Businesses with large catalogues tend to end up at one of two extremes, and both of them are miserable.

At one end, everybody gets everything. There are four thousand products, no sensible way to choose between them, so the emails becomes a rotating window display and every subscriber receives the same eight items regardless of whether they have ever shown interest in that part of the range.

At the other end, somebody has built forty segments. They made sense on the day they were built, half of them have not been reviewed since, three refer to a category that no longer exists, and the person who understood the logic left in March. Nobody can safely delete any of it, so it all stays, generating collisions.

The reason both happen is the same, and it is the thing worth fixing before you touch a single audience definition. Your product catalogue is a taxonomy built for your business, for merchandising, buying, stock and reporting. Your customer does not think in your categories, has never seen them, and would not recognise the distinctions if you showed them.

So segmenting by product category means segmenting your audience along lines that exist inside your operation rather than inside their head, which is why it produces so much work for so little result.

Why it breaks

The specific ways category segmentation falls over

• People do not belong to one category. Somebody who bought a tent is not a camping person; they are a person who bought a tent. They may also be a runner, a parent buying school shoes and somebody who orders the same coffee every month. Filing them under one heading is a decision about your reporting, not about them.

• One purchase is a weak signal. Gifts, one-offs, experiments, panic buys and things bought for somebody else all look identical to a first purchase in a category. Building a content stream on top of a single data point is how somebody who bought their nephew a football gets nine months of football emails.

• The catalogue turns over, and the segments do not. Products get discontinued, categories get restructured, ranges get renamed. Every one of those changes silently breaks a segment somewhere, and nobody finds out until performance drifts.

• More segments means more overlap. Every new segment increases the number of people who qualify for several things at once, so the more granular you get the more collisions you create, and collisions are experienced by the customer as being shouted at.

• Every segment is a maintenance liability forever. Somebody has to own it, review it, and know why it exists. If you cannot name that person, you have not built a segment, you have built a future problem.

• Merchandising wants something different from what you want. The business needs to move stock and protect margin, which is legitimate and pulls directly against relevance. Segments built to serve the stock position will always drift away from serving the customer.

The layers

Build dimensions, not folders

The mental shift that makes this manageable is to stop thinking about segments as boxes that people go into, and start thinking about dimensions that every person has a value on. A subscriber is not in the camping segment. A subscriber has a lifecycle stage, an affinity picture, a current intent state, a position in a cycle, and a risk profile, all at once.

Which means you stop maintaining forty audiences and start maintaining five fields, then combine them at send time. Far less work, far more precision, and the whole thing survives a catalogue restructure.

One: relationship stage

Where they sit with you, independent of anything they bought. Never purchased, first purchase, repeat, high value, lapsed, and dormant. Cuts across every product in the range, changes what you should be saying more than any category does, and it is the layer that most catalogue businesses skip entirely.

A first-time buyer and a customer of six years should not receive the same email about the same product, and getting that right is worth more than any amount of category precision.

Two: affinity, built properly

What they lean toward, accumulated over time rather than inferred from one order. More on how to build this in the next section, because it is where most of the mistakes live.

Three: intent, which is about right now

What they are doing this week. Browsing, searching, adding to basket, checking stock, opening product emails and clicking through. Time-limited, powerful while it lasts, and worthless a fortnight later.

Four: cycle position

When they are next likely to need something. Replenishment intervals differ enormously by product, so a business selling both consumables and durables has two entirely different clocks running through the same customer base. Add seasonality, and occasions where you know the date, and you have a calendar built around them rather than around your promotional plan.

Five: value and risk

Engagement level, deliverability risk, and discount dependency. The last one is underrated and slightly painful to look at. Somebody who has only ever bought on promotion is a different commercial proposition from somebody who buys at full price, and treating them the same trains the second group to wait.

The combination

Five dimensions, combined at send time, give you more useful precision than forty standing segments, and they do it without creating forty things to maintain.

A repeat customer with strong affinity in one part of the range, no current intent signal, sitting mid-cycle, and paying full price is a specific person you can write to. No category segment would ever have found them.

Affinity

How to know what somebody is interested in without guessing

Affinity is where product businesses do the most damage, because the temptation to act on thin evidence is enormous and the tools make it easy.

One purchase is not affinity

Affinity needs either repetition or breadth. Somebody who has bought three things from a part of your range has told you something. Somebody who has bought once has told you almost nothing, and the odds that it was a gift, a one-off or an experiment are far higher than most segmentation logic assumes.

The tell for a gift purchase is that it sits outside everything else you know about that person, at a moment that lines up with an occasion, often at a price point they never otherwise touch. If your rules cannot distinguish that from a real preference, you will spend months emailing people about somebody else's hobby.

Build it from several signals, weighted

• Purchase carries the most weight, and needs repetition to count. Two or more in the same area, or one plus supporting behaviour elsewhere.

• Repeated browsing counts, single browsing barely does. One product page view is noise. Four visits to the same part of the range in a fortnight is a statement.

• Clicks from email are useful and cheap to collect. Tag every link with a topic or category code and your click data becomes an affinity feed you were generating anyway and probably throwing away.

• Declared preferences count, and count less than behaviour. What people tick in a preference centre is what they aspire to want. What they click is what they want.
  
  

Decay it, or it lies to you

Affinity from two years ago is history, not preference. People move house, change jobs, take up and abandon hobbies, and have children who grow out of things. Every affinity score needs to fade with time, so that recent behaviour outweighs old behaviour automatically rather than requiring somebody to remember to clean it up.

And be willing to say you do not know

The most useful thing you can build into an affinity model is an explicit unknown state, because for a large part of your list you have no idea, and pretending otherwise produces worse email than admitting it. High confidence, medium confidence and unknown are three different treatments, and unknown is not an embarrassment, it is a prompt to go and find out.

The unknown

What to send the majority of your list, about whom you know nothing

Worth saying plainly, because segmentation articles tend to skip past it. In most catalogue businesses, the majority of the list has never bought, or bought once a long time ago, and has generated almost no usable behaviour. That group is not the exception, it is the bulk of your file.

Guessing at them produces the catalogue email everybody hates. So the better approach is to use the send itself to learn, which turns your least targeted email into your best data collection tool.

1. Give the email several doors. Three or four distinct directions rather than one product push, each clearly labelled, each linking somewhere different.

2. Tag every link with a topic code. So the click writes to the affinity model automatically rather than being reported and forgotten.

3. Act on the click within days. The signal is worth most immediately, so the follow-up should be automated rather than waiting for next month’s plan.

4. Ask directly, once, and make it change something. A single preference question is fine if answering it visibly alters what arrives next. If it does not, you have collected a field and taught somebody that answering you is pointless.

5. Let people self-select into a stream. Rather than assigning them, offer the choice, and treat taking it as a strong affinity signal in its own right.

Collisons

The problem you create by segmenting well

The better your segmentation gets, the more often one person qualifies for several things at once, and a catalogue business is where this gets worst. Somebody can plausibly be due a replenishment reminder, sitting in a browse abandonment flow, inside a category campaign, receiving the newsletter and getting a back-in-stock alert, all in the same forty eight hours.

Each of those was correct on its own terms. Together they are a business shouting, and the recipient does not experience them as five well-targeted messages, they experience them as too much email from you.

• Cap frequency across campaigns and automations together. Flows do not pause because the campaign calendar got busy, and the person receiving both is the one who complains.

• Decide precedence before you need it. When two things could go, one of them wins by rule rather than by whoever scheduled first. Triggered beats broadcast almost always, because triggered is responding to something the person did.

• Exclude recent purchasers from the thing they just bought. Obvious, routinely missed, and the fastest way to look like you are not paying attention.

• Suppress across the whole system, not per campaign. Exclusions applied campaign by campaign will eventually be forgotten on the campaign that matters most.

There is a whole post on sending hierarchy, because deciding which email wins is a bigger subject than it looks and it is the thing that quietly wrecks otherwise well-built programmes.

Restraint

What not to build

• Do not build a segment nobody owns. If you cannot name the person responsible for reviewing it, do not create it.

• Do not replicate your merchandising taxonomy. Your internal category tree exists for buying and stock control. It is not a customer model and it was never designed to be one.

• Do not segment to satisfy a slide. Granularity looks impressive in a deck and costs you every week afterwards. The right number of segments is the smallest number you can maintain and prove.

• Do not send product recommendations you cannot stand behind. A weak recommendation is worse than none, because it tells somebody you do not know them while appearing to claim that you do.

• Do not use the full catalogue as a fallback. When you do not know what to send, sending everything is not a neutral choice, it is the choice that teaches people your emails are not worth opening.

Where to start

1. Pick your three highest-volume products or categories, and nothing else. Get those right before you attempt breadth. Most of the value in a large catalogue sits in a small part of it.

2. Build relationship stage first. It applies to everybody, it never goes stale, and it will improve your results before you have modelled a single affinity score.

3. Start tagging your links. Costs almost nothing, starts building affinity data from the next send onward, and you will wish you had done it a year ago.

4. Add one intent trigger. Browse or basket abandonment, properly excluded, before you build any more.

5. Write down what you have built and who owns it. The documentation is what stops this becoming the forty-segment mess in eighteen months.

The conclusion

Segmentation in a catalogue business is not a modelling problem, it is a restraint problem. The tooling will happily let you build a hundred audiences and the business will happily ask you to, and neither of those is the same as knowing your customers better.

Five dimensions you maintain properly will beat forty segments you maintain badly, every time, and they will still be working after the catalogue gets restructured.