Translating a website used to mean a per-word quote and a few weeks of waiting. Today DeepL or ChatGPT will produce a readable French version in an afternoon. When we build bilingual sites as part of our services, the question clients ask has changed: nobody asks whether the machine can translate anymore. They ask whether they can publish what it produced.
The short answer is no, not as is. Translation has stopped being the expensive step. Review has taken its place, and review is exactly what Quebec's Charter of the French language and Google both look at, each in its own way.
Our article on the bilingual website and Bill 96 covers the legal framework, OQLF complaints and hreflang. This one covers the workshop: how to produce the French (or English) version with AI without creating the problem the law targets.
The short answer, for the busy
You can translate a website with DeepL or ChatGPT, as long as you treat the machine's output as a first draft. These tools write fluent, grammatically clean French, and DeepL offers a "French (Canadian)" target separate from French as written in France. What they miss shows up in review: sentence structures copied from English, European vocabulary, typography, brand tone. Quebec law targets precisely that quality. Section 52 of the Charter of the French language forbids offering a commercial document in another language if the French version is not available on terms at least as favourable, and the regulation requires quality at least equal. On Google's side, the documentation recommends a separate URL for each language, and pages mass-produced with little value for readers, translation included, fall under its spam policy. The workflow that holds up: glossary, machine translation, review by someone who knows the business, then publication.
What the machine does well, and why the problem moved
For website copy (service pages, product pages, FAQs), DeepL and large language models deliver sentences that are correct and rarely wrong in the schoolbook sense. They get through a hundred pages in minutes.
DeepL has two specific strengths for a Quebec site. Its translator offers two French targets, "French (France)" and "French (Canadian)", and its documentation lists glossary support for the Canadian target: you fix, once, how your service names and trade terms are translated. ChatGPT, Claude and other language models work differently. They translate less sentence by sentence and follow instructions instead: audience, register, a glossary pasted into the request. That is their strength for marketing copy and their risk, because a model that rephrases can also add a sentence or drop one.
Either way, the output looks finished, and that is the trap. A poor human translation is visible. A poor machine translation reads smoothly, and its errors are the ones a hurried reader misses and a Quebec customer notices at once. The OQLF's own terminology record for machine translation still notes that it "only produces an approximate translation". The record is harsher than 2026 tools deserve, but the practical conclusion stands: the machine writes, someone has to read.
What it misses in Quebec French
Review looks for more than spelling mistakes.
French from elsewhere. If the tool does not know the reader is in Quebec, nothing forces it to write like Quebec. The simplest test is "courriel", the term the OQLF prefers for email: a page that says "e-mail" or "mail" was not written for a reader here.
Calques from English. A machine translating from English carries English structures across. "Adresser un problème" (to address a problem) is the textbook case: the OQLF's Banque de dépannage linguistique advises against this use, borrowed from to address, and suggests "régler", "traiter" or "s'attaquer à". The text stays grammatical and still sounds translated.
Typography. Headings with a capital on every word, prices written "$24.99" instead of "24,99 $", missing or misplaced spaces around punctuation. The OQLF's guidance for the web applies the same rules as print.
Tone. A model can use vous on one page and tu on the next. Your brand picks one: the instruction has to be written down, and checked.
Commitments. Lead times, warranties, prices. A nuance lost in a delivery promise becomes a commitment you never made.
Which way you translate matters more than the tool
The Charter does not protect your two versions equally. Section 52 requires commercial documents, "regardless of the medium used", to be drawn up in French, and forbids making them available in another language "if the French version is not available on terms that are at least as favourable." Section 10 of the Regulation respecting the language of commerce and business adds, for two separate versions, "no less favourable conditions of accessibility and quality". That is all the workflow needs; the rest of the framework is in our Bill 96 article.
The consequence is direct. If your team writes in English and the French comes out of the machine, the version the law protects is the one nobody wrote. If you write in French and translate into English, the protected version is the one a person thought through, and machine translation touches the version where errors cost the least.
So when you can, write in French first. When you can't (English-speaking founders, copy supplied by a head office, a manufacturer's product sheets), reviewing the French stops being a precaution and becomes the core of the job.
Glossary, review, sign-off
The workflow has four steps, in this order.
Glossary first. Before translating a single page, list your service names, product names and trade terms with their required translation, plus the words that are never translated (brand names). That file is what you load into DeepL or paste into the model's instructions. Without it, the same service will carry three different names on three pages.
Machine translation next, page by page, with the full instruction: Quebec French, register, intended readers.
Review by someone who knows the business. Writing well is not enough: the reviewer has to be able to tell whether the French page promises what you sell, at the right price, on the right timeline. A good habit: read the French page without looking at the English, the way a customer would. What snags on reading is what sounds translated.
A named sign-off. Someone in the company approves the published version. Without a name, review is the first thing dropped when a change is urgent.
Legal pages leave this workflow. Privacy policy, terms of sale, terms of use: an approximate translation changes the meaning of an obligation. Hand them to a lawyer or to a certified translator from the OTTIAQ (Quebec's order of certified translators, terminologists and interpreters). Terms people accept without negotiating also carry an extra language rule, covered in our article on website legal pages.
Translation plugin, or pages built into the site
That leaves the technical side. Google's documentation is clear on three points. It recommends a different URL for each language rather than content that changes with a cookie or browser settings, noting that Googlebot usually crawls from the United States and sends no language preference. It determines a page's language from its visible content, not from the lang attribute in the code. And it advises against automatically redirecting visitors based on the language you assume they speak.
A script that translates the page inside the browser fails the first test: Google may only ever see the original language, and your French version does not exist for it.
Plugins that serve each language at its own URL pass that test. Weglot, for instance, creates a subdirectory per language and adds the hreflang tags itself. But these tools can put machine output online without anyone reading it. In Weglot, unticking the option that displays automatic translations shows the original content in their place: if your site is written in English, the unreviewed parts of your French page appear in English. And when you edit source text, the tool treats it as new content and generates a fresh automatic translation. WPML, on WordPress, publishes immediately if the administrator chose to publish without review, and otherwise sends the translation to a review queue. In both cases a setting decides whether review happens. Check it.
The approach we prefer builds each language as real pages of the site: one URL per language, text written in that language, and nothing live before sign-off. The machine works in the workshop, never on the public page. Google does not penalize machine translation as such: its scaled content abuse policy targets pages mass-produced with little value for readers, and names translation among the techniques involved.
If you change architecture along the way, read our article on redesigning without losing SEO; the platform comparison shows how each one handles bilingual sites.
The three-page test
Open your site in the language you don't write in and read three pages like a customer. Look for "e-mail", English-style capitals, a badly formatted price, a sentence that snags. Then check your translation plugin's setting: does it publish without review? Then draw up the glossary and name the person who signs off.
This work gets done properly once, then again at every update. We build it into every bilingual site we make.
This article explains language obligations in plain terms to help a small business ask the right questions. It is not legal advice: check your situation with a lawyer. The official text of the Charter of the French language and its regulations, on LégisQuébec, prevails.
Written byXavier Peich