Lead scoring is the practice of assigning a numeric value to each lead, usually 0 to 100, based on how likely they are to convert, so your team always knows who to prioritize. Instead of treating every lead the same and working whoever shouted last, you focus first on the leads most likely to buy. It is one of the highest-leverage things a sales or marketing team can do, because the same effort directed at higher-probability leads produces more revenue.
This guide explains what lead scoring is, how it works, the difference between rule-based and AI/predictive scoring, what signals to score, and how it helps.
TL;DR
- Lead scoring assigns each lead a value based on how likely they are to convert.
- It lets your team work the hottest leads first instead of treating all leads equally.
- Two approaches: rule-based (you assign points to attributes/actions) and AI/predictive (the system learns from your data).
- Score on fit (do they match your ICP?) and engagement (how interested are they?).
- Good scoring lifts conversion because effort goes where it is most likely to pay off.
- Negative scoring matters as much as positive scoring: a model that can only add points eventually scores everyone as hot.
- Scores must decay, or a lead who was interested six months ago outranks one who is interested today.
How lead scoring works
Lead scoring assigns points based on two broad dimensions:
- Fit (who they are): how closely the lead matches your ideal customer profile, company size, industry, role, location, budget signals. A perfect-fit lead scores higher.
- Engagement (what they do): how much interest they have shown, replied to a DM, opened emails, visited the pricing page, booked a call, used buying-intent language. More and stronger engagement scores higher.
A lead's total score combines both. High scores get flagged for immediate follow-up; lower scores enter nurture until they warm up. The point is to make prioritization automatic and data-driven rather than guesswork.
Lead scoring models: demographic, behavioural, and predictive
Almost every scoring system is one of these three, or a blend of the first two. The model type (which signals you read) is independent of the method (where the point values come from): a behavioural model can be rule-based or predictive.
Demographic and firmographic scoring
Scores who the lead is, from attributes true before the lead does anything: job title, seniority, company size, industry, country. Sometimes called explicit scoring, because the data is stated on a form or enriched from a company record. It answers one question: if this person wanted to buy, could they?
Where it fails: it is a ceiling, not a trigger. Firmographics are static, so a purely demographic score never changes, which means it can rank your list but cannot tell you who to call today. It is also the model most vulnerable to bad data, because a self-reported job title is an unverified string and "Director" means something different in a nine-person company than in a nine-thousand-person one.
Behavioural and engagement scoring
Scores what the lead did: replied, opened emails, clicked, visited pricing, booked a call, attended it, started a trial, used buying-intent language. Sometimes called implicit scoring, because interest is inferred from action rather than stated. This is the part of the score that changes daily, so it is the part that drives the work queue.
Where it fails: raw activity is not intent. Someone reading your blog every day may be a competitor, a job applicant, or a student. One pricing-page visit should usually outweigh ten blog views. It also breaks without decay, because points accumulate and never expire.
Predictive scoring
Does not ask which signals matter. It takes historical leads, labels them by whether they closed, and derives weights from the pattern, usually outputting a probability or a rank rather than a hand-assembled total.
Where it fails: it needs history, and a model learning from a few dozen outcomes is learning noise. It also learns whatever your past behaviour taught it, so if reps only ever called large accounts, the model concludes small accounts do not close, because in your data they never got the chance. It is also hard to argue with, which is why it is often run alongside a transparent rule-based score rather than instead of one.
| Model | Scores | Changes over time | Needs history |
|---|---|---|---|
| Demographic / firmographic | Who they are | No | No |
| Behavioural / engagement | What they did | Yes, constantly | No |
| Predictive | What similar leads did | Yes, as it retrains | Yes, substantially |
The practical sequence: start with demographic plus behavioural in one hand-built table, run it long enough to accumulate outcomes, then let a predictive model correct your weights. Starting with predictive scoring before you have outcome data means the system has nothing to learn from.
Rule-based vs AI/predictive lead scoring
- Rule-based scoring: you manually assign points to attributes and actions ("+10 for replying to a DM, +20 for visiting pricing, -10 for being out of ICP"). Simple, transparent, and easy to start, but static, you have to maintain the rules, and they reflect your assumptions rather than reality.
- AI / predictive scoring: the system analyzes your historical data, which leads actually converted, and learns the signals that predict conversion, then scores new leads accordingly. More accurate once it has enough closed outcomes to learn from, because it is based on what actually closed rather than what you guessed matters. This is what Inflowave's AI lead scoring does, analyzing conversation content, response patterns, and engagement to score every lead automatically.
What signals to score
Common high-value signals: replying quickly to outreach, using buying-intent language ("how much," "when can we start"), visiting key pages (pricing, demo), booking or attending a call, engaging repeatedly with content, matching your ICP firmographics, and (negatively) signals like being out of ICP or going silent. For DM-first businesses, conversation signals, what someone actually says in the DM, are among the strongest predictors, which is why scoring that reads conversation content outperforms scoring based only on clicks.
A lead scoring point table you can copy
A complete starting model on a 0 to 100 scale. Treat the numbers as a structure to argue with, not as benchmarks. They are chosen to satisfy three design rules, and those rules matter more than the specific integers.
Fit points (maximum 40)
| Signal | Points |
|---|---|
| Job title can sign the contract | +15 |
| Job title influences but cannot sign | +8 |
| Job title unrelated to the buying decision | 0 |
| Company size inside your target band | +10 |
| Company size one band above or below | +4 |
| Industry on your target list | +10 |
| Country you actively sell and support in | +5 |
Behaviour points (maximum 60)
| Signal | Points | Cap |
|---|---|---|
| Replied to an outbound message | +12 | once |
| Asked about price, contract, or start date | +15 | once |
| Booked a call | +20 | once |
| Attended the call | +15 | once |
| Started a trial or created an account | +18 | once |
| Attended a live demo or webinar | +10 | once |
| Visited the pricing page | +8 first, +3 after | +14 |
| Clicked a link in an email | +3 | +9 |
| Downloaded a gated asset | +4 | +8 |
| Opened three or more emails in 14 days | +5 | +5 |
The three rules the numbers obey
No single action can reach the sales-ready threshold alone. With a threshold of 50, the largest single award is 20, which stops one webinar registration from paging a rep. If a line clears your threshold by itself, it is not a score, it is a routing rule, so implement it as one.
Fit is capped lower than behaviour. Fit maxes at 40 and behaviour at 60, so a perfect-fit lead who has done nothing tops out below the threshold. Fit tells you who is worth working; behaviour tells you when. Let fit dominate and the queue never changes.
Every repeatable action has a cap. Anything a lead can do indefinitely, page views, opens, clicks, needs a ceiling, or the score measures persistence rather than intent. The cap also removes the incentive for a bot to climb your list.
A fourth rule applies once you are live: keep the table readable on one screen. A forty-row table is not more accurate than a fifteen-row one, only harder to audit, and nobody will ever revise it.
Negative lead scoring: what to subtract points for
A model that can only add points has one inevitable outcome: given enough time, everyone is hot. Negative scoring is what keeps the scale honest, and it does more work than the positive side because it removes leads from the queue rather than reordering them.
| Category | Signal | Points |
|---|---|---|
| Disqualifying fit | Country you do not sell or support in | -20 |
| Disqualifying fit | Company size far outside your serviceable range | -15 |
| Disqualifying fit | Title contains student, intern or professor, for a B2B product | -10 |
| Disqualifying fit | Free or personal email domain, for a B2B product | -8 |
| Explicit disinterest | Replied declining, or asked not to be contacted | -30 |
| Explicit disinterest | Unsubscribed from email | -25 |
| Explicit disinterest | Marked a message as spam | -25 |
| Explicit disinterest | Lost to a competitor in the last 90 days | -20 |
| Wrong intent | Email domain matches a known competitor | -30 |
| Wrong intent | Visited the careers or jobs page | -10 |
| Wrong intent | Only ever visited support or documentation pages | -5 |
Explicit disinterest is the strongest negative signal available and the one easiest to leave out, because it arrives as a reply or an unsubscribe rather than as a field. A fourth category, data quality, is better handled at the form than in the score: keyboard-run strings in a name field, phone numbers of repeated digits, a dialling code that contradicts the stated country.
Subtract, or hard-disqualify?
Some negatives should not be points at all. If the lead unsubscribed, or sits in a country you cannot serve, no amount of later engagement should make them sales-ready, yet under a pure points system enough clicking eventually claws the score back above the threshold. Split the mechanisms:
- Hard disqualify anything permanently true or legally binding: unsubscribed, do-not-contact, unserviceable region, competitor. Set a flag, exclude from the queue, stop scoring. Do not rely on a large negative number.
- Subtract points for anything that is a matter of degree or could change: one band off on company size, support-only browsing, a personal email address.
The test: would you want this lead back at the top of a rep's queue after three weeks of activity? If no, it is a flag, not a penalty.
How lead scores decay over time
An interest signal is a statement about a moment. A pricing-page visit today and one in March are not the same fact, but a plain running total records them identically. Without decay your highest scores drift towards whoever has been in the database longest, which is the opposite of what a work queue should surface. Three mechanisms, and the choice matters.
Flat inactivity decay. Subtract a fixed amount each interval a lead is idle, for example minus 5 points every 14 days, down to a floor. Simplest to build and the version most likely to be available without custom work. Its weakness is that it is blind to what earned the points: it drains a booked-call lead at the same rate as a blog reader.
Half-life decay on behaviour only. Every 60 idle days the behaviour subtotal halves, while fit never decays, because firmographics do not get stale the same way. The most faithful to reality, since interest fades gradually and fit does not fade at all, and it degrades smoothly rather than falling off a cliff.
Sliding window on each event. Timestamp every behavioural point and count only events inside a window, say 90 days. Points do not shrink, they expire. The most precise and the easiest to explain to a rep, but it requires storing events rather than a running total, and it drops suddenly as an old event leaves the window.
What decay looks like on one lead
Half-life decay, 60-day half-life on behaviour, fit fixed at 30.
| Day | Event | Fit | Behaviour | Total |
|---|---|---|---|---|
| 0 | Downloaded a guide, visited pricing | 30 | 12 | 42 |
| 5 | Replied to outreach, asked about price | 30 | 39 | 69 |
| 30 | Nothing since day 5 | 30 | 29 | 59 |
| 65 | Still nothing | 30 | 19 | 49 |
| 125 | Still nothing | 30 | 10 | 40 |
| 130 | Books a call | 30 | 30 | 60 |
The lead crossed the sales-ready line on day 5, dropped back under it around day 65 without anyone doing anything, and returned on day 130 on one strong new action rather than on residue from four months earlier. That is the behaviour you want. Three details cause most of the trouble:
Set a floor, usually the fit score. Never decay to a negative, or a long-dormant good-fit lead ends up scored like someone who told you to go away, which makes negative scoring meaningless.
Do not restore old points on re-engagement. Award the new action at full value on top of the decayed total. Resetting the clock so old points return rewards the lead for having once been interested.
Snapshot the score at handoff. If scores move on their own, a lead that converted at 71 may read 34 by the time you analyse it, and every retrospective about what score deals close at becomes wrong. Write the score and its contributing signals onto the record at handoff. That snapshot is also the training data a predictive model will need later.
Lead score thresholds and what happens at each band
A score with no threshold is a report. A score with thresholds is a routing system. Four bands are usually enough.
| Band | Reading | Action | Owner |
|---|---|---|---|
| 75 to 100 | Actively buying | Personal contact fast, same working hour | Closer |
| 50 to 74 | Sales-ready | Added to the call queue for the day, personal sequence | Rep or SDR |
| 25 to 49 | Warming | Automated nurture, alert a human on the next positive signal | Automation |
| 0 to 24 | Not ready | Low-frequency nurture, no sales time | Automation |
| Disqualified | Cannot or will not buy | Suppress, exclude from the queue and from reporting | None |
Set the threshold from capacity, not from a benchmark
There is no correct number for "sales-ready", and copying one from an article gives you either an empty queue or one nobody can finish. Derive it instead: count how many leads one rep can properly work in a week including follow-ups, multiply by headcount to get weekly capacity, then sort last month's leads by score and find the point at which the volume above it matches that capacity. Recheck whenever headcount or lead volume moves.
The value of the band table is entirely in the Action column. If the score is only a column on a list view, it will be sorted by once and then ignored. Wire each band crossing to something that happens on its own: a task created, a sequence started, a notification sent, an owner reassigned.
Lead scoring vs lead grading vs qualification frameworks
Grading is the fit half on its own, usually a letter from A to D. Some teams run grade and score as two axes rather than summing them, so a lead reads as "A2" or "C1". That is unambiguous in a way a single total is not: 55 could mean a perfect-fit lead who has done nothing or a poor-fit lead who is very active, and those need opposite responses. The cost is that two axes cannot be sorted into one queue without a rule for combining them.
Qualification frameworks like BANT or MEDDIC are checklists a human completes during a conversation. They cannot be computed from clicks, because the answers only exist once someone asks. Scoring decides who gets the conversation; qualification is what happens inside it.
How to set lead scoring points from your own closed-won data
The point values above are a starting position. Deriving them from your own outcomes is the single change that most improves a model, and it needs no machine learning.
- Pull a cohort with known outcomes. Leads created in a period long enough that most have resolved into won or lost. On a two-month sales cycle, a cohort from six months ago is safe; last month's is not, because the deals still open are disproportionately the good ones.
- Compute your baseline. Closed-won divided by total leads in that cohort. Every signal is compared against this.
- Compute the conversion rate of leads carrying each candidate signal, one at a time.
- Convert lift into points. A signal converting at three times baseline earns roughly three times the points of one converting at baseline. The exact scaling matters far less than the ordering.
- Drop thin signals. If only a handful of leads had it, the rate is noise. Eight trustworthy weights beat thirty invented ones.
- Validate on a cohort you did not use. If won and lost leads do not separate, the signals are not predictive and adding more will not help.
- Re-run it quarterly. Product, pricing and channel changes silently invalidate weights.
One caution about signals reps control: if "attended a call" scores highly, the rep chose to book that call, so you may be measuring rep judgement rather than lead intent. The same caution applies to the score overall. It says who to work next; it does not say whether the deal will be large, whether the customer will stay, or whether the channel paid for itself.
Why lead scoring matters
Without scoring, teams waste time: they chase low-probability leads, let hot leads go cold, and prioritize by recency or volume instead of likelihood to buy. Scoring fixes this by surfacing the best leads first, so the same effort yields more deals. It also enables automation, high-scoring leads can trigger instant follow-up or routing to a closer, while low-scoring leads enter a nurture sequence automatically. The result is higher conversion from the same lead volume, which is pure recovered revenue.
FAQ
What is lead scoring?
Lead scoring is the practice of assigning each lead a numeric value, typically 0 to 100, based on how likely they are to become a customer, so your team can prioritize the most promising leads. Scores are based on a combination of fit (how well the lead matches your ideal customer) and engagement (how much interest they have shown). High-scoring leads get worked first or routed for immediate follow-up, while lower-scoring leads are nurtured until they warm up.
How does lead scoring work?
It assigns points across two dimensions: fit (firmographic match to your ideal customer profile, like company size, industry, and role) and engagement (actions that signal interest, like replying, visiting pricing, or booking a call). These combine into a total score. In rule-based scoring you assign the points manually; in AI/predictive scoring the system learns from your historical conversion data which signals actually matter and scores automatically. High scores trigger fast follow-up; low scores enter nurture.
What is the difference between rule-based and predictive lead scoring?
Rule-based scoring is where you manually define point values for attributes and actions, simple and transparent, but static and based on your assumptions, and it requires ongoing maintenance. Predictive (AI) scoring analyzes your historical data to learn which signals actually predicted conversions, then scores new leads based on those real patterns. Predictive scoring is generally more accurate and improves over time because it reflects what truly closes rather than what you guessed would matter, though it needs enough historical data to learn from.
What signals should I use for lead scoring?
Score on both fit and engagement. Fit signals include matching your ideal customer profile, company size, industry, role, and budget indicators. Engagement signals include replying quickly, using buying-intent language, visiting key pages like pricing, booking or attending calls, and repeated content engagement, with negative points for out-of-ICP traits or going silent. For conversation-driven businesses, what a lead actually says in a DM or chat is one of the strongest signals, often more predictive than clicks alone.
Does lead scoring actually improve conversion?
Yes, when it is reasonably accurate, because it directs your finite sales effort toward the leads most likely to buy instead of spreading it evenly or by recency. Working high-probability leads first, and following up with them fastest, converts more of the same lead volume, which is recovered revenue at no extra acquisition cost. Scoring also powers automation (instant follow-up for hot leads, nurture for cold ones), compounding the benefit. The accuracy of the scoring model is what determines how much lift you get.
What is negative lead scoring?
Negative lead scoring subtracts points for signals that make a lead less likely or less desirable to close: a country you do not serve, a competitor domain, a careers-page visit, an unsubscribe. Without it points only accumulate, so given enough time every lead reads as hot. Keep penalties separate from disqualifications: anything permanently true or legally binding should set a disqualified flag rather than subtract a large number, or enough later activity pushes the lead back into the queue.
Should lead scores expire or decay?
Yes. A score answers "who should we contact today", and a total that never decreases answers "who has been in the database longest". Pick one mechanism: a fixed subtraction per idle period, a half-life on the behavioural component while fit stays fixed, or a rolling window counting only recent events. Set a floor so a dormant good-fit lead cannot sink below someone who told you to go away, and snapshot the score at handoff so later analysis is not reading a number that has moved since.
What is a good lead score threshold for sales handoff?
There is no universally correct number. Derive it from capacity: how many leads one rep can properly work in a week including follow-ups, multiplied by headcount. Sort last month's leads by score, find the point where the volume above it matches that capacity, and use it. Revisit whenever headcount or lead volume changes.
How many points should each action be worth?
Start from structure, not specific numbers. No single action should reach the sales-ready threshold alone, or it is a routing rule rather than a score. Cap fit below behaviour, so a perfect-fit lead who has done nothing does not top the queue. Put a ceiling on anything repeatable, so the score measures intent rather than persistence. Then replace the guessed values with weights derived from each signal's conversion rate against your baseline.
Related reading
Explore More
- What Is an SDR? Sales Development Rep Role, Pay & Future - What an SDR (Sales Development Representative) is, what they do day to day, SDR vs BDR vs AE
- What Is a Sales Engagement Platform? Definition & Examples - What a sales engagement platform is, what it does, how it differs from a CRM, examples
- Feature: Analytics - See which content and channels actually convert
- Feature: Sales Pipeline - Track every deal from first DM to closed
- Pricing - Plans built for Instagram businesses
