In our test of 15 listings on 10 October 2026, ChatGPT wrote fluent, persuasive copy but added claims the product data did not support in 14 of them, and Nordova’s validator found rule errors in 13. Nordova added unsupported claims in at most 1 and had no validator errors; two blind reviews preferred Nordova’s listing in 13 of 15 tasks. Here is how we tested, what ChatGPT does well, where it fails and how to use it safely.
How we tested
We gave both tools the same 15 tasks: 5 products (a Bluetooth speaker, a handmade ceramic mug, a children’s striped long-sleeve shirt, a solid oak coffee table and a rosehip face oil) for Amazon.de, eBay.de, Etsy, OTTO, Kaufland and an own shop, in German, English, French, Italian, Spanish and Polish.
- ChatGPT: OpenAI’s ChatGPT model, one request per listing. The prompt asked for a listing optimised for the marketplace, using only the product information given, returned as separate fields.
- Nordova: the production pipeline, with the same product data.
- Two independent blind reviews by AI evaluators saw the listings in random order without knowing which tool wrote which. They listed every claim not supported by the product data (a direct consequence of a stated fact or a typical use did not count), noted rule breaches, scored persuasion from 1 to 10 and chose the listing they would rather publish.
- Marketplace rules were checked with Nordova’s own validator. Nordova revises its listings against that validator, so its zero there is expected — read that number as “ChatGPT’s output does not pass these checks”, not as an independent score.
A second test on the same day gave both tools 8 products whose supplier data already contained banned wording, such as “nachhaltig”, “klimaneutral” or “CE-geprüft”. Fifteen and eight tasks are small samples; results vary with products and prompts.
The results
| Measure | ChatGPT · Nordova |
|---|---|
| Listings with unsupported claims (review 1 / 2) | 14 / 14 of 15 · 0 / 1 of 15 |
| Unsupported claims counted in total (review 1 / 2) | 58 / 54 · 0 / 1 |
| Listing the reviewer would rather publish (review 1 / 2) | 1 / 1 · 13 / 13 |
| Persuasion, mean score (review 1 / 2) | 6.27 / 6.53 · 7.07 / 6.93 |
| Validator errors | 26 errors in 13 of 15 listings · 0 |
| Repeated banned wording (second test) | 7 of 8 listings · 0 of 8 |
The remaining preferred task in each review went to an earlier Nordova version, which was also in the comparison. Of ChatGPT’s 26 validator errors, 16 concerned length, format and keyword rules, 8 were claims not backed by the data and 2 were superlatives.
What ChatGPT does well
ChatGPT’s copy reads well in all six languages, and it carried the hard facts — 4,000 mAh, IPX6, 350 ml — correctly into the listings. It writes persuasively: in a first review against the earlier Nordova version, the reviewer preferred ChatGPT’s copy in 10 of 15 tasks, partly because of the added claims. That result is why we improved our own copy, and why we only claim the points measured above, not “better writing” in general. In one task in each review, the coffee table for an own shop in French, the reviewers preferred ChatGPT’s listing.
Where ChatGPT fails
Invented properties. Even with “use only the product information” in the prompt, ChatGPT filled gaps with plausible details:
| ChatGPT wrote | The product data said |
|---|---|
| „Robustes Gehäuse aus ABS-Kunststoff“ | housing made of ABS — nothing about robustness |
| « Étanchéité IPX6 » | IPX6: protected against water jets, not waterproof |
| “Lavable a máquina a 40 °C” | “Waschen bei 40 °C” — not machine washable |
| „speziell für die Gesichtspflege trockener Haut geeignet“ | “für trockene Haut” — no “specially suited” |
| “traditional wheel throwing techniques” | wheel-thrown stoneware — nothing about tradition |
Each of these can be a misleading statement about the product’s characteristics under § 5 UWG, and claims for cosmetics need evidence (Regulation (EU) No 655/2013).
Marketplace limits. Backend search terms that repeat title words on Amazon.de, an Amazon.de bullet of 276 characters (Nordova’s Amazon limit is 250), Markdown asterisks in shop descriptions meant to be pasted as plain text, a title over OTTO’s length limit, repeated Etsy tags.
Banned wording. Given supplier texts with “nachhaltig”, “klimaneutral”, “biodégradable” or “CE-geprüft”, ChatGPT repeated such wording in 7 of 8 listings. Generic environmental claims have been banned in Germany since 27 September 2026 (Directive (EU) 2024/825), and “CE-geprüft” presents a manufacturer’s own declaration as a test (§ 5 UWG) — see the green claims ban and CE and test claims.
When ChatGPT is enough
- You write an occasional listing and have the time to check every sentence against your data.
- The product has no regulated wording: no textiles, cosmetics, toys, electrical goods or environmental properties.
- The marketplace is lenient on structure, such as your own shop, and you count limits yourself.
- You want ideas — angles, synonyms, a first draft — not a listing to paste.
Nordova is built for the opposite case: many listings, several marketplaces and languages, and copy you want to paste without rereading every claim.
How to use ChatGPT safely
- Give it complete data, including what the product is not (“not waterproof”, “hand wash only”). Gaps are where it invents.
- Tell it to use only the given facts, and still expect additions: our prompt said exactly that.
- Read every adjective against your data: robust, high-quality, waterproof, machine washable, suitable for sensitive skin, ideal for … Delete what you cannot prove.
- Do not trust its own counts. Count the backend search terms with the byte counter and the title with the title checker.
- Fill the GPSR fields — manufacturer, EU responsible person, safety information — from your records, not from the model; the GPSR generator builds the block.
Then paste the result into the free listing check. It counts each marketplace’s limits and finds banned environmental claims, “CE-geprüft” and other wording rules in all six languages, without an account. What it cannot do is spot invented properties — that needs your product data, which only you, or Nordova when it writes the listing, can compare against.