Contents7
In short
A product page is translated field by field: title, description, attributes, instructions. Value lists (materials, colours, sizes) are translated only once.
Safety information must be readable in the language of the country of sale, and the page must speak the language of the feed sent to Google.
The machine translates in batches, a glossary and automated checks frame the result, and human review goes where the page commits you and where it sells.
A catalogue of three thousand product pages cannot be translated by hand at a per-page rate: neither the budget nor the deadline holds. Nor can it be published unchecked: a wrong dimension or badly rendered safety information is paid for in returns, disputes, even a product withdrawal. Between the two lies a method that handles the volume with the machine and places human review where it matters.
This article describes it. It extends our articles on website translation and opening new markets, and multilingual and multi-currency store: where to start for the order of work on an online store.
Translate fields, not pages
A product page is not a single block of text. It is an assembly of fields, and each is handled in its own way:
- The title: short, packed with words buyers search for, limited in length by the platforms (150 characters in Merchant Center, for example). It is translated with the glossary, then reviewed.
- The description: the only truly written field. This is where machine translation saves the most time, and where tone matters.
- The attributes: material, colour, size, weight. These are values drawn from closed lists: each value is translated once, then applies to every page that uses it.
- Instructions and statements: safety, care, composition. They commit you, and are reviewed in full.
- The hidden fields: image alt text, title tag, meta description, structured data. They show in search results and are translated like the rest.
The gain from closed lists is considerable: a catalogue has far fewer distinct attribute values than product pages. Translated and reviewed once, they are right everywhere.
Before translating, check that the source page stands up: a vague page in English will be vague in every language. The product page that converts sets out what buyers need to find on it, in every language.
What the law and the platforms require
Two rules frame the language of a product page sold abroad.
Safety information. The EU General Product Safety Regulation, applicable since 13 December 2024, requires the online offer of a product to show its warnings and safety information in the language of the country where it is sold. Its Article 19 puts it this way:
Any warning or safety information […] in a language which can be easily understood by consumers, as determined by the Member State in which the product is made available on the market.
The language of the feed and the page. For Google Shopping, Merchant Center help on languages and currencies asks for landing pages in the same language as the product data, and warns that products may be disapproved if most of their data does not match the data source language. A translated feed that points to a page still in French does not pass.
One more detail for titles generated by AI: the title specification in Merchant Center offers an optional attribute, the structured title, to state that a title was created using generative AI. In our reading, a title translated by a generative AI model falls into that case: the question is worth settling before sending the feed.
Translating at volume: batches, glossary, checks
Translating a catalogue is a bulk job, and it is organised as one:
The path of a translated catalogue
- Export
The fields to translate leave the catalogue, one row per product.
- Translation
In batches, with the glossary and the memory.
- Checks
Figures, units, terms and lengths verified by code.
- Review
Templates, safety information, best sellers, sample.
- Import
The translated fields join each language’s product pages.
Batches cost less. AI model providers offer batch processing, slower but cheaper: Google announces a 50% reduction for the Gemini Batch API, and OpenAI lists a Batch rate at half its standard rate, for GPT-5 mini for example. A catalogue does not need an answer within the second. As for the choice of tool, it is made on your own product pages: see DeepL, Google or an LLM: which translates your trade best.
The glossary and the memory make it consistent. The brand glossary fixes range names and trade terms, the translation memory avoids retranslating a sentence that comes back from one product to the next.
Automated checks catch what the eye misses. Across thousands of pages, human review does not see everything; code does. The useful checks:
- numbers identical between source and translation (dimensions, weights, quantities, references);
- units present and, if the market requires it, converted;
- required terms present and banned terms absent;
- lengths respected, title included;
- no empty field, no sentence left in the source language.
Watch out
A translation must not add anything
A wrong figure is not a style mistake: a wrong size means a return, a wrong dosage means an accident. An AI model can also add what was not in the source. A client learned this the hard way on another kind of text: on a real-estate portal, a model tasked with writing published three property developments that did not exist, since unpublished. Check that the translation contains neither more nor less than the original.
Where to place human review
Reviewing three thousand pages in every language is not realistic; reviewing none is not defensible. Review is placed according to what an error costs:
- Templates and value lists, in full.
An error there is repeated on every page.
- Safety information and regulatory statements, in full.
They engage your liability.
- Best sellers, in full.
In many catalogues, a small share of the pages makes most of the revenue: check it in your sales data, these are the pages buyers read.
- A sample of each batch, in each language.
Record the share of corrections: if it rises, review more.
- Signals after publication.
Questions to customer service, returns for “item not as described”: a rise in one language points to pages to revisit.
This review is the work of a translator or a reviewer who reads the language, and it cannot be replaced. What changes with the machine is that they spend their time on what matters instead of translating lists of colours. We detail that split for a whole website in translating your website with AI, automatic or reviewed.
From the field
On Palmora Property, a real-estate listings site in six languages that we operate, 1,726 property pages were rewritten by AI in 36 batches, with a retry on error: a single failure in the whole run. The translation pipeline that takes over then brings 95% of segments out of the cache after the first pass, without calling the AI again. These figures hold for that catalogue.
What it costs
The machine’s cost, on a catalogue, stays low: from under a dollar to about twenty dollars per million characters depending on the tool, according to the public price lists we compare in website translation cost: per word, per page, per market. What costs is the review, and it is managed with the priorities above. Formats, currencies and units belong to localization: see localize, don’t just translate.
Where to start
Export your hundred best sellers, translate them into the language of your first market with a glossary written beforehand, and have them reviewed in full. You will have a validated template, a first memory, and a measure of the tool’s quality on your own product pages. The rest of the catalogue follows in batches.
To translate a full catalogue with its checks, describe your catalogue and your markets: we start from your best sellers and your safety information.
Frequently asked questions
Should the whole catalogue be translated at once?
No. Start with the pages that sell and those that carry safety information, review them in full, then translate the rest in batches, with automated checks and a reviewed sample. A product page missing in one language does less harm than a wrong one.
Is it allowed to translate product pages with AI?
Nothing prohibits it, but responsibility for the content remains yours. Safety information must be readable in the language of the country of sale, and an error there engages the seller, whatever the tool. Hence the full review of those statements, and a check of everything that contains a figure.
Should sizes and units be converted?
When the market does not read them otherwise, yes: a clothing size or a unit of measurement that means nothing to the buyer causes hesitation, then a return. Conversion is done with verified tables, not by translation, and is checked like any figure.
How many product pages should be reviewed?
All those that commit you (instructions, statements) and those that sell the most, in full; for the rest, a sample per batch and per language. If the share of corrections rises in a sample, widen the review: it signals a glossary or template problem.
Sources
- Regulation (EU) 2023/988 on general product safety (new tab)eur-lex.europa.eu
- Google Merchant Center Help, supported languages and currencies (new tab)support.google.com
- Google Merchant Center Help, title and structured title (new tab)support.google.com
- Google, Gemini API pricing (new tab)ai.google.dev
- OpenAI, API pricing (new tab)developers.openai.com
Google Merchant Center Help, supported languages and currencies
Read on support.google.com(new tab)Jim · founder of Brain Plus AI
Also in translation and new markets
Choose and price7 min
Translating your website with AI: automatic or reviewed?
Having everything reviewed is expensive, and reviewing nothing lets errors through in languages that nobody at your company reads. The split is decided page by page, depending on what each page commits you to.
Practical guide8 min
Brand glossary: keeping translations consistent
The same product called three different things on three pages, and readers wonder whether there are three products. A brand glossary fixes the terms that are not up for debate, before the first translation, whatever the tool or the translator.
Understand8 min
Localize, don't just translate: prices, proof and forms
A page can be perfectly translated and still feel foreign: a price in the wrong currency, a date read backwards, a form that rejects a name, a meeting with no time zone. Localizing means adapting what changes from one country to another, not just the words.