Cross-Border Translation
Rewrite an English listing for DE/FR/IT/ES/NL — not a word-for-word translation
What this tool does
Cross-Border Translation rewrites an English product listing for one of five European markets. You paste the listing, choose a target language and the platform you sell on, and it returns a translated title, a full description, bullet points, a meta description, local keywords, a short note on what it changed, and a self-assigned SEO score out of 100.
The difference from a plain translator is the instruction it runs under. The model is told not to work word for word: convert inches to centimetres and US sizes to EU sizes, state prices in euros, keep the formal address a German, French or Italian shopper expects from a store, and fold local search terms into the copy while leaving your product specifications untouched. It is machine translation with a shopping context attached, which is a smaller claim than "AI localization" but closer to what actually happens.
How it works
- You paste the English listing. One box, everything in it: title, description, features, specs. The server rejects anything under 20 characters, so a bare product name is not enough to work with.
- You pick a market. German, French, Italian, Spanish or Dutch, and one of Amazon, Shopify, Etsy or eBay. The platform is passed to the model as context for register and length; it does not change which fields come back.
- You sign in. The translate endpoint checks for a session and returns 401 without one. Access is free and uncapped after that. Your email, the language, the platform and the first 100 characters of your listing go into the usage log at this point.
- One request goes to the model. Your listing, the language and the platform are packed into a single DeepSeek call at temperature 0.3, capped at 1500 output tokens with a 30-second timeout per attempt. On a timeout or a transient 429/5xx response the call is retried twice, so three attempts in total, waiting one second and then two seconds between them. The prompt carries six rules: translate naturally rather than literally, work in local search keywords, keep your emojis and formatting structure, convert measurements and currency and sizing, use the formal form (Sie for German, vous and Lei for French and Italian), and keep the specifications and prices identical.
- The JSON block is pulled out of the reply and rendered. If the answer comes back unparseable, you get a short placeholder result scored 60 rather than an error screen — three generic bullets, two keywords and no real translation. If a run looks like that, that is what happened, and it is worth running again. The translated text is not stored anywhere, so copy what you want before you leave the page.
What you get back
The prompt asks the model for a title, a description, five bullets, a meta description, five local keywords, a cultural note and a score. Nothing in the code checks the reply against that list, so the counts are a request rather than a guarantee, and the placeholder result used when the reply will not parse carries only three bullets and two keywords. The run below is written out as an illustration of the shape and register of the output. It is not a real user's result.
The prompt asks for a title under 150 characters and a meta description under 160. Those are requests to the model, not limits the code enforces, so count them before you publish.
Where it helps and where it does not
It helps most when the English listing already works and you want a version for a DACH, French, Italian, Spanish or Dutch store without commissioning a translation from scratch. The realistic workflow is to generate the draft here and have a native speaker edit it, so the paid work is an edit pass rather than a translation from nothing.
Here is what it will not do:
- Five languages, one direction. German, French, Italian, Spanish and Dutch, from English. No Polish, Portuguese, Swedish, Japanese or Chinese, and no way back into English.
- One listing per run. No CSV import, no bulk queue, no Shopify or Amazon connection. You paste in and copy out.
- No currency lookup. The prompt tells the model to state prices in euros. Nothing in the code calls an exchange rate feed, so every converted price and every converted measurement is a language model's arithmetic. Check the numbers.
- The keywords are not researched. They come from training data, not from search volume in that country. Plausible starting terms, nothing more.
- The score grades itself. The model that wrote the copy assigns the score in the same response. It has no view of the listings you are competing against.
- Long listings get cut. The reply is capped at 1500 tokens. Paste in a 2,000-word description and expect the end to be missing.
- It knows nothing about your legal obligations. EU marketplaces carry requirements around labelling, safety marking, ingredient disclosure and distance-selling text. None of that is checked here. If you sell cosmetics, food, electronics or anything with a compliance claim, that wording is your job.
The output also has no memory between runs, so translating twenty product variants gives you twenty independently worded listings rather than a consistent catalogue. Keep a glossary of your own brand and product terms and apply it during the edit pass.
What to do before and after this step
Translation is close to the last step, not the first. Get the English source right with Perfect Listing before you localize it, because a weak listing translates into a weak listing. Decide the product deserves a second market at all by running it through the Product Validator, since demand in the US says little about demand in Italy. And if the plan is to sell into DACH or France properly, delivery times decide whether the listing converts, so check what Supplier Finder turns up for local and EU-based supply before you commit to the market.
Frequently asked questions
Is the listing translator free?
Yes. There is no card field anywhere in the flow and no daily cap in the code once you are signed in. Each run costs us one model call, which we pay for.
Do I need an account to translate a listing?
Yes. The translate request returns a 401 without a session, so the sign-in box appears before anything is sent. Your email, the target language and the platform you picked are written to the usage log, along with the first 100 characters of your listing. The translated text itself is not stored, so copy it before you close the tab.
Which languages and platforms are supported?
Five target languages: German, French, Italian, Spanish and Dutch. Four platform presets: Amazon, Shopify, Etsy and eBay. The source listing is assumed to be English. There is no option for Polish, Portuguese, Swedish or any non-European language, and no reverse direction back into English.
How is this different from Google Translate or DeepL?
Those return your sentences in another language. This one runs the listing through a language model with instructions to rewrite rather than translate: convert inches to centimetres and US sizes to EU sizes, state prices in euros, keep the formal register that German, French and Italian shoppers expect from a store, and work local search terms into the title and description. You also get keywords, a meta description and a note on what changed, which a translator does not produce.
How reliable are the local keywords and the SEO score?
Treat both as suggestions. The keywords come from the model's training data, not from a live search-volume tool for that country, so check them in a real keyword tool before you build a strategy on them. The score is the same model grading its own output in the same response. It is useful for comparing two of your own drafts and useless as a prediction of where you will rank on Amazon.de.
Can I publish the output without a native speaker reading it?
For a test listing, probably. For a store you are spending ad budget on, no. Machine output in German, French and Italian tends to be grammatical but slightly off in register and product terminology, and shoppers notice. The realistic use is as a first draft for a native proofreader to correct, rather than a translation written from scratch. Check every converted number yourself, since nothing here looks up an exchange rate.