Using AI to Write Product Descriptions That Do Not Sound Like AI
By The Velocity Wear Team
Ask a language model for a product description for a black hoodie and you will get something about elevating your everyday and effortless style that could describe any hoodie ever made. It is fluent, it is on-brand for nobody, and shoppers skip it. The problem is not the model. It is that you gave it nothing to work with.
Why generated apparel copy fails
A product description has one job: to answer the questions that stop someone buying. For clothing bought online those questions are always the same — how does it fit, what is it made of, how heavy is it, will the print last, and how do I care for it. Adjectives answer none of them.
When you prompt with "write a description for a black hoodie", the model has no specifications, so it fills the space with mood. That is the entire failure mode, and it is an input problem.
Feed it facts, not vibes
The same model produces genuinely useful copy when the prompt contains the things a customer needs to know.
- **Fabric and weight.** "400 GSM brushed-back cotton-rich fleece" is a fact. "Premium heavyweight feel" is a claim.
- **Fit and cut.** Relaxed, boxy, true to size, runs small. Include the measurements you actually have.
- **Decoration method.** Screen printed, embroidered, DTF. Each implies a different durability story you can tell honestly.
- **Care.** Wash cold, inside out, do not tumble. Care instructions in the description reduce returns and complaints.
- **Who it is for.** A gym, a band, a corporate rollout. Specific audience produces specific language.
The four tells to edit out
Even with good inputs, generated copy has habits. These four are the ones customers register as "this was written by a machine", usually without being able to say why.
- 1**The rule of three everywhere.** "Comfortable, versatile and timeless." Once is fine. Three times in four sentences is a rhythm no human writes by accident.
- 2**Elevation language.** Elevate, effortless, curated, essential, must-have. These words have been used so heavily that they now signal filler.
- 3**Claims with no source.** "Loved by thousands" and "our bestselling piece" are inventions unless they are true. On top of reading as filler, publishing invented social proof is unlawful in the UK.
- 4**Symmetrical paragraphs.** Three paragraphs of near-identical length is a generated shape. Vary it. Real writing is lumpy.
A structure that works
One sentence on what it is and who it is for. Two or three on the specification, in plain language. One on the decoration and how it will wear. One on care. Then the practical detail — sizing, delivery, what happens if it is wrong.
That is a description a person can act on, and it is also the structure search engines can parse into a rich result, because every claim maps to a property rather than a mood.
The honest limit
AI is genuinely good at turning your specification into readable prose, at producing forty variations for forty SKUs, and at maintaining a consistent voice across a catalogue. It is bad at knowing anything about your product that you did not tell it. If you do not know the fabric weight of the garment you are selling, no prompt will rescue the copy — ask your manufacturer, and put the real number in.
