Build Lead Scoring That Actually Predicts Revenue
Actus · October 6, 2026
Build Lead Scoring That Actually Predicts Revenue
Most lead scoring feels like grading homework. Points for a job title, points for company size, points for opening an email. The score goes up, the lead gets routed to sales, and then nothing happens because the score measured activity, not intent to buy.
This guide is for teams that want lead scoring to predict which conversations will close, not which contacts clicked the most links. It covers what to score, what not to score, how to test whether the model works, and how an AI agent can run the scoring without a person tagging every field.
Actus Agent can read a lead's website, check whether they match your ICP, pull signals from their online presence, and write a brief that includes a numeric score. That score is only useful if it reflects things that actually correlate with revenue in your data. If you score everyone high, the sales team ignores the score. If you score based on vanity signals, you waste time on leads that were never going to close.
What Lead Scoring Should Predict
A lead score is a shortcut. It should answer: if I call this person today, what is the probability they become a customer in the next 90 days?
That probability comes from historical patterns. Look at the last 50 deals that closed. What did those leads have in common before the first call? Look at the last 100 leads that went quiet after one conversation. What did they have in common?
The score predicts the outcome. It does not create the outcome. A high score on a terrible offer will not save the deal.
Signals That Usually Matter
These patterns show up in most B2B and local service workflows:
They already tried to solve the problem. A lead who Googled your competitor, signed up for a trial, or asked a question in a Facebook group is further along than a lead who just downloaded a guide. Evidence of prior attempts is a strong signal.
They can describe the cost of inaction. In discovery calls, leads who say "we lose $X per week" or "the owner is doing this manually" close faster than leads who say "it would be nice to improve." If you can infer that cost from their site, their industry, or their team size, score it.
They match your best customers. If 70% of your revenue comes from HVAC companies with 5–15 employees in the Southeast, score those attributes. Do not score Fortune 500 enterprise just because it sounds impressive.
They have budget authority or access. A lead who is the owner, the department head, or the person currently paying for the thing you replace is more likely to close than an intern doing research.
They engaged with content that shows the problem is urgent. A lead who read "how to fix [specific painful thing]" is hotter than a lead who read "10 tips for general improvement." Score specificity, not traffic.
They reached out first. Inbound form fills, demo requests, and replies to a cold email are not equally hot. A demo request is hotter. A reply that asks a specific question is hotter still. Score initiation, not just acknowledgment.
Signals That Usually Do Not Matter
Email opens. Opens are noisy. Spam filters open mail. Previews open mail. The lead may have opened it and decided you are not relevant. Opens measure deliverability more than intent.
LinkedIn profile views. Unless you can prove the lead viewed your profile before they requested a demo, this is correlation without causation. People browse LinkedIn. It does not mean they are evaluating your product.
Company headcount in a range you do not serve well. If you have never closed a deal above 200 employees, do not give extra points to a 500-person company. That is aspiration, not prediction.
Content downloads that are unrelated to activation. A lead who downloaded an industry report may be doing research for school. A lead who downloaded a calculator or a checklist that diagnoses the exact problem you solve is warmer. Score relevance, not volume.
Page views on generic pages. Time on your About page does not predict revenue. Time on your pricing page, case studies, or implementation docs does.
How to Build a Scoring Model
Start with closed deals, not guesses.
- Pull every lead that became a customer in the last six months. Note what you knew about them before the first sales conversation. Do not include information you only learned later.
- Pull every lead that went cold after one or two touches. Note the same fields.
- Look for differences. If 80% of closed deals were small business owners and 80% of cold leads were mid-level employees, score decision-making authority. If 70% of closed deals visited your pricing page and only 20% of cold leads did, score pricing page visits.
- Assign points only to attributes that show a meaningful gap between closed and cold. If there is no gap, the attribute does not predict.
- Test the model on the next 20 leads before you route anyone based on it. Check whether high-scoring leads actually converted at a higher rate. If not, revise the weights.
Do not copy another company's scoring model. Their ICP is not yours. Their funnel is not yours. A model that works for a SaaS platform selling to marketing managers will fail for an HVAC shop selling to homeowners.
How an AI Agent Can Score Leads
Manual scoring breaks when the volume goes up. A human can score five leads a day by reading their LinkedIn and their website. A human cannot score 200 leads a day.
An agent can visit the site, check the attributes, and return a score. Here is a workflow that works:
- Give the agent a checklist. Example: Does the business have 5–15 employees? Is the owner listed on the site? Is there evidence they currently use [category] but not [your solution]? Did they fill a demo form or reply to outreach?
- For each yes, assign a point value you derived from closed deals.
- Ask the agent to cite where it found each signal. If it says the business has ten employees, it should link to the About page or the LinkedIn company page. If it cannot cite a source, it should not score the attribute.
- Return a total score and a one-paragraph brief. The brief should summarize the evidence, not restate the score.
- Route leads above a threshold to immediate follow-up. Route leads below the threshold to nurture or disqualify.
Actus Agent can run this as a scheduled workflow. Every morning it scores new leads, writes the briefs, and updates your pipeline. A person reviews the briefs, not the raw data.
Scoring Mistakes That Kill Trust
Scoring every lead high to avoid missing one. If the median score is 85 out of 100, the score is useless. Sales will treat all leads the same because the score does not differentiate.
Scoring based on effort instead of fit. A lead who opened five emails and visited six pages is not automatically qualified. They might be a student writing a report. Activity measures engagement, not buying intent.
Changing the model every week. When closed deals are slow, teams panic and reweight the model. A score that changes meaning every sprint confuses the sales team and makes historical analysis impossible. Change the model quarterly, after enough deals to see a pattern.
Scoring without a feedback loop. If sales never tells you whether high-scoring leads actually closed, you are scoring in the dark. Require closed-loop reporting: the score, the outcome, and the time to close.
Hiding the scoring logic. If the sales team does not know why a lead scored 40 versus 80, they will ignore the score and route based on gut feel. Publish the rubric. When someone asks why a lead scored low, you should be able to show the calculation.
Example: Scoring Local Service Leads
A digital agency sells websites to contractors and HVAC companies in Florida. Closed deals share these traits:
- Business has 5–20 employees.
- Website is more than three years old or is a single-page Facebook presence.
- They rank on Google Maps but have fewer than 10 reviews.
- The owner's name is on the site or in the Google Business Profile.
- They serve residential customers, not just commercial.
A lead that matches all five gets 100 points. A lead that matches three gets 60 points. A lead that matches one gets 20 points. Leads under 40 go to a nurture sequence. Leads above 60 get a call within 24 hours.
The agent workflow:
- Google the business name and open the site.
- Check the About page or footer for employee count. If not listed, check LinkedIn company page.
- Check domain age with a lookup or note if the site is a Facebook page only.
- Check Google Maps review count.
- Check for residential service mentions on the homepage or services page.
- Calculate score, write a three-sentence brief, and save to the CRM with a tag for score range.
That workflow runs in under two minutes per lead. A person doing it manually would take ten minutes and get bored after the fifth lead.
When to Disqualify Instead of Score
Some leads should not enter scoring at all. If a lead is outside your service area, if they are a competitor doing research, if they asked to be removed from contact, or if they are clearly a student or job seeker, do not score them. Disqualify immediately and suppress future outreach.
A disqualification rule is stricter than a low score. A low score might warm up over time. A disqualified lead should not re-enter unless something changes.
Testing Whether Your Model Works
After 30 days of live scoring, pull the data:
- What percentage of leads scored above 70?
- Of those, how many closed, how many are still in process, and how many went cold?
- What percentage of leads scored below 40?
- Of those, did any close? If yes, why did they score low?
A working model will show a clear conversion gap between high and low scores. If there is no gap, the model is not predictive. Revise the attributes.
If high-scoring leads are closing but the volume is too small, your ICP filter is too narrow. Loosen one attribute and test again.
If high-scoring leads are plentiful but not closing, your model is scoring the wrong thing. Go back to closed deals and find what you missed.
FAQ
Should I score leads in real time or in batches? Batch scoring works when leads come in slowly and you have time to review briefs. Real-time scoring works when leads are urgent (inbound demo requests, hot replies) and you need to route immediately. Most teams should batch score nurture leads and real-time score hot inbound.
How often should I update scores? Do not update unless the lead's situation changes. A lead who scored 50 last week and has done nothing since is still a 50. Recalculate only when they engage again, update their website, or reply to outreach.
What if I do not have 50 closed deals yet? Use your last ten and note the patterns. Your model will be rough. Revise it every 20 deals until you have enough data to stabilize.
Can I score leads I have not contacted yet? Yes, if the signals are public. You can score website age, employee count, and online presence before you send the first email. Use that score to decide whether to send at all.
Conclusion
Lead scoring predicts revenue only when it is built from revenue data. Points for opens, clicks, and page views measure activity. Points for ICP match, problem evidence, and authority measure likelihood to close.
An AI agent can visit sites, check signals, and calculate scores faster than a person. It should cite its sources and return a brief, not just a number. The score routes the lead. The brief tells the rep what to say.
Start with a simple model. Test it on 20 leads. Revise the weights when closed deals prove you were wrong. A scoring system that improves every quarter beats a perfect model you never ship.
If you want an agent that can research leads and return scored briefs, start with Actus Agent.