Writing product descriptions is one of the first jobs store owners hand to AI, and for good reason. A shop with 500 items and thin, copied or missing descriptions can have usable text for all of them in days rather than months. The risk is equally clear. Language models are very good at writing confident, pleasant sentences, and they will happily write that a shirt is “100% organic cotton” or a pan is “dishwasher safe” when nobody told them so. A description that invents facts is worse than no description, because it leads to returns, complaints and, in some cases, legal trouble. The goal is speed with honesty.
Give the model facts, not freedom
The single most important rule: the model should only describe what you have told it. It should not guess. That means the input for each product is a set of facts, not just a product name and a photo.
For each item, prepare a short fact sheet:
- Product name and category.
- Materials or ingredients, exactly as on the label or supplier sheet.
- Measurements, weight, sizes available.
- Colour names.
- Care or usage instructions, if known.
- Origin, if known and relevant (for example, a tea region).
- Anything you are sure of that makes it stand out.
Then instruct the model clearly: write only from these facts; if something is not in the facts, do not mention it; do not add health, safety or environmental claims. A well-written instruction like this cuts invention sharply, though not to zero, which is why the next step matters.
Check every description against the facts
Even with good instructions, some invented details will slip through. A person must read each description before it goes live. That sounds like it defeats the purpose, but reading and checking is much faster than writing from scratch.
Take a worked example. Say writing one description by hand takes 15 minutes, and checking an AI draft against its fact sheet takes three. For 500 products, that is roughly 125 hours of writing against 25 hours of checking, plus some hours to prepare the fact sheets. The saving is real, but it only holds if the checking is actually done.
Make the check specific. The reviewer looks for:
- Any material, size or feature not on the fact sheet.
- Claims such as “hypoallergenic”, “waterproof”, “eco-friendly” or “best quality” that you cannot prove.
- Wrong units or rounded measurements.
- Words that do not fit your shop’s voice.
A second AI pass can help here: ask a model to list every factual claim in the description and mark which ones appear in the fact sheet. Anything unmatched goes to the human reviewer. This is a helper, not a replacement for the person.
Keep a consistent structure
Descriptions are read by three audiences: shoppers skimming on their phones, search engines, and increasingly AI shopping assistants. All three prefer structure.
A simple pattern works well for most shops:
- One or two sentences on what it is and who it suits.
- A short list of key facts: material, size, weight, care.
- One sentence on delivery or returns if it differs from your standard terms.
Resist long, flowery paragraphs. “Crafted with love for your special everyday moments” tells the buyer nothing. “Handwoven cotton, 180 cm by 90 cm, machine wash cold” tells them what they need.
Languages and local buyers
If you sell to customers who prefer Sinhala or Tamil, AI can produce translated descriptions as well. The same honesty rule applies, and a fluent speaker should check each one, because models are less reliable in these languages than in English and can produce stiff or wrong phrasing. Product terms, especially for textiles, spices and local foods, often have everyday names that a model may not know. Keep a small glossary of the words your customers actually use and include it in the instructions.
Where this approach falls short
AI descriptions work best for products with clear, listable facts. They are weaker for items where the selling point is feel, craft or story, such as handmade jewellery or art. A model can arrange the facts, but the owner’s voice and knowledge are what make those pages convincing. For those, use AI for a first draft of the structure and write the heart of it yourself.
They also do not fix missing data. If you do not know what a product is made of, the model does not either, and it should say nothing rather than guess. Sometimes the real work is asking your supplier for a proper specification sheet.
Finally, avoid publishing hundreds of near-identical descriptions for very similar items. Vary them by the facts that genuinely differ, and keep shared text short.
A sensible way to start
Pick 20 products from one category. Build fact sheets for them, write clear instructions, generate the descriptions, and time how long the checking takes. Count how many invented details the reviewer caught. That small trial tells you whether the process works for your catalogue and how much checking to budget for. If you want a starting point, the product description tool drafts from facts you supply, and every draft still needs your eyes before it goes on the site.
