Segmentation: sending different email to different people

A single broadcast to an entire list has to serve people at incompatible stages at once. Someone who subscribed yesterday needs to know who you are and what they signed up for. Someone who has bought from you four times does not, and an email explaining the basics to them reads as though you have not noticed they are a customer.

The usual response is to aim at the middle, which produces an email that is slightly wrong for everybody. It is too basic for the people most likely to buy and too advanced for the people who have not yet decided to trust you, and because it offends nobody it is easy to keep sending indefinitely while performance drifts down.

Segmentation is not a sophistication for large senders. It is the recognition that relevance is the main variable you control, and that relevance is a relationship between a message and a particular reader rather than a property of the message alone. A list of four hundred people usually contains at least two groups that want different things, and separating them is the change with the most leverage available.

Most useful segmentation reduces to four dimensions, and the first two carry most of the value.

Lifecycle stage: not yet bought, bought once, bought repeatedly, lapsed. This maps directly onto what the reader needs to hear next, and every business already knows it.

Behaviour: what someone did rather than who they are. A Pune fitness studio can separate people by which class they attended; a Surat textile wholesaler by which catalogue someone requested. This is the axis that produces email a reader recognises as being about them.

Purchase history: category, value, recency. Useful for deciding what to show, and easy to over-engineer.

Engagement recency: how recently someone did anything at all. This one is dual-purpose — it shapes content, and it is the input to list hygiene decisions later. Note that it has to be built on clicks and purchases rather than opens, since preloaded tracking pixels make open data an unreliable measure of whether a human was involved.

The constraint nobody accounts for is not the tool, it is the writing. Every segment is a message somebody has to draft, proofread, schedule and eventually maintain. Twelve segments is twelve pieces of copy per campaign, and the predictable outcome is that two get written properly and the other ten receive a lightly edited version of the same email — which is the broadcast you were trying to escape, with more administrative overhead attached.

Two segments, genuinely served, beat twelve neglected ones. The pair that earns its place first is almost always customers and non-customers, because the difference in what those two groups need is the largest difference on the list and the cheapest to act on.

Add a third only when you can name what it would receive that the existing two would not. If the answer is 'roughly the same email', the segment is not real yet. This test also protects you from the failure mode where the segmentation exists in the tool but not in the sending, which looks like progress in a dashboard and changes nothing for any reader.

Most small businesses can build the first useful segments from data already sitting in their systems: order history, signup source, which form somebody filled in, which link they clicked, which page they came from, whether they have ever booked. None of this requires asking anybody anything.

The temptation is to jump to a preference survey instead — a form asking subscribers to select interests. These have two problems. Stated preferences and actual behaviour diverge, sometimes sharply. And a survey generates fields that must be maintained and acted on, which is a commitment most senders quietly abandon, leaving data that is worse than none because it looks authoritative.

The discipline is to collect only what changes what you send. If knowing somebody's industry would not alter a single email, the field is administrative overhead wearing the costume of personalisation. Behavioural data has the opposite property: it accumulates without asking, and it is a record of what people did rather than what they said they would do.

These get used interchangeably and they are different operations with different risk profiles. Segmentation decides which email a person receives. Personalisation changes fields inside a single email — a name, a product, a recommended item.

Segmentation is the higher-leverage of the two and the more robust. A well-chosen segment gets an email written specifically for its situation, from subject line to call to action, and there is no template machinery to fail. Personalisation is a shallower change and it breaks in public: an empty merge field, a fallback nobody set, a recommendation engine suggesting the item somebody just bought.

The common mistake is reaching for personalisation because it is the feature the tool advertises, while sending everybody the same email. A generic message addressed to Priya by name is still a generic message. A message actually written for people who booked a room last winter does not need her name in it to feel relevant, because relevance came from the choice of audience rather than from a token.

The comparison that misleads people is segment against aggregate. A small, high-intent segment will outperform your old broadcast on almost any metric, and that tells you nothing about the segmentation — it tells you that you selected engaged people and then measured how engaged they were.

The honest comparison is each segment against its own prior baseline. Did the customers segment do better than the same customers did when they were receiving the broadcast? That question is answerable and the answer is sometimes no, which is the point of asking.

Watch unsubscribes and complaints per segment as closely as clicks, because segmentation can go wrong in a specific way: a segment defined by an assumption the readers do not share receives email that feels intrusive rather than relevant. And give it time. Segmentation changes the relationship over a sequence of sends, not within one, so a fortnight of data mostly measures which day you sent on.