; one with 1.1x lift is noise. The weighting conversation is a data pull from this step, not a workshop opinion.

Step 4

Assign weights proportional to lift.

Weight each criterion in rough proportion to its lift over baseline. A criterion with 3x lift gets a weight three times a criterion with 1.5x lift. The precise weight does not need to be perfect; the ordering has to be right.

Step 5

Pick band thresholds against routing capacity.

Pick thresholds that produce hot-band Lead volume matching the capacity of the hot-routing team. If the top-performing rep can handle 20 hot leads a week, the hot threshold should produce roughly 20 per week. Capacity constrains the thresholds; demand does not.

Step 6

Launch, measure, review quarterly.

Launch the model, measure close rate by band weekly, hold a quarterly review to re-weight and re-threshold. The model gets durable the moment the review cadence is on the calendar. Teams that commit to the quarterly review end up with a model that still works in year two.

The best practices only work when the surface supports them.

Strkr ships lead scoring natively with explain-every-score, three-band routing, manager override, quarterly-review reports, and criterion-lift analysis built in. The best practices translate to a configuration surface, not a project.

Start free trial See Lead Scoring Criteria

Common questions

What buyers ask about this feature.

How many criteria should a lead scoring model have?

Five to ten. More than ten and the model loses transparency; nobody can predict a Lead's score by looking at the Lead, which breaks rep trust. Fewer than five and the model loses discrimination; too few criteria means the bands end up based on one or two signals. The sweet spot is six to eight criteria, each with clear predictive power, each explainable in a sentence. Teams running 20-criteria models typically find on review that six of them produce 90% of the discrimination.

How often should the scoring model be reviewed and tuned?

Quarterly. The review is a 30-minute meeting that pulls three reports: close rate by band (should show hot > warm > cold with meaningful spread), criterion-lift analysis (which criteria still have predictive power), and override frequency (which rules the team is routing around). The review adjusts weights, thresholds, and criteria based on the data. Monthly reviews invite overfitting; semi-annual reviews let drift accumulate. Quarterly matches the business rhythm most mid-market teams run on.

What criteria should NOT be in a scoring model?

Noisy signals: email opens (contaminated by image-loading services), time-on-site (correlates with research intent not buying intent), social-media activity (does not predict close), and generic activity counts (opens, clicks). The exception is explicit intent signals (demo request, pricing-page view, contact-sales click) which do predict. the frame: most of the signals vendors push into scoring templates do not predict close reliably; the predictive signals are fit, intent, and buyer-authority.

How should scoring interact with manual overrides from managers?

The override is a feature, not a bug. A manager seeing a Lead the system rated cold but which personal context (friend of a past client, referred by a VIP, inbound from a target account) rates hot needs an override path. The override requires a reason that writes to the audit log. The weekly override report drives the next quarterly review: an override reason appearing three or more times is a candidate for a new scoring criterion. Overrides are the correction mechanism that keeps the model honest.

How do I know when the scoring model has stopped working?

Five signals: close rate by band has flattened (hot and warm closing at similar rates); the hot band is closing above 30% (too few Leads in hot) or below 10% (too many); a criterion's lift has dropped by half (the signal is decaying); override frequency has spiked (the model is lagging reality); or reps are working the pipeline without filtering by score (they do not trust the score anymore). Any one signal justifies a mid-cycle review; two or more means the model needs re-weighting now.

Can AI improve lead scoring?

Yes, as a second opinion rather than a replacement. Strkr AI reads the Lead's first interaction (the inquiry email, the demo request, the chat transcript) and emits a confidence rating that sits next to the rule-based score. When the rule-based model and the AI agree, the Lead routes cleanly. When they disagree, the signal in dispute is investigated. The arrangement captures nuance the rule-based model cannot (buyer tone, implied urgency, decision-maker language) while preserving the transparency of a scored model that reps can read.

Related features

Pair this with...

Lead Scoring Criteria AI Lead Scoring Automated Lead Scoring B2B Lead Scoring

Try it free. Bring your team next week.

No sales call, no migration consultant, no four-month implementation. Enter your card, get 14 days of the full Pro tier, cancel any time before day 14 with zero charge. Spin up a workspace, import your CSV, and have something useful before lunch.

Start free trial See pricing