E-Commerce Translation in 2026: What Actually Works

A shopper in Munich lands on a fashion site they’ve never seen before. The hero banner is in German, but as they scroll the product titles stay in German while the descriptions are in English.

They click into a product page and the size guide is half in each language. They add to basket: the overlay loads in German, the checkout drops back to English. The order confirmation email arrives in English again.

This is roughly the modal experience for mid-sized retailers expanding across Europe. The translation work got done, only halfway. The customer notices, trust evaporates, and the conversion rate tells the story a few weeks later.

What follows is what the grown-up version of e-commerce translation looks like in 2026, and the operational decisions that separate a translated site from a site that earns trust in every market. The principles apply to any multilingual storefront, and they sit alongside the broader website translation work we structure for retailers expanding cross-border.

What’s actually changed in e-commerce translation since 2020

Three shifts since 2020 matter for how you scope and budget your translation programme today: large language model translation has absorbed the bottom 60% of e-commerce content, EU regulation has caught up with cross-border online retail, and buyer expectations on integration have moved from “send us a spreadsheet export” to “connect directly to our PIM.”

LLM translation has eaten the easy stuff

Neural machine translation became operationally usable around 2018. Large language model translation, since 2023, has done to the next tier what neural did to the first.

Spec-heavy product attributes, repetitive descriptions, FAQ content, and most user-generated content can now move through machine translation with light human review and reach acceptable quality. The cost per word for that band has dropped by a factor of five or six since 2020.

This doesn’t mean you can machine-translate everything. It means the cost curve has bent enough that the operational question now is which content type belongs in which tier of your stack, not whether to use machine translation at all.

Regulation has caught up

Three regulations now shape what gets translated and how:

• General Product Safety Regulation (Regulation 2023/988), in force 13 December 2024, replacing the 2001 directive. Safety information must be provided in the official language(s) of every member state where the product is offered.
• Omnibus Directive (Directive 2019/2161), transposed across the EU during 2022 to 2023. Changes the rules on “reduced from” price claims and consumer review authenticity.
• Digital Services Act (Regulation 2022/2065), fully applicable from 17 February 2024. Transparency obligations for marketplaces and platforms, including local-language information for traders.

None of these were on the radar in 2020. All three now sit alongside translation procurement. Section 5 unpacks each one.

Integration is the new baseline

In 2020, the typical translation workflow for a mid-sized retailer involved a marketing coordinator exporting product data from the storefront, emailing it to a vendor, receiving translated files back, and re-uploading them. That worked at small scale and one or two markets. It doesn’t scale to eight markets running weekly content drops.

The new baseline is connector-driven translation. Source content changes trigger a translation job, completed translations write back automatically, and a localisation manager monitors throughput in a dashboard. Section 6 covers the data flows in detail.

The content-type decision matrix: what to translate, how, and with what QA

Not every word in your e-commerce stack needs the same translation treatment. The matrix below maps each content type against a recommended translation route, the QA stack to apply, and a realistic cost band per 1,000 source words. It’s the lookup table your localisation manager actually needs.

How to read the matrix

Routes: raw MT, MTPE (ISO 18587), professional human translation (ISO 17100), and transcreation. QA options: MT-only, MTPE-light review, ISO 17100 four-eyes (translation plus revision), in-country review, and legal review. Cost bands are in EUR per 1,000 source words, assuming Western European target languages and a documented QA cycle.

Cost bands assume Western European target languages. Asian and Nordic targets sit 15 to 30% higher. Specialist sectors (regulated pharma, complex legal) sit higher still. Sources: CSA Research 2024 market rate data and our operational reference rates.

Where buyers tend to get this wrong

Three common errors come up in vendor handover audits:

• Treating the whole catalogue as one content type. Spec-heavy attributes get MTPE’d correctly, and the premium hero descriptions on the same pages get MTPE’d the same way, so the brand voice flattens across both.
• Skipping in-context review for UI strings. Strings fit in English, then overflow the button in German.
• Letting raw MT touch policy pages. One client mismatch on a returns policy can create chargeback exposure that costs more than the entire translation budget.

The matrix above plus a multilingual style guide and a maintained termbase close most of those gaps.

Machine translation, MTPE, and human translation: choosing the right route

For e-commerce content, the route question is rarely “MT or human.” It’s which content gets raw MT, which gets MTPE, which gets human, and which gets transcreated. The right answer for each content type depends on three factors: brand-voice sensitivity, regulatory exposure, and content lifespan.

Where machine translation works for e-commerce

High-volume, low-stakes, short-shelf-life content is where machine translation pays off:

• User reviews displayed inline.
• Internal documentation.
• Customer support drafts before human review.

Shoppers don’t expect UGC to be edited, internal users tolerate ambiguity, and for both, speed beats polish. The current ceiling on neural and LLM-based machine translation for retail content is high enough that fluency is rarely the problem. Errors are now mostly terminology and tone.

Where machine translation breaks for e-commerce

Four content types are where buyers run into trouble:

• Brand-led product descriptions. The output reads as grammatical and dull. The model has no read on the brand voice.
• Sizing, fit, and dimension content. Translators see subtle drift across languages and cultural edges that machine translation misses. “Slim fit” lands differently in markets where the word reads as an insult about the wearer.
• Policy and legal pages. Machine translation confidently invents wording that doesn’t exist in the target jurisdiction.
• Marketing copy with idioms or cultural references. Output gets translated literally and lands flat in market.

MTPE: what good post-editing actually involves

ISO 18587:2017 defines post-editing as editing and correcting machine-translated output. Two levels exist:

• Light post-editing. Comprehensibility and accuracy, no stylistic polish. Suitable for internal content, support drafts, UGC display.
• Full post-editing. Comprehensibility, accuracy, terminology, style, and brand voice. Suitable for product content and FAQ articles.

Most translation suppliers default to full MTPE in pricing. Ask which level you’re getting and at what rate, and ask what the post-editor’s instructions actually say.

When transcreation, not translation, is the right call

Transcreation is the right call wherever conversion depends on emotional response, not information transfer:

• Hero pages and campaign creative.
• Slogans, taglines, and headline copy.
• Push notifications and email subject lines.

Transcreation is a separate skill, typically priced by hour or by project, not by word. The slogans article goes deeper on when and why this matters.

Multilingual SEO that actually drives traffic to your store

Good multilingual e-commerce SEO needs three layers in this order: a URL architecture that signals language and region clearly to Google, correct hreflang implementation across every translated page, and localised metadata and structured data on product pages. Skip a layer and the whole stack underperforms.

URL architecture: subdirectories, subdomains, or ccTLDs

Most retailers should start with subdirectories. ccTLDs only make sense when you already have country-level platform separation or strong legal reasons to own the country domain.

The four hreflang mistakes that break ranking

1. Missing x-default. Google falls back to the wrong language for unmatched locales.
2. Conflicting canonicals. Pages canonicalised to the English version and also marked as the French version. Pick one.
3. Mismatched return tags. Every hreflang on page A must point to a page that hreflang-points back to A. One break and the cluster degrades.
4. ISO code errors. “en-uk” doesn’t exist (it’s en-GB). “es-LA” isn’t standard. Use ISO 639-1 plus ISO 3166-1 alpha-2 only.

Localising metadata and product structured data

• Translate, don’t transliterate, page titles and meta descriptions. Apply local keyword research, not literal source-language keywords translated.
• Localise alt text per market. A “running shoe” image alt-text needs to change when the local search term is “chaussures de course”.
• Product schema (Product, Offer, AggregateRating) needs language-specific values for name, description, and priceCurrency. Aggregate reviews per market where possible.
• Image filenames don’t need translating. Image alt attributes do.

For the longer treatment of multilingual ranking strategy, including the technical SEO patterns we apply at scale, see our multilingual SEO article. The dedicated SEO translation service covers the operational side, from keyword research per market through to metadata localisation.

EU compliance: what you can’t translate around in 2026

Three regulations now shape what you must translate, how prices and reviews must be displayed, and what transparency you owe customers in the EU. None of them were on the radar in 2020. All are in force or fully applicable as of early 2026.

GPSR and the language obligation

Regulation (EU) 2023/988, the General Product Safety Regulation, has been in force since 13 December 2024. It replaces the 2001 General Product Safety Directive.

• Article 9 requires safety information, warnings, and instructions to be provided in the official language(s) of the member state where the product is made available. Not optional. Not waivable on grounds that the consumer reads English.
• Applies to all consumer products sold online. Resellers, marketplaces, and dropshippers are in scope alongside manufacturers.
• Online marketplaces carry additional traceability obligations.
• Translation scope: safety warnings, hazard information, usage instructions, age-restriction notices, and any text affecting safe use.

Omnibus Directive: price-display, “reduced from”, and reviews

Directive (EU) 2019/2161, transposed across the EU during 2022 and 2023.

• Article 6a requires that any “previously” or “reduced from” price refer to the lowest price applied during the 30 days before the reduction. Local language obligations apply to the price label and any explanatory text.
• Reviews must be verified as genuine when presented as “consumer reviews”. Translated reviews need authenticity signalling.
• Personalised pricing must be disclosed. The disclosure must be in the local language.

Digital Services Act: transparency for marketplaces and platforms

Regulation (EU) 2022/2065, the DSA, fully applicable since 17 February 2024.

• Online marketplaces must verify trader identity and display certain trader information to consumers. All trader information in the local language.
• Very Large Online Platforms (45 million plus EU monthly active users) carry additional content moderation transparency obligations.
• Recommender systems must be explained to users. Explanations available in the local language.

Compliance obligations at a glance

For policy pages, returns content, and trader information sitting under these obligations, scope the work through the legal translation service rather than the general translation workflow. Different reviewer pool, different sign-off, different audit trail.

How translation actually moves through your e-commerce tech stack

Translation in a mid-sized e-commerce setup moves through up to five systems: a storefront platform, a PIM, a DAM, an optional headless CMS, and a TMS. The translation operation is only as fast and clean as the connectors between them. Manual file exports work for two markets and break at four.

Storefront platforms and their translation hooks

• Shopify Plus: native Markets feature, multilingual capabilities, JSON-based content export, third-party apps for deeper TMS connection.
• Adobe Commerce (Magento 2): store views per language, native multistore, API access to product attributes.
• BigCommerce: Multi-Storefront, Stencil themes, multi-language support via API.
• Salesforce Commerce Cloud: Site Languages, content slots per locale, deep API access.
• commercetools: headless from the ground up, locale-aware product data model.
• Centra: built for fashion brands, native multi-market and multi-currency.
• Shopware: store inheritance and multi-language storefronts via the Storefront API.

PIM systems and where translation belongs in the catalogue lifecycle

Translation in a PIM-led setup happens at the product attribute level, not the storefront level. This is the right architectural pattern: translate once in PIM, syndicate to every channel.

• Akeneo: locale-aware attributes, dedicated translation workflow, market-leading among mid-market retailers.
• Salsify: Productxp, dedicated translation app, broad connector library.
• Pimcore: open-source, flexible, more setup work.
• inriver: enterprise-focused, syndication-first model.

Headless CMS and the API-first pattern

Contentful, Sanity, and Storyblok all support locale-per-entry and machine-readable APIs. The pattern: content created or edited in the CMS triggers a translation job via webhook; completed translations write back via API. End-to-end latency drops from days to hours.

Connector logic in practice

A connector sits between your source system (storefront, PIM, DAM, CMS) and the TMS. Its job: detect changes, send to translation, receive completed translations, write back to source. Our SmartConnect is our connector layer, built for this pattern.

Two examples from our client base illustrate the spread:

• Universal Robots uses SmartConnect to keep its website content in sync across multiple languages without manual file movement. Catalogue updates flow from the source CMS straight through the translation workflow and back to the live site. That’s the manufacturer end of the spectrum: technical content, structured documentation, controlled velocity.
• Our client Ziener, a German sportswear brand, sits closer to the fashion-retail end. Their setup runs on a SmartDesk core, with translation memories and a termbase covering product names, fabric types, technical features, and garment details, across French, Italian, and English.

The architectural principle holds across both cases: locale-aware attributes in the source data model, translation jobs triggered at the field or attribute level, completed translations writing back to the same record. The connector logic is the same; only the volume and content sensitivity differ.

Both patterns share three properties worth pulling out:

• One source of truth for content. Translation work targets that source, not the channels.
• Change detection at attribute or field level. Whole-page translation is the wasteful option once you’re past the first launch.
• Bidirectional connectors, not one-way exports. Translated content writes back automatically, with audit trail.

Operational benchmarks for an e-commerce translation programme

If you’re scoping a translation programme or evaluating a vendor’s pitch, the benchmarks below set the bar for what realistic throughput and turnaround look like in 2026. Numbers assume Western European target languages and a documented QA cycle. Ask any vendor whose figures fall outside these bands to explain why.

Source: CSA Research 2024 market data, Slator 2024 industry report, our operational reference rates.

How to use the benchmarks in an RFP

Map your annual content volume by content type to the matrix in section 2. Apply throughput benchmarks to estimate linguist days and cycle time. Add latency benchmarks to your launch plan for new markets.

Compare any vendor proposal against these bands. Numbers way better are aspirational; numbers way worse are inefficient.

Programmes that hit the benchmark ranges typically run on disciplined terminology management with a maintained termbase and translation memory across markets.

How we handle e-commerce translation

Everything in the sections above, the matrix, the route choices, the SEO architecture, the compliance overlays, the integration patterns, the benchmarks, is how we structure e-commerce translation programmes for clients. Our delivery platform is SmartDesk, our connector layer is SmartConnect, and our in-context layout review tool is SmartEdit.

Our SmartDesk is the central platform for managing multilingual content. Content upload and assignment across workflows, real-time progress and budget tracking, approvals, translation memory and termbase management, and the connection point for connector-driven integrations.

Our SmartConnect ties our SmartDesk to the storefront, PIM, DAM, or headless CMS. The Universal Robots integration is documented in the published case study. The Ziener workflow operates on the same connector pattern at fashion-and-sportswear retail scale.

Our SmartEdit handles in-context layout review for UI strings, banners, push notifications, and any content where layout matters. Reviewers see the translated content in the rendered context, not as a string in a spreadsheet.

For more on the buyer side of how we support retailers and marketplaces, see the e-commerce industry page.

If you’re scoping a multilingual e-commerce programme, or want a second opinion on an existing setup, speak with our localisation team. We can walk through where the throughput and integration gains live in your current stack.

Get in touch: Speak with our localisation team.

Frequently asked questions

How much does e-commerce translation cost per language?

There’s no per-language number that covers the question. Cost depends on content type, route (raw MT through to transcreation), QA stack, and volume. The matrix in section 2 shows realistic 2026 bands per 1,000 source words for Western European targets, from EUR 15–30 for raw MT on UGC up to EUR 200–320 for transcreated campaign pages. In a mature programme with translation memory, your effective rate per language drops 30 to 50% over 12 to 18 months.

What’s the best e-commerce platform for multilingual stores?

There’s no single right answer. Shopify Plus, Adobe Commerce, BigCommerce, Salesforce Commerce Cloud, commercetools, Centra, and Shopware all support multilingual setups, with different trade-offs on flexibility, headless capability, and operating cost. The platform decision should follow your existing stack, team capability, and PIM strategy, not the other way round. Section 6 maps each platform’s translation hooks.

Can you machine-translate an entire e-commerce site?

Technically yes, operationally no. The matrix in section 2 shows where raw MT is acceptable (UGC, live chat, internal docs) and where it isn’t (checkout strings, marketing copy, legal pages). Full-site MT carries two risks worth pricing: conversion drop on the content types that need human work, and compliance exposure under GPSR and Omnibus for safety information and policy pages.

How long does it take to launch an e-commerce store in a new language?

Four to eight weeks for a clean launch with proper QA and SEO localisation, assuming an existing source-content base and a defined scope. The variables that move that figure: content volume, integration depth (manual versus connector-driven), regulatory-content workload (cosmetics, food, toys, medical devices add weeks), and review-cycle capacity in the target market.

How do you keep translations in sync when products change daily?

Connector-driven workflow. A change in your source system (PIM or CMS) triggers a translation job, completed translations write back automatically, and the storefront picks up the updated content via syndication or direct integration. Section 6 maps the data flow. Latency target: under 48 hours from source edit to live in market, with full connector integration.

Sources

• Regulation (EU) 2023/988 (General Product Safety Regulation), EUR-Lex. In force 13 December 2024.
• Directive (EU) 2019/2161 (Omnibus Directive), EUR-Lex.
• Regulation (EU) 2022/2065 (Digital Services Act), EUR-Lex. Fully applicable 17 February 2024.
• ISO 17100:2015 Translation services – Requirements for translation services.
• ISO 18587:2017 Translation services – Post-editing of machine translation output – Requirements.
• CSA Research, market rate data, 2024 release.
• Slator 2024 Language Industry Market Report.
• Google Search Central, International and multilingual sites documentation (current version).
• Schema.org Product / Offer / AggregateRating specifications (current versions).
• AdHoc Translations ISO 17100 and ISO 18587 certificates, 2025.