Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

Experiment Library

Lead Scoring Threshold Test

A design for testing the lead score handover threshold: leads just below it randomised to sales or nurture, read on pipeline per lead.
Experimentadvanced7 min readUpdated

Hypothesis

If leads scoring just below the current handover threshold are passed to sales instead of nurture, enough of them will become pipeline to justify sales time, which would show the threshold is set too high. The decision behind it: lower, raise or keep the lead score at which marketing hands leads to sales.

Thresholds are usually set once, from a model and a judgement, and then defended by both sides. Marketing suspects good leads are stuck in nurture; sales suspects it is already receiving too many weak ones. Historical data cannot settle it, because leads below the threshold never reached sales, so nobody knows what sales would have made of them. The lead scoring guide explains how to calibrate a model against closed deals; this test supplies the missing evidence by sending some below-threshold leads to sales on purpose. Define the band, the routing and the guardrail limits in the experiment hypothesis template, and agree them with sales, before launch.

Metrics and guardrails

The primary metric is pipeline created per lead in the score band: the value of opportunities created from each arm's band leads within the follow-up window, divided by the number of band leads assigned to that arm. Pipeline is the outcome that justifies sales time, and the Marketing-Sourced Pipeline definition the business already uses should apply unchanged. The denominator includes every band lead assigned, so leads that sales rejects still count. In the nurture arm, band leads that later cross the threshold and reach sales in the normal way count too, which is what makes the comparison fair.

The guardrails protect sales capacity and the rest of the funnel. Sales acceptance rate for band leads shows whether sales considers them worth working; a low rate warns that the extra leads cost time. Rep time per accepted lead, logged or estimated consistently in both arms, turns that warning into a cost. Pipeline from leads above the threshold must not fall: if band leads crowd out reps' time, the better leads may be worked less well, and a gain in the band would be paid for elsewhere. The MQL, SQL and lead stages guide defines the stages the acceptance rate and the Sales-Qualified Lead count depend on.

Design

Define a band of scores just below the current threshold, wide enough to collect leads at a useful rate and narrow enough that its leads are plausibly close to ready. Randomise each lead at the moment its score first enters the band: half are routed to sales as if they had crossed the threshold, half stay in nurture. Leads above the threshold are routed as usual, and leads below the band are untouched. Freeze the scoring model for the test's duration.

Design elementChoiceReason
Unit of randomisationLead, assigned when its score first enters the bandOnly leads near the line are in doubt
ArmsBand leads to sales versus band leads to nurtureTests what moving the line down would do
Leads above the thresholdRouted as usual, not randomisedKeeps the rest of the funnel normal
Scoring modelFrozen for the whole testA changing model moves leads in and out of the band
Sales visibilityArm and band hidden on the lead recordPrevents reps working band leads differently
Primary metricPipeline per band lead within a fixed windowCounts outcomes, including nurture leads that later cross the threshold

The lead response time test uses the same principle of hiding the arm from reps for a different question, follow-up speed. Here, hiding the band matters even more, because a score is a visible signal of expected quality.

Sample size reasoning

Only the band is randomised, so the sample is the number of band leads, and pipeline takes a sales cycle to appear. Pipeline per lead is also skewed: most band leads create none and a few create large opportunities, so the spread is wide and modest differences are hard to separate from chance. The inputs are the band's expected pipeline per lead, how widely it varies, the smallest difference that would justify sales time and the confidence required; the sample size and significance guide explains how they combine and why skewed outcomes need larger samples. A thin band needs a long enrolment. If that is impractical, read opportunity creation rate, which is less skewed than pipeline value, and name it as the primary metric before launch rather than switching later.

Duration

Enrol until the planned number of band leads is reached, across full weekly and monthly cycles, since lead mix often follows campaign calendars. Then allow a follow-up window matched to the early stages of the sales cycle, long enough for accepted leads to become opportunities. Reading earlier favours the nurture arm, whose leads have not yet been asked to do anything. Set the end date before launch and avoid enrolling across a pricing change or a reorganisation of the sales team.

Reading the results and the decision rule

Read once, after the follow-up window closes for the last band lead enrolled. Compare pipeline per band lead between arms with its confidence interval, then check acceptance, rep time and pipeline above the threshold. The rule, written before launch:

ResultDecision
Band leads sent to sales create pipeline at an acceptable rep cost, and acceptance and above-threshold pipeline holdLower the threshold to the bottom of the band
Band leads create some pipeline but acceptance is poor or rep cost too highKeep the threshold; route band leads to a lighter-touch sales motion if one exists
Band leads sent to sales mostly stallKeep the threshold
Pipeline from leads above the threshold fallsKeep the threshold; the band is taking capacity from better leads

The B2B measurement guide explains how to report the new threshold's effect on MQL volume and pipeline afterwards, so that a change in the count of qualified leads is not mistaken for a change in performance.

Illustrative example — not a real result

An energy analytics firm tests a lower handover score

Northpane, an invented building-energy analytics company, hands leads to sales at a score of 60. It defines a band from 50 to 59 and, over five months, randomises 900 band leads: 450 to sales and 450 to nurture. After a 90-day follow-up window, band leads sent to sales have created £1,850 of pipeline per lead, against £620 in the nurture arm, where some leads crossed 60 later; the 95% interval for the difference is wide, roughly £400 to £2,050. Sales accepts 58% of band leads, rep time per accepted lead is similar to leads above 60, and pipeline from above-threshold leads does not fall. Northpane lowers its threshold to 50.

Pitfalls

Sales treating band leads differently once they know is the first pitfall. A score shown on the lead record, or a note that the lead is part of a test, invites reps to deprioritise it. Hide the score band and the arm, and give reps one rule for all leads they receive. The scoring model changing mid-test is the second: a new rule, a decay change or a re-weighting moves leads in and out of the band and changes what the band means. Freeze the model, and if it must change, end the enrolment and analyse up to that point. Reading too early for pipeline to form is the third, and it always flatters nurture.

If the handover problem is that qualified leads are not converting at all, rather than that the threshold is in the wrong place, start with the guide to MQLs not converting to SQLs before testing the threshold.

Common mistakes

Threshold tests fail when the band is visible or the clock is short. The experiment library lists the other lead handling designs.

  • Moving the threshold for everyone at once. Randomise the band and leave the rest of the funnel alone.
  • Showing reps the band. Expectations then decide the outcome.
  • Changing the scoring model during the test. Freeze it until the enrolment ends.
  • Reading before the sales cycle has started. Pipeline needs time to form.
  • Ignoring above-threshold pipeline. Capacity spent on the band may cost better leads.
  • Switching to a faster metric halfway through. Name any proxy before launch.

Frequently asked questions

Why randomise only a band below the threshold?

Because the question is where the line should sit, and only leads near the line are in doubt. Leads well above it already go to sales and leads far below it are clearly not ready. Randomising a band just below the threshold tests what happens if the line moves down to the bottom of that band, without disrupting the rest of the funnel.

Can the same design test raising the threshold?

Yes. Randomise a band just above the threshold between sales, as today, and nurture. If band leads routed to nurture lose little pipeline compared with those sent to sales, the threshold can rise and free sales time. The read and the guardrails are the same.

Should sales be told which leads are in the test?

Sales should know a test is running and agree its rules, but reps should not see a lead's arm or band on the record. If they can tell a lead came from below the usual threshold, they may work it less hard, and the test would then measure their expectations rather than the leads.

Terms used on this page

Frameworks

Also related

Related