Ecommerce translation services: what to look for

Ecommerce translation services are providers that convert an online store's content into other languages and keep it that way as the catalogue changes. That second half is the part worth paying for, and it is the part almost nobody sells against. Most providers price the first translation of your product copy, category pages, navigation, policies and transactional email. Very few will tell you who does the work in month seven, when four hundred new SKUs have landed and half of them are English again.

This is a buyer's guide rather than a pitch. It sets out what to look for in an ecommerce translation service, the question to ask a vendor under each criterion, and what an evasive answer sounds like. Where we make a claim about our own results, it is marked and sourced.

What does an ecommerce translation service actually cover?

An ecommerce translation service covers every piece of text a shopper meets between landing and confirmation. In practice that means product titles and descriptions, collection and category pages, site navigation and filters, search results, size and care information, policy and legal pages, transactional email such as order and shipping notifications, and support macros. It usually does not cover text baked into images, banners or video.

The scope question matters more than it sounds, because scope is where quotes diverge without anyone noticing. Two providers can quote the same rate per word against very different definitions of the job. One is pricing your product catalogue. The other is pricing your product catalogue plus the eleven Shopify sections, the returns policy, the abandoned-cart email and the customer-service macros, which is a materially larger number and a materially more complete store.

It is also worth being clear about what separates this from a translation project in general: a store is not a document. A document is finished. A catalogue changes weekly, which is why the difference between translation and localisation for a store ends up being a question about operations rather than vocabulary.

What should you look for in an ecommerce translation service?

Look for seven things: defined scope, ownership of ongoing work, a stated quality process, a way to detect bad translations rather than only produce good ones, evidence of measured outcomes, transparent pricing with no per-language surprises, and genuine ecommerce integration rather than a file handover. Most buyers instead compare languages offered, years in business and price per word, which are the three criteria on which every provider looks broadly the same.

That is not an accident. Translation quality has become cheap and widely available, so the things that used to differentiate providers have flattened. What has not flattened is the operational load of running a multilingual store, and that is where providers genuinely differ.

Who owns the work after launch?

Ask who owns the work after launch, and get the answer in writing, because it is the single criterion that predicts whether the project still works in twelve months. Three models exist. An agency translates a snapshot and the work stops when they invoice. A self-serve app translates on demand but leaves your team running glossaries, reviewing strings and chasing what changed. A managed service takes the operating burden itself. All three are legitimate. Only one of them survives a catalogue that turns over weekly without consuming somebody's Thursday.

The failure mode is specific and quiet. Products land, collections retire, a seasonal page goes up, and a French shopper meets a half-English page. Nobody on the team notices, because nobody on the team is reading French. On Shopify this shows up often enough that merchants describe it in public reviews: pages published to live international storefronts partly translated, or reverted to English after a change, and found weeks later.

Our own answer to this is continuous localisation: you add a product on your source store, and it is live in every language without anyone touching it. That is the standard worth holding any provider to, whether or not the provider is us. If the answer to "who translates next month's products" is "you re-run it", the operating cost has been moved onto your team and priced at zero.

What is "AI plus human" actually describing?

"AI plus human" almost always describes machine translation followed by human post-editing, and that workflow has its own international standard, separate from the one most providers display. ISO 17100:2015 sets requirements for translation services, including translator competence and a revision step by a second person, and it explicitly places raw machine-translation output plus post-editing outside its scope. The standard covering that workflow is ISO 18587:2017, which applies only to content processed by machine translation systems and distinguishes full post-editing from light post-editing.

This gives a buyer an unusually sharp question. If a provider markets AI speed with human quality and displays ISO 17100, ask which standard covers the work you are buying. There is nothing wrong with post-editing, and nothing wrong with ISO 18587. The useful signal is whether the provider knows the difference and will tell you plainly which one describes their delivery.

Our own delivery is a stack rather than a single step: AI translation, refined by models specialised per language, reviewed by human proofreaders. Nearly all our merchants choose human proofreading, because brand voice is the point. The reason to state the architecture rather than a quality adjective is that architecture is checkable and adjectives are not.

How do you check quality in a language nobody on your team reads?

Producing an acceptable translation is close to solved; finding the one wrong sentence in ten thousand, when nobody in the building reads the language, is not. This is the question most quality conversations skip, and it is the one that actually protects a brand. Ask a provider how errors surface, not how translations are made.

Two mechanisms are worth asking about. The first is automated drift detection: daily checks that flag wording which has moved away from versions already approved for your brand, so reviewers are pointed at candidates instead of reading everything. The second is scored sampling. Native linguists score representative samples using a framework such as MQM, weighting terminology, accuracy, linguistic conventions and style, with each error graded from neutral to critical, and several independent scorers per sample so the result reflects the language rather than one person's taste.

Neither mechanism is exotic. What is unusual is a provider volunteering them, because both imply that errors exist and will be found, which is a less comfortable pitch than "our quality is very high".

What outcome evidence should you ask for?

Ask for a result measured against an untranslated control, and treat everything else as background. The statistics circulating in this category are almost entirely preference surveys quoted without a study, an edition, a sample or a date. CSA Research's Can't Read, Won't Buy: B2C 2020 edition, based on 8,709 consumers across 29 countries, found 76% of consumers prefer to purchase products with information in their own language. That is a real, dateable figure, and it still only tells you what people say they prefer, not what changes when you translate.

The spread is the honest part of that number, and any provider quoting an average without one is hiding the interesting information. A brand whose category is already understood sits near the bottom of the range; one whose product needs explaining sits near the top.

Timelines are the other evidence worth asking for, because greenfield buyers fear a nine-month programme. Real deployments do not look like that. Sabre Paris launched five markets in two weeks and doubled German revenue. Rodier launched eight markets in two months and grew organic revenue 22%. Ask for launch windows in the same shape, with the market count attached.

What does it cost, and what drives the bill?

Publicly quoted rates for professional human translation run from roughly $0.09 to $0.40 per word, with most commercial work in the $0.15 to $0.30 band. Those are vendor-published ranges rather than an independent survey, so treat them as orientation rather than a benchmark. What actually drives the bill is not the rate: it is words multiplied by languages multiplied by how often your catalogue changes.

That third term is what makes multi-country expensive under a per-word or per-word-per-language model, and it is the term nobody puts in the quote. A model that tiers investment by market priority, spending on premium quality where the market justifies it and running unlimited machine translation everywhere else, is a different shape of bill rather than a cheaper version of the same one.

Where does a translation service stop?

Text inside images, banners and video is not translated, so campaign visuals stay in the original language unless you produce local versions yourself. Most premium brands prefer that, because art direction is not something to hand over. It is still worth naming at the quote stage rather than discovering in week three.

Several other things stay with you regardless of provider: brand and terminology decisions, market and pricing strategy, commercial and legal positions, and the final judgement of whether the result sounds like you. Any provider claiming otherwise is overselling. The honest version is that a good service absorbs the chasing, the re-running, the glossary reconciliation and the hunt for the half-translated page, and leaves you the decisions.

When is a translation agency the right answer?

A translation agency is the right answer when you have a defined, high-stakes body of copy in a small number of languages and you want people rather than a pipeline. Brand manifestos, campaign lines, a flagship collection story, regulated product information: these reward a named translator who learns your brand, and they do not change weekly.

The mismatch appears when that model is applied to a moving catalogue. An agency prices a snapshot honestly and delivers it well, and then your store keeps changing. Neither party did anything wrong; the shape of the engagement simply did not match the shape of the work. If most of your volume is product content that turns over, you are buying an operating model, not a body of translation, and the comparison worth running is between platforms rather than between agencies. Our rundown of the Shopify translation apps compared covers the self-serve end of that market, and the wider guide to ecommerce translation sets out the four approaches side by side.

If what you want is managed ecommerce localisation rather than a supplier to brief, the questions in this guide are the ones to put to us as readily as to anyone else.

Frequently asked questions