Lead scoring for LinkedIn outreach: a simple model

A simple LinkedIn lead score has four parts: fit (does the person match your ideal customer), source (where the name came from), signal (a recent reason to write) and behaviour (what they did after you wrote). Weight the source and the signal highly, because on our own LinkedIn outbound a list built from an intent signal replied at 54.6% and booked 4.54 meetings per 100 messaged, against 21.7% and 0.71 for a Sales Navigator search.

Most lead scoring models were built for marketing teams with thousands of inbound leads and a CRM full of page views. LinkedIn outreach is different: you choose the list, the volume is a few hundred a week, and the best information is what you can see on a profile. A small model fits that job better than a big one.

Why score leads for LinkedIn at all?

Because your daily invitations are limited. Sellinger's defaults are 20 a day and 150 a week per account, and every slot spent on a poor fit is a slot not spent on a good one. A score gives you an order: contact the top first.

It also gives you something to learn from. Without a score, a campaign is one big list and you learn only its average. With a score, you can ask whether the people you rated high really did better, and if they did not, change the score.

What are the four parts?

Fit comes from your ICP template: sector, size, country, title. Give 3 for a clear match, 1 or 2 for a partial one, 0 for a blocker.

Source is the part most models forget, and it carries the most weight in our data. A name that came from a person engaging with a relevant post is a different thing from a name that matched a title filter. See below for the numbers.

Signal is a recent event: a job change, a post, a funding round, a hire. It answers why now.

Behaviour starts at zero, because you have not written yet. After you write, it moves: accepted the invitation, replied, asked a question, visited your site, booked.

What do the numbers say about source?

On our own LinkedIn account, 27,737 people were messaged and 10,978 replied (40%). By source, the October 2025 cohort showed intent signal lists replying at 54.6% and booking 4.54 meetings per 100 messaged, against 37.2% and 1.16 for an untagged list. A Sales Navigator search across the whole period replied at 21.7% and booked 0.71. The agents and the offer were the same.

Inside the intent lists the audience mattered too: people who engaged with posts from AI SDR vendors and automation creators booked 2.6% to 3.7% meetings per person messaged, and people who engaged with enterprise sales voices under 0.4%.

The ShortVibe study points the same way at small scale. 16 of its 19 booked discovery calls came from people engaging with relevant creators' and competitors' posts, and title and keyword searches booked 1 call in 225 conversations.

These are results of one account with its own audience, offer and definitions. Read them as a reason to test source in your own data, not as a rule.

How do you weight them?

Start with a simple sum out of 10 and adjust.

• Fit 3, source 3, signal 3, behaviour 1. Before you write, behaviour is zero, so the maximum is 9. After a reply it can reach 10.
• A blocker overrides the sum. If the person is in an excluded group, the score is zero.
• Source gets 3 only when the source has been shown to work. If you have not tested a source, give it 1 until you have.

Give these example scores to a few real leads:

The first gets written to today. The second goes in a later batch or gets a signal added first. The third is skipped.

How do you use the score?

Use three bands, for example 7 to 10, 4 to 6 and 0 to 3.

• 7 to 10: write now, with the personal opener, from your best account.
• 4 to 6: wait for a signal. Watch for a post or a job change, or move them to a different channel such as email if you have a verified address.
• 0 to 3: skip or park. Do not delete the lead. Contacts stay in your database, and you can revisit them when your ICP changes.

In Sellinger you can keep these bands as segments or tags and point each at a campaign. The post on intent signals worth acting on shows which signals tend to be worth a message.

How do you check the score?

Once a month, split your scored leads into thirds and compare the top third with the bottom third on two numbers: replies per person messaged and meetings per person messaged. Use the reply and acceptance rate calculator for the arithmetic.

If the top third is clearly better, keep the score. If there is no difference, one of your four parts is not predicting anything. Remove one part at a time and re run the comparison. A score that does not separate good from poor is only a ritual.

Beware of the trap in our own data: reply rates can hold steady while meetings drift. On our account replies stayed between 35% and 45% in each of twelve monthly cohorts, while meetings per person messaged drifted below 1% from June 2026. If you score on replies alone, you will miss it.

What are the common mistakes?

• Scoring on what is easy to see. Title and company size are easy to filter, so they get the most weight. In our data they predicted less than the source of the name did.
• Too many points. A model with fifteen traits cannot be explained or tested. Four parts and about seven traits are enough.
• Never changing it. A score is a hypothesis. After every hundred meetings, rerun the comparison of the top and bottom thirds.
• Scoring behaviour before you have any. Before you write, behaviour is zero for everyone. Do not invent a signal to fill the column.
• Letting the score decide alone. A person should read the top fifty by hand. If the sheet and your eyes disagree, learn why.

The glossary entry on lead scoring gives the short definition, and the ICP template shows how to write the fit part.

A checklist

• [ ] Four parts: fit, source, signal, behaviour.
• [ ] Source carries weight only after I test it.
• [ ] A blocker sets the score to zero.
• [ ] Three bands with a clear action for each.
• [ ] I park low scores, I do not delete them.
• [ ] Once a month I compare top and bottom thirds on meetings, not only replies.