A store with two hundred products has two hundred small writing problems, and they all arrive on the same afternoon. The obvious answer is to hand them to a language model. That works for about a week — until you notice every page opens on the same cadence, praises the same craftsmanship, and closes on the same promise about a perfect gift. The bottleneck was never typing speed. It was the brief. What follows is a workflow for writing product descriptions with AI at a catalogue size where writing each one by hand is not realistic, without ending up with two hundred pages that could belong to any store.
Why AI product descriptions all sound alike
AI product descriptions read alike because a language model returns an average of what it has read. Ask it to describe a ceramic mug and you get the average ceramic mug page: elevate your morning ritual, crafted with care, a nod to durability, a closing line about gifting. No single sentence is wrong. The problem is that your competitor received it too, from the same model, five minutes earlier.
Only specificity breaks the average, and specificity has to come from you. There are three things the model cannot know and will therefore invent or omit:
- What your customers ask before buying. The question your inbox receives every week — will it fit a standard shelf, does the colour match the photo, how long before it arrives. That question belongs in the description, not in your replies.
- What the product does that the spec sheet does not say. The handle stays cool. The zip is the part that fails on cheaper versions. The scent fades by lunchtime, which is the point.
- How your store sounds. Not “our brand voice” as an abstraction — the actual register you use with customers, and the words you would never write.
Everything below is a way of getting those three into the prompt reliably, at volume, without writing an essay per product.
The brief: five inputs that change the output
Before generating anything, assemble a short input block for each product. Five fields, one line each. This is the entire difference between usable copy and filler.
- The buyer and the moment. Not a persona document — a sentence. “Buying a gift, in a hurry, does not know the brand.” A page written for a hurried gift-buyer looks nothing like one written for a returning enthusiast.
- The single reason to choose this one. One reason, chosen deliberately. A description that argues four advantages argues none.
- Three concrete facts. Material, dimension, capacity, origin, timing. Nouns and numbers, never adjectives. These are what stop the model from inventing.
- The objection. The reason someone puts the item back. Price, sizing doubt, a comparison with a cheaper version. Naming it in the copy is what makes a page feel written by someone who has actually sold the thing.
- The constraint. Word count, reading level, banned words, and above all the claims you are not allowed to make — anything about health, safety, origin, or performance that you cannot evidence.
The fifth field is the one most people skip, and the one that saves the most editing. A model asked for enthusiasm will happily produce a medical claim.
AI product descriptions: a workflow that survives a real catalogue
Write the voice document once
One page, reused in every prompt: three sentences of description, five words you use, five you refuse, and two short samples of copy you already consider good — one product page, one email. Samples do more work than adjectives. “Confident but not loud” means little to a model; a paragraph you actually published means everything.
Fill a spec sheet, not a prompt
Put the five fields in a spreadsheet, one row per product. Filling forty rows takes an afternoon and is genuinely dull, which is why most catalogues never get good copy. It is also the only part that cannot be delegated to the model, because it is the only part that contains information the model does not have.
Generate in batches of ten
Ten rows per request, with the voice document at the top. Batching does two things: it keeps you reading output critically instead of accepting it, and it exposes repetition immediately. If three of the ten open with the same construction, you can see it on one screen and fix the instruction rather than the copy.
Edit for the thing the model cannot know
Expect to change one sentence in four. Usually it is a fact softened into a claim, a comparison you would not make about a competitor, or a piece of confidence the product has not earned. This edit pass is not a failure of the method — it is the method. Copy that leaves without a human read is how a catalogue acquires two hundred small inaccuracies.
If you would rather not build the prompt scaffolding yourself, The E-Commerce Prompt Pack contains this logic pre-written — description frameworks by angle, titles, bullets, comparison tables and FAQ blocks, each with the variables marked so you fill in what only you know. It is the same structure described above, already argued out.
Structure AI product descriptions for two readers
Every product page is read twice: by a person scanning on a phone, and by a crawler deciding what the page is about. One layout serves both.
- The first line carries the weight. It should say what the thing is and who it suits, in plain words. Most shoppers read nothing else.
- Three or four bullets, facts only. This is where the numbers live. Bullets are what a scanning reader actually consumes.
- One paragraph handling the objection. The honest one, not a rhetorical one.
- A short specification block. Repetitive, unglamorous, and the part returning customers use.
Keep the structured data on the page honest — price, availability and title must match what the visitor sees. Search engines read the markup, and a mismatch between markup and page is one of the few technical errors that reliably causes trouble.
The checks before any AI product descriptions go live
Six checks, applied to every batch of AI product descriptions. They take a few minutes and they are the reason this scales without incident.
- Invented facts. Anything the model asserted that was not in your three concrete facts is a claim you now own.
- Regulated claims. Health, safety, environmental and origin statements need evidence, and in several markets they need specific wording. If in doubt, remove it.
- Sameness sweep. Read the opening sentences of twenty descriptions in one column. Repetition is invisible page by page and obvious in a list.
- Keyword stuffing. If the product name appears in every sentence, the model was optimising, not writing.
- Structured data match. Price, stock and title consistent between markup and page.
- One read aloud. The oldest test there is, and still the one that catches the sentence no rule would.
Where to start with AI product descriptions if the catalogue already exists
Do not rewrite every product at once: AI product descriptions are worth the most where the traffic already is. Sort your products by revenue, take the twenty that carry the store, and rewrite only those. Leave the rest untouched for a full season so you have something to compare against — a catalogue rewritten all at once teaches you nothing about whether the rewriting worked.
Manufacturer-supplied descriptions deserve particular attention: identical text sitting on dozens of retailer pages gives a search engine no reason to prefer yours. Those pages are usually where the first rewrite pays for itself.
FAQ
Will Google penalise AI-written product descriptions?
Google’s published guidance on AI-generated content is about quality and intent rather than production method: what its spam policies target is content mass-produced to manipulate rankings. A description that answers a real buyer’s question is fine regardless of how it was drafted. A thousand near-identical pages generated to occupy search results is the thing that gets penalised — and it was penalised long before language models existed.
How long should a product description be?
Long enough to answer the buying question and stop. For a simple, familiar item, forty words and a spec block are plenty. For an expensive or unfamiliar one, two hundred words earn their place because the objection needs handling. Length has no independent value; the model will happily give you six hundred words that say the same thing four times.
Can I use the same copy on my store, Etsy and Amazon?
You can, but the buyer’s state of mind differs on each and so should the emphasis. On a marketplace the visitor is comparing several near-identical listings, so the differentiator belongs in the first line. On your own store they arrived with some intent already, so context and objection-handling matter more. The spec sheet stays identical; the angle changes. That is one extra line in the prompt, not a rewrite.
Go further
Once the pages are written, the same specificity problem reappears in your advertising — and there it costs money per impression. The E-Commerce Ad Creative Pack gives you thirty studio-style backgrounds built around an empty product zone, so one cutout of your own product becomes thirty testable ads for Meta and TikTok. If your work is closer to writing than to design, the rest of our prompt packs follow the same principle as this article: the strategy sits inside the prompt, and you supply the details only you have.