- A translated shop is four jobs wearing one name, and ai translation is excellent at one of them, risky at two and unsuitable alone at the fourth.
- The dangerous errors are the fluent ones. Research on post editing found coherence problems and meaning shifts cost the most effort, not obvious grammar mistakes.
- Under the EU product safety rules that apply from 13 December 2024, instructions and safety information must be in a language buyers there can easily understand.
- If you import stock and resell it, that obligation is yours. It falls on importers and distributors, not only on the original manufacturer.
- You are not required to translate your site to sell into the EU, and you are not allowed to redirect a visitor by country without asking first.
- Google ignores hreflang tags that are not reciprocal, so a one directional set of annotations does nothing at all.
The cheapest growth idea available to a small shop in 2026 is a second language. Translation that used to cost a few thousand and take three weeks now costs a few pounds and takes an afternoon, which means the question has moved from whether you can afford it to whether you can afford to get it wrong.
Getting it wrong is not a matter of clumsy phrasing. Clumsy phrasing costs you a little credibility and nothing else. The expensive failures are quieter: a size that means something different in the target market, a safety instruction that no longer says what the original said, a set of annotations that tells Google nothing because it points only one way.
A product page is four documents pretending to be one
This is the frame that makes the rest of the decisions obvious. The words on a product page do very different work, carry very different consequences when wrong, and deserve very different treatment.
| Layer | What happens when it is wrong | Who should check it |
|---|---|---|
| Marketing copy | Reads oddly, converts slightly worse | Nobody, or a native speaking customer |
| Product facts and attributes | Wrong item ordered, returns, disputes | You, against the source data |
| Safety and legal text | Regulatory exposure, product withdrawal | A human who reads that language |
| Units, sizes and currency | Silent conversion errors, refunds | Rules in your system, not the model |
| Hreflang markup | Wrong language ranks, or nothing does | A crawler, before launch |
Most shops translate all five with the same tool at the same setting, then evaluate the result by reading the marketing copy, which is the layer where the tool is strongest. That is why the failures show up later and elsewhere.
Why are fluent translations more dangerous than clumsy ones?
Because you cannot see them. A sentence that reads naturally and says the wrong thing passes every check a non speaker can perform, and it is also the error type the research says costs the most to fix.
A study published in Frontiers in Psychology on which machine translation error types drive post editing effort put 23 participants, 13 professional translators and 10 masters students, through 721 post edited sentences with keystroke logging and eye tracking. The pattern it found is the useful part. Coherence problems had the greatest effect on how long a sentence took to fix. Meaning shifts drove fixation duration, which is the measure of cognitive load rather than typing. Grammatical faults, the ones a reader notices immediately, mostly affected the technical edit distance.
Translate that into shop terms. The errors that feel worst are cheap, because you spot them in a second and correct them in five. The errors that cost you are the ones where the sentence flows and the meaning has drifted, because nobody flags them and they reach a customer intact. That is exactly the profile of a mistranslated material, a mistranslated care instruction, or a compatibility claim that quietly became broader than the original.
The same study found professionals absorbed meaning shifts without much extra cognitive effort while students did not. Experience in the language is what catches this class of error. That is an argument for paying a native speaker to read your fifty highest revenue pages once, rather than paying anyone to read all nine hundred.
What does the law actually require you to translate?
Less of your site than you think, and more of your product documentation than you probably realise. Those two facts pull in opposite directions and both surprise sellers.
On the site itself, the answer is nothing. The EU geo-blocking rules give customers in other member states the right to access your online interface and buy on the same terms, and as the European Consumer Centres explain in their guide to the end of geo-blocking, that does not create an obligation to translate. A trader may keep the site in its own language, and is not compelled to deliver to every country either. What the rules do require is that you stop automatically redirecting a visitor to a national version without their consent, which is a feature many storefronts still ship switched on by default.
That last point is worth sitting with, because it inverts the usual instinct. The helpful thing your site does, detecting a German visitor and sending them to the German page, is the thing that needs permission. Offering the choice is fine. Making it for them is not.
Who carries the safety text obligation?
Whoever puts the product in front of the buyer, which in a small import business is you. This is the part that catches resellers, because the wording of the obligation does not stop at the manufacturer.
The EU general product safety rules apply from 13 December 2024, and the UK government's detailed guidance on Regulation 2023/988 sets out how the duties distribute. Manufacturers must ensure products come with clear instructions and safety information in language consumers can easily understand. Importers carry the same duty and may add the missing information themselves where the manufacturer failed to. Distributors have to verify that the information is there, and may also add compliant labelling or documentation if it is not.
Read as a chain, that means the obligation follows the product rather than resting with its maker. A shop importing goods from outside the EU and selling them to European consumers occupies the importer position and inherits the duty. An AI translation of a safety instruction is a perfectly reasonable first draft of that text. It is not a reasonable last draft, and the difference is one person who reads the language confirming that the warning still warns about the same thing. We looked at a narrower version of this problem, where the stakes are highest, in the piece on allergen wording in product copy.
The technical layer that quietly does nothing
Hreflang is where most multilingual shops lose the benefit they paid for, and the failure is silent. There is no warning, no error, no drop in a dashboard. The annotations simply get ignored.
Google's documentation on localized versions of your pages gives three equivalent methods: HTML link elements in the head, HTTP Link headers, or entries in an XML sitemap. You pick one. Using all three is allowed and gives no benefit in search, while tripling what you have to keep in sync.
The rule that breaks setups is reciprocity. If two pages do not both point to each other, the tags are ignored. Each version has to reference itself as well as every sibling. So a shop that adds German annotations to its English pages, and forgets to add English annotations to the German ones, has done nothing at all. This is extremely common because the second half of the job happens in a different template.
Three further details are worth knowing before launch. The x default value catches any language you have not explicitly listed, and it works best on a selector page or a homepage that redirects. Google does not use hreflang or the HTML lang attribute to work out what language a page is in; algorithms do that from the content, so the annotation is organisational rather than declarative. And localized versions only count as duplicates where the main content is untranslated, so a translated template wrapped around identical core content is a genuine problem while a fully translated page is not.
Units and sizes are not a translation problem at all
They are a data problem wearing a translation costume, and treating them as text is how shops end up shipping the wrong item. A model asked to translate a page will happily render a measurement, and it has no obligation to convert one.
The categories that bite are predictable. Clothing and shoe sizes follow incompatible national systems, so a number that survives translation intact is usually wrong. Paper sizes, screw threads, plug types and voltages are national standards rather than words. Volumes and weights move between metric and imperial in ways that produce plausible looking numbers. Dates written in figures reverse meaning between markets, which matters for expiry and for delivery estimates.
The fix is structural rather than linguistic. Keep measurements as numbers with units in your product data, render them through conversion rules at display time, and never let free text carry a figure that matters. If a size chart exists, translate the labels and leave the numbers to the system. This is much easier when your product data lives in a database you control rather than inside prose, which is one of the practical reasons to run the storefront on your own code, as our approach to building the shop assumes.
What happens when the replies start arriving in that language?
You get a support inbox you cannot read, which is the consequence almost nobody budgets for. Translating the shop is a one time cost. Translating the conversation is a permanent one, and it starts the week the first order lands.
The volume is manageable, and the shape of it is not what sellers expect. A second language does not double your support load, because most of the extra questions are the ones your translated pages failed to answer. The first month of foreign language messages is therefore a bug report on your translation, and it is worth reading as one. If eleven German customers ask whether a cable fits a particular socket, the German page did not say, and fixing the page removes eleven messages a month permanently.
For the replies themselves, the working arrangement for a one person business is to draft in your own language, translate, and send both versions. Including the original underneath is not a confession of weakness. It gives a customer who reads some English the ability to resolve an ambiguity themselves, and it gives you a record of what you actually meant if the exchange is ever disputed. The cost is a few lines of extra text.
Two categories should never be handled this way. Anything about a safety issue and anything about a refund or legal right needs a person who reads the language, for the same reason the safety text did. A fluent mistranslation in a refund reply is a promise you did not intend to make, in a jurisdiction whose consumer rules you probably do not know in detail.
How do you tell whether the second language is earning anything?
Separate the traffic before you launch, not afterwards. Any measurement that starts after the fact will be contaminated by the traffic you already had, and the question you actually care about is whether new people arrived.
Three numbers answer it. Sessions on the translated URLs from the target country, which tells you whether search found and served the new pages. Conversion rate on those sessions against your baseline, which tells you whether the pages persuade anyone once they arrive. And the return rate on orders from that market, which is the number that exposes translation failures in product facts, because a customer who received something other than what the page described sends it back.
That third number is the one to watch hardest in the first quarter. A translated shop with healthy traffic, decent conversion and an elevated return rate has a data problem rather than a marketing problem, and the returns will point at the specific pages where the facts drifted. It is a slower feedback loop than you would like, which is the argument for getting the structured fields right at the start rather than discovering them through refunds.
Give it two full quarters before judging. Search takes time to index and rank a new language section, and the first weeks are dominated by whatever direct and referral traffic you send yourself. A verdict reached in week three is a verdict about your own announcement rather than about the market.
A sequence that works for a shop of one
Do it in this order and you will not have to redo it.
Pick one market rather than five. The cost of a second language is not the translation, it is the support, the returns and the tax handling that arrive with real customers, and those scale with the number of countries rather than the number of words. Cross border logistics and customs paperwork carry their own surprises, which we set out separately in the piece on what customs actually asks of a small seller.
Translate the marketing copy with the model and do not agonise over it. This is the layer where the tool genuinely performs and where an imperfect result costs you very little.
Extract every fact into structured fields before translating anything: dimensions, materials, compatibility, quantities, care instructions. Translate the labels, never the values. A field called material holds a code that maps to a translated word in each language. A sentence that says the strap is leather holds nothing, and has to be re translated and re checked every time you touch it.
Route the safety and legal text to a human who reads the language. Fifty pages of high revenue products, once, is an affordable engagement and it is the part that carries real exposure. A broader audit of what your listings claim, across languages, follows the same method we described for checking listing compliance at scale.
Build the hreflang set last, generate it from one source rather than by hand, and crawl the site before launch to confirm every page points back. Then leave it alone. It is the one part of this job that stays correct without maintenance, provided you got it right once.
That sentence has cost more international SEO effort than any other in the documentation. It is also the easiest thing on this list to verify, which makes it a good place to start: crawl your own site, check that every language version names all the others including itself, and you have eliminated the most common reason a translated shop earns nothing from being translated. Everything else on this list needs judgment about your products and your market. That one needs a crawler and twenty minutes, which makes it the highest return check available to anyone reading this.