Writing OTTO product copy with ChatGPT: workflow, prompt and checks

What you actually write on OTTO

OTTO assembles the product title on otto.de from your data; the OTTO Market API's category interface shows, per category group, how it is built. What you write yourself is mainly:

• Description – "relevant for SEO" according to the API, should stay under 2,000 characters, simple HTML allowed.
• Bullet points – 3 to 180 characters each.
• Product line – the product's proper name, at most 70 characters, identical for all variations.

Attributes aren't copy, they are data. They drive the filters and, according to reports, OTTO's AI assistant too. Background in Writing product descriptions for OTTO and Optimising product titles for otto.de.

The workflow in five steps

1. Collect the facts: export the maintained attributes, scope of delivery and care instructions from your system. That is the only source for the copy.
2. Set the search terms: the main term plus two or three synonyms shoppers use – see Semantic search on otto.de.
3. Generate a draft with the prompt template below.
4. Check: fact check and legal check, line by line.
5. Upload and measure: note the date, compare positions.

Work product by product, not a hundred at once. Otherwise a mistake the AI makes once is repeated across the whole catalogue.

A prompt template for a generic product

The template forces the model to stick to your facts and returns a list of the facts it used, so you can check them. Replace the brackets with your data, for example for men's running shoes. Since the copy on otto.de is German, the prompt asks for German output:

You write product copy for the OTTO marketplace in German, formal "Sie". Use only the facts below. Do not invent dimensions, materials, certificates, test results or features. If information is missing, write [MISSING] instead of filling it in. Do not mention other brands. No unsupported superlatives, no marketing filler. Deliver: 1) a description under 2,000 characters in short paragraphs, 2) five bullet points of at most 180 characters each, one statement per point, 3) a list of the facts you used. Target group and use: [e.g. beginners, jogging on tarmac]. Terms shoppers search for: [main term, 2–3 synonyms]. Facts: [attributes from your data set, scope of delivery, care instructions].

If useful, ask for two versions and pick the more factual one. Never paste competitors' copy into the prompt.

Fact check: what AI likes to invent

Language models write fluently even when information is missing. Typical errors in product copy:

• Dimensions and weights: "only 280 g" with no basis, or centimetres instead of millimetres.
• Materials: "breathable mesh" or "genuine leather" although your data says nothing about it.
• Features: "waterproof", "dishwasher-safe", "suitable for allergy sufferers".
• Labels and tests: certificates or test results that don't exist.
• Scope of delivery: accessories that aren't in the box.

Check every number and feature against your attributes. Anything not backed up goes. If the copy contradicts an attribute, the attribute wins – and the copy is corrected.

Legal pitfalls

You offer the product in your own name and are therefore responsible for the copy. In the "Himalaya Salz" case (I ZR 86/13, 2016), Germany's Federal Court of Justice ruled that an online retailer is liable for its offer even when a supplier entered the details; the Händlerbund concludes that incorrect product information must be removed. By the same logic, AI-written copy is no different.

• No untrue or misleading statements about essential features – that is misleading advertising under Section 5 of the German Unfair Competition Act (UWG). Invented dimensions, materials or features fall under it.
• No factual claims you can't prove, such as "test winner" or "lasts ten years".
• No other brands in the copy, for example as a comparison – that can cause trademark problems.

If in doubt, have wording checked by a lawyer. This article is not legal advice.

Attributes remain the source of truth

Good copy doesn't replace a missing attribute. When a shopper filters by material or size, the attribute value counts, not the sentence in the description. And according to reports, AI assistants also draw on structured attributes – see OTTO in ChatGPT and the AI assistant.

So use AI the other way round as well: have it draft a list of possible attribute values from a manufacturer's data sheet – and check every value yourself before entering it.

Measuring the effect on rankings

Whether new copy makes a difference shows in the organic position, not in gut feeling. Before uploading, record the positions for the main term and the synonyms, change only the copy and compare over several weeks. MarktRank stores the positions per keyword at every check – on demand and automatically each week –, so you see before and after side by side. The method: Measuring rankings after listing changes.

Frequently asked questions