Descriptions for 500 products in a week: how to do it with an LLM without ending up with garbage

A catalog of 500 products with no descriptions is a common situation. A copywriter charging per product listing adds up to thousands of dollars and months of waiting. Writing them yourself never actually happens.

An LLM closes that gap in a week for a few dollars of API cost. But only if you do it right: the naive "generate a product description" approach produces text that is worse than having none at all.

Why the naive approach fails

Ask the model to write a description for "Glass vase, 25 cm" and you will get a paragraph about how the vase "will be the perfect centerpiece for your home and reflect your refined taste." That kind of text:

• contains no fact that is not already in the title;
• reads identically across all 500 products, give or take a synonym;
• does not help anyone decide, so it does not help conversion;
• looks exactly like the kind of content search engines devalue.

The problem is not the model. The problem is nobody gave it any facts.

Step 1. Prepare your data — 70% of the work

Export your catalog to a spreadsheet and make sure every product has at least this:

• an exact name and category;
• material, dimensions, volume, weight;
• color and variants;
• manufacturer or country of origin;
• notable features: dishwasher-safe, microwave-safe, handmade.

If the specs do not exist, there is nothing to generate from, and no model can fix that. Filling in the specs is the real work; generation itself takes hours afterward.

A nice side effect: filled-in specs also improve your filters and your Merchant Center feed.

Step 2. A prompt with real constraints

A good catalog prompt has four parts:

1. Role and context: what kind of store this is, who it is for, in what language and tone.
2. A strict structure: first paragraph — what it is and what it is for; second — the specs in prose; third — practical advice on choosing or caring for it.
3. Bans: no "wide range," no "great prices," no "our store," no invented facts, no adjectives that are not backed by data.
4. Product data — every spec from your spreadsheet.

The core rule: the model is not allowed to add facts. Only to rephrase what you gave it. That is what separates a useful description from a hallucination you will be the one answering for.

Step 3. Pilot on 20 products

Do not run the whole catalog at once. Take 20 different products from different categories, generate them, and read all 20 back to back.

Look for two things: whether they all follow the same sentence structure, and whether any facts were invented. If every description opens the same way, fix the prompt, not the texts.

Step 4. Generation

In batches of 50–100 products, saving results back into the same spreadsheet. Cost-wise this is a few dollars, maybe a few tens of dollars for the whole catalog — orders of magnitude cheaper than a copywriter.

Step 5. Proofreading — non-negotiable

Reading 500 descriptions is a single working day. You cannot skip this step, because it is where you catch:

• specs mixed up between similar products;
• facts that were not in the source data;
• awkward phrasing, especially in translated names;
• repetition that is only visible once you see them all together.

The rule is simple: a human reads everything that goes on the site. Not because the model is bad, but because the text on your domain is your responsibility.

Where LLMs actually break

• Dimensions and fit. The model will happily write "fits a standard shelf" — with no idea what your shelf looks like.
• Care instructions. "Dishwasher safe" on a product that is not is already a customer complaint waiting to happen.
• Local terminology in niche categories. Check what your actual customers call the thing.
• One-of-a-kind handmade items. Where the value is in the story, generation does not work — write those yourself.

Will Google penalize AI-written text

Google evaluates how useful a text is, not how it was made. A description with real specs that helps someone decide works. A description that pads out the title with fluff does not — and would not have worked even if a human wrote it.

The risk is not that the text was generated. The risk is that it is empty.

What this costs, all in

Stage	Time	Money
Filling in specs	2–4 days	your time or an assistant’s
Prompt and pilot	a few hours	pocket change
Generating 500 descriptions	hours	a few dollars to a few tens of dollars
Proofreading	1 day	your time

Against thousands of dollars and months for the traditional approach.