Multilingual e-commerce: AI translation and hreflang

In short: for an online store or B2B website selling abroad, AI translation is a strong first step but not enough on its own. Copy that sells or commits your company should be reviewed by a native speaker. Each language needs its own URLs, translated slugs and reciprocal hreflang tags, or Google will show buyers the wrong version. Choose your tool by testing it on your own catalog, and only translate confidential documents or personal data on plans that do not reuse your text.

Key takeaways

Machine translation works for low-risk content; sales copy, key pages and legal text need human review.
Google recommends a separate URL for each language and advises against redirecting users automatically based on the language you assume they speak.
Hreflang only works when every page lists itself and all its alternates, with return links.
There is no single best translation software: machine translation engines, language models and platform apps do different jobs, and you pick them by testing on your own content.
Free versions may use your text to train their models: price lists, quotes and personal data belong on business plans with written data-use terms.
Publishing pages in bulk that are machine-generated or machine-translated and add no value can fall under the scaled content abuse described in Google's spam policies.

Why a multilingual website matters if you export

Online buyers prefer to read in their own language. In a survey of 8,709 consumers in 29 countries, CSA Research found that 76% prefer to buy products with information in their native language and 40% never buy from websites in other languages. The data is B2C and dates from 2020, but the logic carries over to B2B: a technical buyer who needs internal sign-off forwards your product page to colleagues, and a page in clumsy English slows the decision down.

The second reason is search: a German distributor searches in German, using the terms of their industry, and if your German version is missing or poorly indexed you simply do not show up. That is why a multilingual site is a B2B e-commerce and international SEO project, not just a translation job, and part of any digital export strategy. You will find more in the websites and e-commerce section.

Machine translation, post-editing, localization: what's the difference?

These are different levels of work, with different costs and results.

Machine translation: translation produced by software, today based on neural networks or large language models, with no human involved.
Post-editing: human review of machine-translated text. The translator fixes errors, terminology and tone instead of translating from scratch.
Localization: full adaptation to a market. Besides the language, it changes units of measure, date and price formats, examples, images, regulatory references and keywords.
Transcreation: creative rewriting of a message for another market. It keeps the intent, not the words.
Glossary: the list of terms with their approved translation, such as product names, technical terms and words that must not be translated. The DeepL API, for example, lets you build glossaries from pairs of source and target terms.

When is machine translation enough, and when do you need a human reviewer?

It is enough when the text is informational, repetitive and low-risk. You need a person when the text has to persuade, commits the company, or contains technical data that causes returns or disputes if it is wrong. The table sums up the rule of thumb I use.

Content	AI alone	AI plus native-speaker review	Professional translation or copywriting
User reviews and questions	Yes, labeled as machine-translated	Not needed	No
Catalog attributes (color, material, size)	Yes, with a glossary and spot checks	On the first rows, to tune the glossary	No
Product and category descriptions	As a draft only	Yes	For flagship products
Homepage, service pages, campaign landing pages	No	At a minimum	Recommended (transcreation)
Data sheets, instructions, safety warnings	No	Yes, by a reviewer who knows the industry	For documents with regulatory weight
Terms of sale, privacy, warranties	No	No	Yes, with a legal check in the target country

Which tools should you use to translate an online store or B2B website with AI?

There is no single best translation software: what matters is where the text ends up and who checks it. The tools fall into four families, often used together.

Type of tool	Examples	Best suited for	Watch out for
Machine translation engines	DeepL, Google Cloud Translation	High volumes of consistent text: attributes, descriptions, documents	Without a glossary, industry terms drift from page to page
General-purpose language models	ChatGPT, Claude, Gemini	Rewriting and adapting tone, suggesting title and FAQ variants, adapting campaign messages	Output changes with the instructions: always compare it with the source
Platform translation features	Shopify Translate & Adapt, multilingual plugins for WordPress and WooCommerce	Managing languages inside the CMS, with per-language URLs	Check that they produce correct URLs, slugs and hreflang and that every text can be reviewed
Automated workflows	n8n or Make connected to translation APIs and the ERP	Large, frequently updated catalogs and multiple channels (website, marketplaces, feeds)	You need a review queue: automation moves text, it does not check it

Translation engines accept glossaries: Google describes the Cloud Translation glossary as a custom dictionary used to translate domain-specific terminology consistently, such as product names. On Shopify, the free Translate & Adapt app translates store content automatically or manually and shows source and translation side by side for review.

The most reliable way to choose is a test on your own catalog: take 20-30 real texts of different kinds (data sheet, description, category page, customer email), translate them with two or three tools and have a native speaker from your industry rate them blind, noting terminology errors, omissions and tone. The winner is the tool that makes the fewest mistakes on your vocabulary, not the one that tops the rankings.

How should you structure URLs, slugs and hreflang on a multilingual site?

The basic rule is one language per page and one URL per language. Google recommends separate URLs for each language version rather than switching language with cookies or browser settings, and advises against redirecting users automatically based on the language you think they speak. A visible language switcher on every page works better. There is a technical reason too: by default Googlebot crawls from IP addresses that appear to be in the USA and without an Accept-Language header, so a site that picks the language from the visitor's IP or browser may hide some versions from it.

The options are a country domain (example.de), a subdomain (de.example.com) or a subfolder (example.com/de/), the simplest for most small and mid-sized companies.

A slug is the last part of the URL that identifies the page. On a multilingual site it should be translated: /en/ball-valves and /de/kugelhaehne are clear to anyone scanning search results, /de/ball-valves is not.

Hreflang is the annotation that tells Google which pages are language or regional versions of the same page. According to Google's documentation, four rules apply:

each version lists itself and every other version;
links must be reciprocal: if page A points to page B, B must point back to A, or Google ignores the annotations;
URLs are fully qualified (https://…) and codes give the language (de) or language plus country (de-CH), never the country alone;
it is worth adding x-default, the page shown to people whose language you do not cover, such as a language selector.

Hreflang can go in the HTML, in HTTP headers or in the XML sitemap. One detail that often gets missed: Google detects language from the visible content, not from the lang attribute or the URL. A /de/ page with a German menu and a product description left in the source language is not a German page.

I have seen this first-hand: on my own website an automatic language redirect led Google to treat some Italian pages as the English version. In October 2026 I removed it and left the choice to visitors, with a visible switcher.

How do you localize product pages for an online store?

Start with the data, not the copy: a mistake in structured attributes repeats across the whole catalog.

Units and formats: centimeters or inches, EU, UK or US sizes, decimal separators, date formats, currency and taxes shown the way the market expects.
Local keywords: the right term is not always the literal translation. Before translating titles and categories, check what people in that country actually search for.
Technical documents: in B2B, data sheets, certifications, lead times and delivery terms matter most. They are what buyers download and forward.

Take a hypothetical scenario: a component manufacturer with 2,000 catalog items wants to enter Germany and France. A sensible workflow translates every attribute with AI using the glossary, has a native speaker from the industry review the highest-revenue product pages and the category pages, and queues the rest with spot checks. Moving data between the ERP, the translation tool and the CMS can be automated with n8n or Make, like any other business process automation; inside the CMS, AI features should be used selectively, as I explain in the guide to AI in WordPress and other CMSs. This is the kind of work covered by my website and e-commerce solutions.

What language requirements apply when selling online in Europe?

Translation is not just marketing. For consumer products, the EU General Product Safety Regulation (EU) 2023/988 (GPSR), which applies from 13 December 2024, requires in Article 19 that online offers clearly and visibly show, among other information, warnings and safety information in a language easily understood by consumers, as determined by the Member State where the product is sold. The same regulation requires manufacturers and importers to provide instructions and safety information in the language set by that Member State.

In practice, warnings and instructions should never go live as unchecked AI translations, and the required language has to be checked country by country with whoever handles product compliance. In pure B2B, with goods not intended for consumers, industry rules and contracts apply: ask your distributors which language they need for data sheets, certificates and manuals.

Beyond the website: product feeds, email and customer support in other languages

A translated website is only the first channel. The same copy is needed elsewhere, and it pays to translate it once from a single source (ERP or PIM) and distribute it everywhere.

Google Shopping and marketplaces: for selling abroad, Google Merchant Center suggests localizing language, price and currency, translating landing pages and checkout into the local language and uploading translated product data as a new data source, with prices in the local currency.
Inquiries and quotes: AI helps you understand a request in Polish or Swedish and draft a reply, but a person checks prices, quantities, lead times and delivery terms before it goes out. The glossary you use for the website applies here too.
Customer support: machine-translated chat and email work for common questions, as long as complaints and technical cases are handed over to a person. I cover this in the article on omnichannel customer experience in B2B.
Campaigns and social media: ads, email subject lines and posts are adapted (transcreation), not translated word for word. A pun translated literally rarely works.
Distributors and sales network: price lists, presentations and training material in the partner's language cut down repeated questions and sales mistakes.

Does Google penalize AI-translated websites?

No, as long as the translations are useful and well looked after. Google rewards content quality, however the content is produced. Its spam policies, however, list two relevant examples of scaled content abuse: using generative AI to produce many pages that add no value, and generating pages from other people's content through automated transformations, translation included.

The risk is not using AI but publishing thousands of pages in ten languages with no quality control. Three well-kept languages beat ten neglected ones.

Is it safe to translate business documents and data with AI?

It depends on your plan, not just the tool. A documented example is DeepL's privacy policy: on its free services, DeepL processes submitted texts and their translations for a limited period to train and improve its neural networks, and its terms of use rule out texts containing personal data of any kind on the free version. With DeepL Pro and the API Pro, texts and translations are processed only temporarily and deleted once the service has been performed, except for what you ask it to keep, such as glossaries and saved translations.

The rule of thumb: confidential price lists, quotes, contracts and customer or employee data only go through business plans with written data-use terms and, where personal data is involved, a data processing agreement under Article 28 GDPR, which DeepL's own privacy policy provides for. The same applies to any tool, ChatGPT included: before uploading documents, check your plan's terms to see whether content is retained or used for training. For the bigger picture, see the article on AI and cybersecurity.

How much does it cost to translate a website or catalog with AI?

The software price is only one line item. To estimate the real cost, look at four factors:

Volume: characters to translate per language, including attributes, filters and system text. Services like DeepL sell subscription plans with a monthly character cap and a number of glossaries that grows with the plan (see DeepL's pricing); some platforms include free tools, such as Shopify's Translate & Adapt app.
Human review: the most variable item, driven by how much of your content, in the table above, needs review or professional translation.
Technical setup: URL structure, hreflang, language switcher, integration with the ERP or PIM.
Maintenance: every new product, promotion or price change has to reach every language. Without an automated workflow, the foreign-language versions fall behind the original.

How to apply this in your company: first steps

Pick markets from data: check in GA4 and Search Console which countries already send visits and inquiries, and start with one or two languages.
Classify your content with the table above: for each type, decide whether AI alone is enough, or you need review or professional translation.
Build the glossary (product names, technical terms, do-not-translate words) and have it approved by someone who knows the market.
Choose the tool with a blind test on 20-30 real texts, rated by a native speaker from your industry, and check that your plan does not reuse your text.
Set up the technical structure before translating: subfolders per language, translated slugs, reciprocal hreflang with x-default, a language switcher and no automatic redirects. If you lack in-house SEO skills, here is how to choose an SEO consultant.
Measure by language: after two or three months, compare impressions and clicks by country in Search Console and conversions by language in GA4. To pick the right metrics, see the guide to KPIs for measuring the value of AI.

Mistakes to avoid

One URL that switches language through cookies or browser settings: Google may only see and index a single version.
A widget that translates the page on the fly in the browser: it creates no per-language URL, so there are no translated versions to index.
Hreflang without return links or with wrong codes, such as "uk" for the United Kingdom: that is the code for Ukrainian, British English is en-GB.
Half-translated pages: menu and footer translated, main content still in the source language. Google sees duplicates, not localized versions.
Confidential text in the free version of an online translator: price lists and quotes end up in a service that may use them to train its models.
No one answering in that language: a translated site generates inquiries, and a German lead who gets a reply in another language a week later is a lost lead.

Frequently asked questions (FAQ)

Can I translate my whole website with ChatGPT or DeepL?

Yes for a first version, and for catalog attributes and reviews that is often enough. Your homepage, service pages, main products, data sheets and terms of sale should still be reviewed by a native speaker who knows your industry.

Subfolders, subdomains or country domains: which is better?

For most small and mid-sized companies a subfolder (example.com/de/) is the simplest option, because every language shares the same domain and technical setup. A country domain (example.de) sends the clearest geographic signal for a single country but costs more to run.

What is x-default and when do I need it?

x-default is the hreflang value that tells Google which page to show people whose language your site does not cover. It usually points to a language selector or to an international version, often in English.

Does machine translation create duplicate content?

Not if the main content is translated: Google treats localized versions as duplicates only when the main content stays in the original language. The risk lies in pages where only the menu and footer change.

How many languages should I start with?

Start with the markets where you already see demand: GA4 visits, Search Console impressions, quote requests, active distributors. Add more once the first ones work and someone can answer customers in that language.

What is the best AI translation software for a business?

It depends on the content. Machine translation engines such as DeepL or Google Cloud Translation, used with a glossary, suit volume and terminology; language models such as ChatGPT are good for rewriting and adapting tone; platform apps such as Translate & Adapt on Shopify manage languages inside the store. The most reliable test is a blind comparison on 20-30 real texts, rated by a native speaker from your industry.

Can I use the free version of DeepL for business documents?

Not for confidential texts or personal data. According to DeepL's privacy policy, texts sent to the free services are processed for a limited period, including to train its models, and the terms of use exclude personal data from the free version. For these documents you need DeepL Pro, where texts and translations are deleted after the service.

Sources

If you are taking your online store or B2B website into new markets and want to get languages, URLs and translations right, book a free assessment.