Skip to content
Antiqore

A signal from March is not a signal in July

Most lead scores only ever go up, so they end up measuring database tenure rather than relevance. What it takes to make a score that fades the way the evidence does.

Antiqore5 min read

Almost every lead score is additive. Visited pricing, plus ten. Opened three emails, plus six. Company matches the profile, plus twenty. Nothing ever comes off, so the highest-scoring accounts are reliably the ones that have been in the system longest — which is not the same as the ones worth calling this week, and is occasionally the opposite.

What an old signal actually tells you

A company posted a role in March that implied your problem. In July, one of three things is true: the role was filled and the problem is now somebody’s job, the role was cancelled and the budget went elsewhere, or it is still open and something is stuck.

All three are interesting. None of them is the thing the score is asserting, which is that the March event is happening now. The score has not aged the evidence, so it is making a claim about the present using a fact about the past.

A score that only adds is a record of everything you have ever noticed, ranked by how long ago you noticed it.

Half-life beats expiry

The obvious fix is an expiry window — drop anything older than ninety days. It is better than nothing and it introduces a new problem: an eighty-nine-day-old signal counts fully and a ninety-one-day-old one counts for nothing. Companies jump around the list for reasons that have nothing to do with them.

A half-life fades instead. Set the period after which a signal is worth half its original weight, and let it decline smoothly. Relevance does not fall off a cliff, and neither should the number modelling it.

One rate per class, not one rate for everything

The second mistake is applying a single half-life to every input. Events resolve at wildly different speeds, and a rate that suits hiring signals will destroy a funding signal months before it stops mattering.

Signal classSuggested half-lifeWhy
Site visit or content read10–14 daysAttention is the fastest-fading thing there is.
Job posting6–8 weeksRoughly how long a role stays genuinely open.
Funding round4–6 monthsThe money is real for a year; the window to influence how it is spent is shorter.
Leadership change in your function3–4 monthsNew leaders re-evaluate tooling early and then settle.
Stated priority in a talk or interview2–3 monthsStated intent is weaker than funded intent and should fade faster.
Regulatory or contractual deadlineNo decayIt does not fade — it intensifies to a date, then stops entirely.
Starting points to check against your own history, not constants. The discipline is having a rate per class at all.

What decay changes about the list

Two things, both of them the point.

The top moves. Companies with recent, dated evidence rise; companies whose case rests on something from two quarters ago sink. This will surface accounts nobody has called, which is uncomfortable and correct.

Doing nothing has a cost. In an additive system a lead left alone stays exactly as attractive as the day it arrived. With decay it fades, which is an accurate description of what is happening to your chances.

Implementing it without a rebuild

  • Store the date of the underlying event, not the date the row was created. Without this, nothing else is possible.
  • Group your signals into four or five classes and give each a half-life.
  • Compute the weight at read time rather than writing a decayed number back — that way changing a rate does not require a backfill.
  • Show the age next to the contribution, so a rep can see why a company is where it is.
  • Re-check the rates once you have enough closed-won data to compare against.

The first item is the one that blocks most teams. If your records carry only an ingestion timestamp, you cannot decay anything, because you do not know how old the evidence is — only how long you have had it. That is the same requirement as the second test for evidence, which is not a coincidence: a claim you cannot date is a claim you cannot age, and a claim you cannot age will eventually be wrong without ever looking wrong.

In Leads, every signal carries the date it was observed and its contribution is computed at read time, so a company’s position reflects what is true this week rather than everything that has ever been noticed about it.

Frequently asked questions

Why should lead scores decay?
Because the underlying events do. A role posted in March has been filled or cancelled by July, but a score that only adds points will still be carrying it as if it happened this morning. Without decay, a score measures how long a company has been in your database.
What is a half-life in scoring?
The period after which a signal contributes half of what it originally did. It is a better model than a hard expiry because relevance fades gradually — a two-month-old funding round is weaker than a two-week-old one but not worthless.
Should every signal decay at the same rate?
No, and using one global rate is the most common mistake. Hiring intent fades in weeks. A funding round stays relevant for a couple of quarters. A regulatory deadline does not decay at all — it accelerates until the date, then stops.
How do I pick a half-life without data?
Start from how long the underlying event takes to resolve. If a posted role is usually filled within eight weeks, that is your half-life. Then check it against your own closed-won data once you have enough, and adjust.
Does decay make scores less stable?
It makes them move, which is the point — a score that never moves is not measuring anything current. What it should not do is oscillate. If a company's score swings weekly, your inputs are too noisy, and decay is exposing that rather than causing it.