AI translation gets the sentence right and the size wrong.
Translating a whole catalog with AI now takes an afternoon, and most of it will read fine. Trouble hides in a handful of words: a size, a washing temperature, a returns promise, your own product names. Know which ones, and send only those to a person.
AI translation of product pages and help content is good enough for most sentences, so don't review everything by hand. Review the words that cost money or trust when they're wrong: units, sizes and number formats; care, safety and ingredient information; promises about returns, warranties, delivery and price; and your own product and brand names. Keep a do-not-translate list and a short glossary of approved terms, route the risky fields to a fluent reviewer, sample the rest by category, and keep the source text and the version you published so you can trace any complaint.
Most of it will be fine. That's the problem.
A few years ago, translating a store into a second language meant a quote from an agency and a wait of weeks. Now you can run a whole catalog through an AI model in an afternoon, and if you read a random page, it'll probably sound natural. Better than natural, sometimes.
That's exactly why the mistakes get through. When nine sentences in ten read well, people stop reading carefully, and the tenth sentence is the one that says the jacket is machine washable at a temperature that shrinks it, or promises free returns in a country where you don't offer them.
Most of a translated product page can be good enough. A few words on it have to be right. The trick is knowing which words, and sending only those to a person.
This isn't an argument against AI translation. For a small store, it can be the difference between selling in one language and selling in four. It's an argument for spending your review time where it counts, instead of spreading it evenly across thousands of sentences that were already fine.
So here's a short glossary, not of translations but of the kinds of words we'd never publish without a human check.
Decision guide
What can be sampled, what needs a person
| Consideration | Sample it | A fluent person approves it |
|---|---|---|
| Product descriptions | Read a few per category per language. | Only for new categories or flagship products. |
| Sizes and measurements | Never just sampled. | Checked against your size chart for that market. |
| Care, safety and ingredients | Never just sampled. | Every item, exact wording approved. |
| Returns, warranty, delivery, "free" | Never just sampled. | Matched to the local policy page. |
| Playful brand copy | Easy to get flat or odd. | Written by a fluent person, not reviewed. |
Units, sizes and numbers
Measurements. Centimeters and inches, kilograms and pounds, milliliters and fluid ounces. A model may convert them, leave them alone, or do one in the title and the other in the description. Decide which you want for each market and check that it happened.
Sizes. Clothing and shoe sizes don't translate; they map, and the mapping depends on your brand's own size chart. "Medium" is not a universal size. Neither is a "9." If you sell apparel, the size chart is the most important page to get right, and it's often a table or an image the translation step skips entirely.
Number formats. In many countries a comma is a decimal point. "1,5 kg" and "1.5 kg" mean the same thing to different readers, and "1.500" can mean fifteen hundred or one and a half. Dates flip too: 03/04 is March or April depending on who's reading.
Prices. Currency symbols, where they go, and whether tax is included all vary. A translated page should never quietly convert a price. Prices should come from your pricing system for that market, not from a model's arithmetic.
Most of a translated product page can be good enough. A few words on it have to be right.
Care, safety and ingredients
Care instructions. Washing temperatures, "do not tumble dry," "dry clean only." These are short, easy to mistranslate, and the cost of getting them wrong is a ruined product and a refund.
Safety warnings. Age labels on toys, choking hazards, electrical warnings, "keep out of reach of children." In many markets some of this wording is regulated. Check the rules for the categories and countries you sell into, and have a fluent person approve the exact wording.
Ingredients and allergens. For food, cosmetics and supplements, an ingredient list isn't marketing copy. A model that "improves" it, merges two ingredients into one word or drops a "may contain" line has created a real risk. Treat these fields as data that a person signs off, not prose.
Promises: returns, warranty, delivery and "free"
Returns and refunds. "30-day free returns" has to match what you actually offer in that country. If the policy is different, the translation should say the local policy, not a faithful rendering of the English one.
Warranty and guarantees. Words like "guarantee," "lifetime," and "waterproof" carry legal and commercial weight. A model may choose a stronger word than you meant.
Delivery times. "Ships in 2 days" is a promise your fulfillment has to keep in that market. Translating it word for word can make a promise you can't keep.
"Free." Free shipping, free gift, free trial. Each one has conditions somewhere. Make sure the conditions came along with the word.
This is the same principle we wrote about for scoring catalog truth before you publish a claim: a confident sentence isn't the same as a true one.
Your own names
Product and collection names. Some should be translated, most shouldn't. A model left to itself will translate "The Harbor Coat" into the local word for harbor in one place and keep it in English in another.
Brand terms and features. Names of your loyalty program, your sizing system, or a fabric you've trademarked. Keep a do-not-translate list and give it to the model with every request.
Approved terms. A short glossary of how you say key things in each language: "add to bag" or "add to cart," formal or informal "you," the word you use for "returns." A model given that glossary is far more consistent than one asked to guess.
Idioms and jokes. Playful copy is where models are often weakest. If your brand voice leans on wordplay, a fluent person should write that line, not review it.
How the work moves
A translation workflow with a few checkpoints
- Send source plus glossary
Do-not-translate list and approved terms go with every request.
- Split by field
Risky fields go to a reviewer; the rest are sampled.
- Approve and publish
Store the source, the translation and the date.
- Listen for flags
Customer reports and support tickets by language.
- Re-run on change
Any source edit triggers a new translation and check.
How to review without reading everything
You don't need a person to read every translated word. You need the risky fields to reach the right person, and a sense of how the rest is doing.
- Split the page by field, not by page. Titles and descriptions can be sampled. Size charts, care, safety, ingredients and policy text go to a reviewer every time.
- Sample by category. Read a handful of products from each category in each language, especially new categories, rather than a random handful overall.
- Keep the source and the version. Store the original text, the translation you published and when it changed. When a customer complains, you can see exactly what they read.
- Let customers flag it. A small "report a translation problem" link on product pages costs little and finds things your sampling won't.
- Re-run the check when the source changes. An updated English description doesn't update the Spanish one on its own. Someone, or something, has to notice and send it through again.
The same thinking applies to images: our note on reviewing AI-generated product imagery separates what can be good enough from what has to be exact.
We build AI applications that keep sources, review steps and accountability visible, and we connect them to the storefront in our ecommerce work. If you're about to translate your store, or already have and want to know what slipped through, let's go through it together.
Related work: Howcast.
We build alongside the people doing the work.
Bring us the problem you’re thinking about.