Hierarchy decisions. Every report you will ever run is shaped by a tree somebody drew in the first week.
What the tree is for
Division, category, subcategory, type. The hierarchy is not filing, and it is not a menu. It is the axis every count, every margin and every sell-through will be grouped by, drawn in the first week by whoever was setting the system up, and inherited by everything downstream from that day on.
There is a straightforward test for whether it is finished. Write down the four or five questions your buying and planning meetings actually ask (how did outerwear do against last year, what is the sell-through on knitwear by subcategory, which division is carrying the margin) and check that the tree answers each of them by grouping alone. If it cannot, the answer does not stop being needed. It gets rebuilt by hand in a spreadsheet, every month, by somebody senior enough to know which rows to move.
It is also the most copied thing you will draw. The shop's navigation, the ERP's product groups, the analytics tool's dimensions and every report anyone builds all take their shape from it. That is the real cost of getting it wrong: not the tree, but the six places that have already agreed with it.
Four levels, and what each one is doing
The shape that survives contact with a real assortment is almost always four levels, each answering a different kind of question.
Two tests decide whether a level has earned its place. The first is that somebody can name the report that needs it. A level exists to answer a question that is actually asked; if nobody can name the question, the level does not exist yet, and adding it now in case it is wanted later is how trees get deep.
The second is harder and matters more: can two people who have never met file the same garment at that level and agree? If they cannot, the level is an opinion, and an opinion is a poor axis to group your numbers by. Every article filed under it is one person's judgement on a day, and the report that adds them up reads like a fact.
Depth costs more than breadth
These two are not the same kind of change, and they are routinely treated as if they were. Adding a category alongside the ones you already have costs one decision, once. Adding a level costs a decision on every article from now on, taken correctly, including by people who were not in the room when the level was defined and who will never read the note explaining it.
What that looks like in practice is always the same sequence.
Nothing in that sequence was a mistake. Padded against shell is a genuine distinction, and separating them was the sensible thing to do that week. The error was earlier and quieter: padding is a property of a jacket, and a property put into the axis has to be decided once per article and can only ever have one answer. As a filter it costs nothing, because a filter is allowed to be true of some jackets and not others without anyone having to adjudicate the boundary.
Four levels that are always right beat six that are guessed at. Add a level when a real question needs it, not in case one does.
One article, one place
If a garment plausibly belongs in two categories, the tree is wrong at that point, not the person filing it. This is worth saying to the person who raises it, because the reflex in most teams is to write a rule about what to do in that case, and a rule about the exception is how the exception becomes permanent.
The two answers that look reasonable in the moment are both expensive.
Let whoever creates it decide
Both placements are defensible, so the choice becomes personal. The axis your reports group by is now partly a record of who happened to create each article, and nobody downstream can see that from the numbers.
Add a level to hold the difference
The overlap gets a node of its own, and every article created after it carries one more decision to get right. You have bought one problem out of the way by paying for it in perpetuity.
The answers that hold are the two nobody suggests first. Either the categories overlap because the distinction was never load-bearing, in which case merge them and the question stops being asked. Or the distinction is real but is not a place: keep one category and move the difference into a tag, where an article can carry both without anyone having to choose.
What is a level, and what is a tag
The test is short. A level is a place. A tag is a property. An article can have many properties and exactly one place, so anything two of which can be true at once is not a level, however naturally it reads as one.
Season, collection, fit, fabric, a sustainability claim, carryover against new, and in most businesses gender: all properties. Each of them will be proposed as a level at some point, usually by someone who has just been asked for a report that splits by it; a report that splits by a tag is the same report, built the same way, without the tree having to change shape to allow it.
If two answers can be true at once, you are looking at a tag wearing a level’s clothes.
The exception worth naming is division. Division is often a property in principle (a garment could be sold into two of them) but it is treated as a place because the business is run that way: separate budgets, separate buyers, separate meetings. That is a legitimate reason for a level, and it is a different reason from the one that produces bad levels. The question is not whether the distinction is real. It is whether the organisation already resolves it.
Moving things later
Articles do get recategorised. A style moves out of a capsule and into the main range, a subcategory is split once it is finally big enough, a division is renamed after a reorganisation. None of that is a failure of the original tree; it is the assortment doing what assortments do. What matters is that the system lets it happen without breaking history.
What stays
The article itself: its number, its tech pack, its sample rounds, its costings and every order ever placed against it. Moving a style is a change to where it sits, not to what it is.
What moves
Its place in the tree, and every report from that day forward. Whether last year’s figures are restated into the new shape or left in the old one is a decision, and it should be made once rather than per move.
That is also the argument for keeping the meaning in the hierarchy and out of the article number. A number that encodes the category cannot follow the article when the category changes: either the number is edited, and every purchase order, label and spreadsheet quoting it is now quoting something that does not exist, or the article is retired and recreated, and its history splits in two. Neither is a good week, and both are avoidable by not asking the number to carry the category in the first place. More on article numbers →
Settle the restatement question in advance, before it comes up under pressure. Restating the past into the current tree makes year-on-year comparisons honest and quietly rewrites what people remember agreeing to. Leaving history in the shape it was reported in keeps every old deck correct and makes the comparison a manual job. Both are defensible. Having no answer is not, because it gets answered differently by whoever is in the room each time.
Before you draw it
Half an hour with the people who read the reports is worth more than a week of tidying the tree afterwards. These are the questions that decide the shape.
- What are the four or five questions the buying and planning meetings actually ask? Does the tree answer each one by grouping alone?
- For every level: can two people who have never met file the same garment there and agree?
- For every level: who can name the report that needs it?
- Which of your current levels is really a property (season, fit, fabric, a claim) and should be a tag instead?
- What does anything downstream need from this shape: shop navigation, ERP product groups, the analytics dimensions?
- When a category changes, does history restate or stay as it was reported?
The tree is cheap to draw and expensive to own. Four levels you are certain of will carry a decade of reporting without an argument. Six you are guessing at will be argued about every season, and rebuilt by hand every month by somebody who has better things to do.