Predictive Segmentation and Engagement Scoring

An engagement score built on opens is partly a score of machine activity, and it errs in the direction that costs most.

Privacy protection pre-fetches content whether or not anyone read the message. So a contact who has not looked at an email in a year can register as consistently engaged — and stay in your active segment, receiving campaigns, generating no interest, and dragging your engagement rate down.

Security gateways compound it from the other side, fetching content and following links during pre-delivery scanning. On a B2B list this produces clicks from people who never saw the message, sometimes on every link.

The result is a score that keeps the wrong people in and cannot identify who to remove. Which matters, because deciding who to stop emailing is the single highest-value segmentation available.

---

What to build a score from

Four signals, in order of how much they mean.

1. Destination events. Purchases, bookings, sign-ups, logins, downloads — anything that happened on your own property. Unambiguous, attributable, and impossible for a proxy to fabricate.

2. Replies. Rare, and the strongest single signal on the list. A reply is a person who chose to spend time on you.

3. Clicks, with the B2B caveat. Better than opens and not clean. The tell for a fabricated click is a burst on multiple links within seconds of delivery, often at an unusual hour. Filter those before scoring.

4. Recency of any of the above. Recent beats frequent. Somebody who bought last month and nothing before matters more than somebody who bought four times two years ago.

What not to use: opens, on their own, for anything.

A workable score without a model: days since last destination event, days since last verified click, and total destination events in the last year. Three numbers, computed in a spreadsheet, and better than a predictive score trained on opens.

---

The segmentation that pays for itself

Deciding who to stop sending to.

This is the highest-value segmentation available and it is the one nobody wants to run, because it means reducing list size on purpose.

Why it pays:

• Engagement rate rises mechanically when you stop sending to people who never engage, and providers weigh recipient engagement in placement decisions
• Complaint rate falls, because complaints come disproportionately from people who forgot subscribing
• Bounce rate falls, since old addresses are concentrated in the dormant segment
• Spam trap risk falls. Recycled traps are former real addresses that providers repurposed precisely because nothing legitimate should still be mailing them. A dormant contact is exactly the profile. What traps do

The order of operations matters:

1. Identify contacts with no destination event and no verified click in a defined period — six or twelve months, depending on your buying cycle
2. Send one, at most two, re-engagement messages with a genuine reason to respond
3. Suppress the non-responders. Not delete — suppress, so they are not re-imported later
4. Do not repeat the attempt in three months. A contact who ignored a re-engagement message has answered

Run it small the first time. Suppressing a large dormant segment in one go changes your sending volume sharply, and sharp volume changes are themselves a reputation signal. Why volume patterns matter.

---

Where predictive segmentation genuinely helps

Three applications that do not depend on inbox signals.

1. Grouping by what people actually asked. Classify a year of replies, support tickets and survey responses into themes. You will find segments you did not know existed, defined by problem rather than by demographic — and those are the segments worth writing different emails for.

2. Purchase-pattern segments. What people bought, in what order, at what interval. Genuinely predictive, and built entirely on your own transaction data.

3. Lifecycle stage. New, active, lapsing, lapsed, defined by destination events. Simple, robust, and it drives most of the value people expect from sophisticated scoring.

What these have in common: the input is something that happened, not something that was fetched.

---

Predictions are inputs, not statements

A churn prediction is a legitimate reason for you to change what you send. It is never something to tell the recipient.

"We noticed you haven't been engaging" is a model's guess about somebody's inner life, delivered as an observation — and it is frequently wrong in the specific case, because the underlying signal contains machine activity. Telling somebody who reads every email that they seem to have lost interest is a poor way to open. What to personalise instead.

Use predictions to decide. Use facts to speak.

---

Over-segmentation

A real cost, and it arrives quietly.

Every additional segment reduces the volume behind each send, which has three consequences:

• Testing becomes impossible. A segment of 200 cannot detect anything
• Maintenance grows. Each segment is a definition that can drift, and a rule that can stop matching
• Reporting fragments, so nothing has enough data to read

A useful ceiling for a small list: as many segments as you can write genuinely different emails for. If two segments receive the same message, they are one segment — and the extra definition is maintenance with no return.

Three or four well-defined segments outperform twelve theoretical ones, and they are still there in a year.

---

Frequently asked questions

Why are engagement scores based on opens unreliable?
Privacy protection pre-fetches message content whether or not anybody opened it, so a contact who has not read an email in a year can register as consistently engaged. The error runs in the expensive direction — it keeps disengaged contacts in your active segment rather than flagging them.

What should an engagement score be built on?
Destination events first — purchases, sign-ups, logins, anything on your own property — then replies, then clicks with fabricated ones filtered out, then recency. Three numbers computed in a spreadsheet beat a predictive score trained on opens.

How do I identify fabricated clicks?
Look for a burst across multiple links within seconds of delivery, often at an unusual hour. That pattern is a security gateway following links during pre-delivery scanning, and it is common on B2B lists. Filter those events before scoring.

Should I remove inactive subscribers?
Yes, and it is the highest-value segmentation available. Engagement rate rises mechanically, complaint and bounce rates fall, and spam trap risk falls — dormant contacts are exactly the profile providers convert into recycled traps. Suppress rather than delete, and do it gradually.

How many email segments should I have?
As many as you can write genuinely different emails for. If two segments receive the same message they are one segment, and the extra definition is maintenance with no return. Three or four well-defined segments outperform twelve theoretical ones.

Can AI predict which subscribers will churn?
It can produce a prediction worth using as an input to your own decisions. What it should not do is tell the recipient — the underlying signal contains machine activity, so it is frequently wrong in the specific case, and informing an attentive reader that they seem disengaged is a poor opening.

How often should I run a re-engagement campaign?
Rarely, and with a hard stop. One or two messages, then suppress non-responders, and do not repeat the attempt a quarter later. A contact who ignored a re-engagement message has answered, and mailing that segment repeatedly is the fastest way to damage a sending reputation.