AI-Translated Product Pages: When Sizes, Care, and Safety Wording Go Wrong
Machine translation of a catalogue is fast and mostly fine, until a size converts wrong, a care instruction flips meaning, or a safety warning softens. Auditing translated pages where errors cost money.
Translating a full catalogue used to take a season. Now a store can launch in a new language over a weekend, and most of the copy is genuinely fine, which is exactly the problem. The errors hide inside fluent text: a chest measurement converted twice, a care instruction that now says the opposite, a warning that reads as a suggestion in the new language.
Nobody on the team reads the new language well enough to catch those, by definition, or they would not have needed the tool. So translated pages need a different kind of audit, one that checks the things that can hurt a customer or a margin first, and leaves the poetry for later.
Numbers first, because numbers travel badly
Audit measurements, sizes, weights, capacities, temperatures, and quantities before any prose. Unit conversion is where automated translation and catalogue tooling combine into real errors: inches converted and then labelled as centimetres without conversion, shoe sizes mapped to the wrong regional table, a 500 ml bottle becoming 500 litres in one field. These are checkable without reading the language, by comparing the numeric values on the translated page against the source and against the physical product data.
Size names need the same check. Small, medium, and large do not map cleanly across markets, and an untranslated size code next to a translated description can leave a shopper guessing which garment they are buying. Returns in a new market are expensive in a way domestic returns are not, and a size error is a return with a tracking number across a border.
Care and safety wording is not marketing copy
Do not bleach becoming a mild suggestion, wash at 40 becoming wash at 140, keep away from children losing its force: these are the mistranslations with consequences beyond a return. Any product with safety, dosage, age, or material warning content needs those specific strings reviewed by a human translator, even if the rest of the catalogue is machine translated and post edited lightly.
Make a list of the strings that matter, per product type: warnings, care, ingredients and allergens, age restrictions, assembly safety. That list is short compared with a catalogue, and paying for human review of exactly those strings is affordable in a way full human translation was not. It is also defensible in a way we translated everything automatically is not, if something goes wrong.
Let customers and data tell you the rest
Watch the new markets search terms, returns reasons, and support contacts for translation shaped problems: questions asking what a specification means, returns citing size, reviews complaining an instruction made no sense. Those signals are your audit queue, ordered by how many customers each error has already reached.
Fix the source pattern, not just the page. A recurring conversion error usually lives in a feed field or a translation memory entry, and it will keep regenerating into new products until it is fixed there. Stores that patch pages one at a time find the same wrong number on next seasons catalogue, wearing a new product name.
Related reading
If this topic is useful, these earlier guides go deeper on connected questions: AI Review Summaries Smooth Away the Patterns You Actually Need, AI Product Descriptions Get the Specs Wrong: Auditing Generated Copy Before It Costs Returns, and AI Recommendations on Out of Stock Pages: Helpful Detour or Dead End.
Frequently asked questions
What should be checked first on AI translated product pages?
Numbers: measurements, sizes and their regional conversions, weights, capacities, temperatures, and quantities. Compare them against source data and the physical product, because numeric errors cause returns and are checkable without fluency in the language.
Which translated content needs a human reviewer?
Safety warnings, care instructions, ingredients and allergens, age restrictions, and dosage or assembly information. Those strings are short in total, and human review of them is affordable compared with translating a whole catalogue by hand.
How do we find translation errors we cannot read?
Use market signals: support questions about specifications, returns citing size or instructions, and search terms in the new language. They surface the errors customers have already hit, in priority order.
Why fix translation memories and feed fields?
Because page level fixes get overwritten. A wrong conversion or phrase stored in a translation memory or feed field regenerates into every new product until the source itself is corrected.