Generative Merchandising Content, Without Losing the Facts

Catalogue copy tends to be shaped like a pyramid. The top hundred products have descriptions somebody wrote, argued about and rewrote. The middle has whatever the supplier feed sent. The long tail has a name, a price and nothing else, because no merchandising team has ever been resourced to write for the tail.

Generation is the obvious tool for that gap. The risk is equally obvious. A language model asked to write engagingly about a product it cannot see will produce a fabric, a sizing note or a compatibility claim that reads perfectly and is not true. In most categories that is a return. In a few, it is a regulatory problem.

So the interesting question is not whether a model can write product copy. It is what the model is allowed to know.

Attributes in, sentences out

Our generator never receives a product name and a prompt. It receives the product record: the structured attributes, dimensions, materials, compatibility, care instructions and category placement you already hold, plus the attributes our tagging module derived from the product photography and the supplier text.

Copy is then written over that record, under one rule. If a fact is not in the record, it does not appear in the sentence. No inferred fabric weight, no invented sizing guidance, no “ideal for” claim the data does not carry. Where a field is genuinely missing, the description gets shorter rather than more imaginative.

Each generated field keeps a trace back to the attributes it drew on. That changes what a reviewer is doing. The question stops being “does this sound right” and becomes “does this match the record”, which is a question a person can answer quickly and consistently.

Voice is a constraint, not a flourish

Two hundred thousand descriptions each written in a slightly different register is not a catalogue, it is noise.

Voice is configured from a sample of copy your team already approves, plus rules stated explicitly: how you address the shopper, how long a sentence runs, the structure a description takes in each category, the vocabulary you avoid, and the claims that need sign-off before anyone writes them. Waterproof against water-resistant is a legal distinction, not a stylistic one, and the generator treats it that way.

Consistency is the thing automation does better than people. The same rules produce the same treatment whether a product landed in March or November, and regardless of who is on the team that month.

Nothing publishes itself

Generated text lands in a review queue. Merchandisers approve, edit or reject, in batches, filtered by category or by the fields the generator was least confident about. Only approved copy reaches the site.

The edits matter more than the approvals. A reviewer who rewrites the same phrasing forty times has just described a rule that belongs in the configuration, so the queue is where the voice actually gets tuned rather than a rubber stamp on the way out.

Copy also goes stale quietly. When an attribute changes, because a supplier switched a material or corrected a specification, the affected descriptions are flagged for regeneration instead of sitting on the page contradicting the product they describe.

The loop back into search

This is the part that pays twice, and it is the reason we treat generation as infrastructure rather than marketing.

Search and recommendation do not look at your photography. They read text and attributes. A product carrying three words of copy is close to invisible to a description-similarity model, awkward to place on a category page, and absent from the long-tail queries that convert best. Filling the tail with accurate, attribute-grounded text gives the retrieval step something real to work with.

It shows up hardest at cold start. A product launched this morning has no sales history, but with a proper description it can be placed next to the items it resembles immediately, and estimated on that basis.

Note the symmetry with automatic product tagging: tagging pulls attributes out of copy and images, generation writes copy from attributes. Run both and the catalogue gets better in both directions, which is why they are usually deployed together.

Generative merchandising is one of three parts of generative commerce, alongside a grounded shopping assistant and bounded post-purchase agents. To talk about your own catalogue, email office@mlart.biz.