From Beginner To Advanced Segmentation

Segmentation gets discussed as though it were one thing you either do or do not do. It is four things, stacked, and each level answers a question the level below it cannot.

The order matters. Level one is a filter and it never goes away, but on its own it cannot tell you a single thing about what to write. Levels two and three decide the message. Level four decides the product inside the message. Skip ahead and you get a beautifully targeted email sent into a spam folder; stop at level one and you get perfect deliverability on an email that says nothing to anyone.

Ideally every person on your list is a repeat customer who buys every single day. That is not going to happen, and we are trying to be realistic here.

The beginner move, and everybody does it. If somebody opened or clicked recently, send them the next campaign. Everybody else gets nothing, and the reporting stays beautiful.

The standard shape of this is a set of nested windows, and they are worth naming because the rest of the guide refers back to them. Engaged 30 days is the workhorse. Engaged 90 and 180 are progressively looser rings around it. Prospects usually get tighter windows than customers, because somebody who has never bought has to show more intent to earn the same reach.

None of that is wrong. Section 9 is entirely about why you keep it. But look at what the sort is actually reading.

So level one is a safety system. It protects the channel and it decides who is safe to mail. It cannot decide what to say, because the axis it sorts on has nothing to do with the thing you want. The axis that does is purchase behaviour, and that is level two.

3.1  What the letters actually mean

People say RFM the way you would name a classified programme. Advanced RFM analysis. We must be data nerds, we have PhDs, we hired a team. It is three words:

Monetary is a word that could not be more simple. We all understand what money means, in every language. All RFM does is read past purchase history and sort people into groups, so that each group can get a different message.

Take somebody who has bought three times a year for five years and spent an insane amount with you, and somebody who placed their first order yesterday. Send them the same email and you are failing one of them. The new buyer wants the natural next thing to buy. The five-year buyer has seen the natural next thing forty times and wants to know what is new. Point the new buyer at your bestsellers, not the experimental flavour, because if their second purchase is a disappointment there is no third.

3.2  You can usually drop the M

Monetary correlates well with recency and frequency. On most catalogues it is a third axis that repeats what the first two already said, and it doubles the number of boxes you have to maintain to get there.

It earns its place when your catalogue spans wildly different price points. If you sell both a pen and a tank, then one tank buyer is worth a thousand pen buyers and you cannot treat frequency as a proxy for value. Sun Tzu might disagree. Sun Tzu is not on the call. Whatever you sell probably sits inside one price band, so drop the M and keep the model simple enough that somebody will actually maintain it.

3.3  Score all three axes when you are setting strategy

When you are setting strategy rather than picking an audience for Thursday, score all three axes. Rank customers into quartiles on each, one being best, which self-calibrates to your own distribution instead of borrowing somebody else’s thresholds. That gives up to 64 cells, which nobody can run a calendar off, so people collapse them into named statuses: champions, loyal, high potential, new, at risk, and the ones worth sunsetting.

4.1  Two axes, nine groups

Drop the monetary axis and you are left with something you can hold in your head. On frequency, people have bought from you zero times, one time, or more than once. On recency, they are recent, at risk, or lapsed. Multiply them and you get nine groups, each of which wants a different email.

4.2  Set the bands from your own data

Thirty and ninety days are placeholders, and they are wrong for plenty of brands. Somebody who buys pens may well need pens again in three months. Somebody who bought a tank hopefully does not need another one unless there is a war on.

The band you want is your own median time between orders, and there are two ways people get this wrong. They use the mean, which a handful of enormous gaps drags upward, when they should use the median. And they pool every customer into one number, when the gap from first to second order is almost always longer than second to third, which is longer than third to fourth. Repeat buyers compress their own cadence. One pooled average is wrong for everybody in the file.

Past twelve months overdue, somebody is not due. They are gone, and they belong in a reactivation campaign rather than a replenishment nudge.

4.3  What each of the nine groups gets

There are four content buckets to allocate across the grid: promotional, product, educational and authenticity. Authenticity is the founder letter, the review, the story about why the thing exists.

The one rule that costs money when you break it: keep recent customers out of your biggest sale. They paid full price a week ago. Show them 40% off the thing sitting in their hallway and you have bought yourself a support ticket and a refund request instead of a second order.

For the at-risk group, do not guess which content works. Klaviyo will break campaign performance down by segment. Look at how that group performs against the others by email type, keep the two that work, and stop sending the rest.

4.4  Level 4: put the product layer on top

The nine boxes tell you what kind of message somebody gets. They do not tell you which product to put in it. That is the fourth level, and it is what personalisation actually means once you strip the word of its marketing.

Take one box, say lapsed repeat customers, and cut it by a product they have already bought. Now recommend whatever people who bought that product buy next. Not what is bought alongside it in the same basket, which is a different question with a different answer. What comes next in time.

One layered send replaces what would otherwise be three generic ones, which matters as much for restraint as for relevance: without the layer you either send one vague email or you inundate the same people three times.

5.1  Count the conversions email actually caused

Lapsed prospects signed up a long time ago and have never bought anything. Before you decide they are worthless, run the numbers, because some of them do convert and the whole argument turns on how many.

The naive version is to count how many buyers last month were lapsed prospects, find some small percentage, and feel reassured. That overstates email’s case badly, because most of those people did not buy because of an email. They saw an ad, or came back on their own, and the email list happened to be holding their address at the time. Those conversions survive suppression untouched.

So the real question is narrower: how many of them clicked an email in the few days before they bought? That click window is what separates conversions your list caused from conversions your list merely witnessed.

5.2  Price both sides

Keep the bands where email-attributable contribution clearly beats the carrying cost. Cut the ones where cost beats even a generous reading of attribution. The band in the middle is a judgment call, and deliverability pressure should break the tie toward cutting.

5.3  Cut without a guillotine

• Start with the free win. Profiles that never subscribed are billed and cannot legally be mailed. Suppressing those costs you nothing at all and is the obvious first move.
• Send a sunset sequence first. Two or three emails saying plainly that this is the last one unless they do something. Suppress the non-responders, not the whole band on day one.
• Check they were ever actually mailed. A profile that never received anything did not fail to engage. It was never asked.
• Suppression is not deletion. History stays, the profile stops being billed, and most platforms will bring somebody back automatically if they purchase.

The deliverability dividend is the part nobody puts in the model. A large disengaged band drags your open and click rates, and those rates are what the inbox providers use to decide where the rest of your list receives you.

A discount is only worth giving to somebody whose behaviour it changes. Give it to a person who buys from you every month anyway and you have not bought an order, you have bought the same order at a worse price.

At the other end of the grid, somebody who has not bought in two years can take a deep one. You had written that revenue off. Anything you get back is a bonus, like finding money in a coat pocket. You were not expecting it, and you will take it.

The interesting thing about these groups is that people move between them, and the movement itself is a trigger. Somebody who always bought crosses from active repeat into at risk. That transition is a moment, and you can fire a flow on it, which beats waiting for a monthly campaign to notice.

What you should not do is send one email, get ignored, and treat it as a personal betrayal. I am so offended you did not take my offer, it is me, not you, I am sunsetting. That is money left on the table over a single non-response. Build the follow-up. Two, three, four touches, spaced off the interval you computed in section 4.2 rather than off a round number.

A missing flow loses people one at a time and you never see it happen, which is exactly why the flow layer gets skipped in favour of the campaign calendar somebody has to fill in on Monday.

8.1  Revenue per unsubscribe, not open rate

Every decision in this guide is a trade: reach more people, burn more of them. So read each expansion as an exchange rate. What did the extra reach add in revenue, and how many subscribers did it cost you permanently?

Unsubscribes, not spam complaints, are usually the binding constraint. Complaint rates normally sit far below the level that gets you in trouble, while every unsubscribe is a permanent loss of all future revenue from that person. Include a ring while the revenue it adds beats the forward value of the subscribers it burns.

Two measurement traps worth naming. Opens are inflated by machine fetching and should not decide anything. And order count is not revenue: a group can look dead on orders and be fine on dollars, because the people out at the edges often buy bigger when they do buy. Go to dollars whenever the data lets you.

8.2  Small segments produce noisy numbers

When you slice a list this finely, the outer boxes get small fast. A dormant-repeat segment might be producing single-digit orders a month, and a rate computed on single digits is noise wearing a percentage sign.

Here is the problem nobody warns you about when you move from level one to levels two and three. You have just stopped sorting people by whether they engage with email and started sorting them by whether they buy. Those are not the same population.

Some of your best customers never open anything. They buy because they saw an ad, or because they ran out, or because they walked in through search. Under the old system they were quietly filtered out. Under the new one they are in your dormant repeat box and you are about to mail them your deepest offer. Good for revenue, potentially bad for the inbox, because Gmail and Yahoo are watching how the people you mail respond and drawing conclusions about you.

9.1  Why the engaged-30-day segment exists in the first place

The reason so many accounts run everything through Engaged 30 Days is that it holds the open rate up. The working target most operators keep to is somewhere around 40% to 50%, and a segment tight enough to clear that is a segment the inbox providers keep delivering.

9.2  Run the exclusion underneath everything

The fix is not to go back to level one. It is to keep level one running as a filter beneath the purchase-based segments. Build the audience once, exclude it from every campaign, and let the nine boxes decide everything else.

Klaviyo now has audience optimization built in, which will restrict the people unlikely to interact in a click and is the fastest way to do this. If you would rather build it yourself, the segment you want is the inverse of an engagement window:

The 10 is arbitrary and you should change it. It has to reflect how much you actually send in ninety days. If you send twice a week, ten campaigns is barely a month of evidence and the segment will be too broad. If you send twice a month, ten campaigns is more than the window holds and the segment will be nearly empty. Raise the number to make the exclusion tighter and your sendable list bigger, lower it to do the opposite.

9.3  Excluding the unengaged costs you a few buyers

You will exclude some people who would have bought. That is the honest trade and it is worth naming rather than pretending the exclusion is free. What you get back is that everybody else, including every customer in every box on the grid, keeps arriving in the inbox instead of the promotions graveyard.

Level one stops being your targeting the moment you reach level two. It becomes your floor. Every purchase-based segment in this guide is sent minus the unengaged, always, no exceptions.