The Glossary Nobody Sets Up, and What It Costs

Ask someone what went wrong with their machine-translated site and they will describe a sentence that reads awkwardly. Ask them to find it again and they usually cannot. The awkward sentence was a one-off. The problem they actually have is somewhere else, repeated four hundred times, and they have stopped noticing it.

That problem is vocabulary. It is the failure mode of modern translation, it is invisible page by page, and it is fixed by a list you can write in ten minutes.

Grammar stopped being the problem

We benchmarked eleven models on real product copy, in several languages, twice each. Not one produced a sentence a native speaker would call broken. The defects were all of one kind: terminology and register.

A product family named one way in the category and another way on the product page. A trademark helpfully translated into a common noun. A technical term rendered with the word from a neighbouring industry, which reads fine to anyone except the customers who know the difference.

None of those look like errors in isolation. Each one is a plausible translation of the words in front of it. They only become visible when you see the whole catalogue at once, which nobody does.

What a glossary is, exactly

A glossary is a short list telling the translation engine two things: which terms must never be translated, and which terms must always be translated in one specific way.

That is it. It is not a style guide and it is not a dictionary. Ten to forty entries covers most businesses.

The value is that it is deterministic. Everything else about machine translation is a judgement call the model makes fresh every time. A glossary is the part you decide once, and it holds across every page, every language and every future translation run.

What belongs in it

Your brand and product names. The obvious ones. Also the ones that are ordinary words in your language, which are precisely the ones that get translated: a product called Compass, Origin or Nova will be rendered as a common noun in half your target languages unless you say otherwise.

Model numbers and references. Anything mixing letters and digits. They usually survive, and when they do not, the damage is disproportionate because a customer cannot order what they cannot name.

Your industry's terms of art. The words your customers use to search. If your market says one specific thing and the general-purpose word is a near synonym, the near synonym costs you visibility.

Terms you have already decided. If your German market manager settled on one word for your main category two years ago, that decision belongs in the glossary. Otherwise every translation run relitigates it.

Names of features and offers. If you have named a service, a warranty or a delivery option, it should either always be translated or never be. Alternating is worse than either.

What does not belong in it

This is where glossaries do damage, and it is worth being blunt about it.

Common words. Locking a frequent word to a single translation fights the grammar of the target language. Nouns decline, adjectives agree, word order changes. Forcing an invariant form into a Polish or German sentence produces text that is wrong in a way a reader immediately notices, unlike the subtle inconsistency you were trying to fix.

Whole phrases. A glossary works on terms. A locked sentence fragment collides with the sentence around it.

Anything you are unsure about. An entry you added on a hunch will be applied hundreds of times. If you would not defend it to a native speaker, leave it out.

The rule of thumb: a glossary entry should be a noun or a name, and forcing it should never change the sentence it sits in.

The traps in practice

A glossary does not repair what is already translated. It applies to translation runs, not to stored text. Adding an entry and expecting the site to change is the most common disappointment, and the fix is to re-translate the affected content deliberately.

Case and inflection. Depending on the tool, an entry may match exactly or loosely. If your brand appears capitalised in titles and lowercase in body text, check that both are covered.

Cached pages. After a re-translation, what you see may be the old page from a full-page cache or a CDN. Before concluding that the glossary did nothing, request the page with a cache-busting query string. This one wastes more support time than any actual bug.

Per-language exceptions. Sometimes a term must stay in English in Dutch and be translated in Spanish. A glossary that only supports one global rule per term will force you into the wrong choice for one of those markets.

Ten minutes, once

Open a spreadsheet. Write your brand name, your product line names, your three or four most important industry terms, and any term someone in your company has already argued about. That is the first version, and it is already worth more than the second version you will never write.

Then improve it from evidence rather than imagination. When a native speaker reviews a page and says a term is wrong, that goes in the glossary, so their one hour of attention applies to every page you translate afterwards. When a customer email misuses a term, look at how you named it. When Search Console shows your pages matching queries you did not expect, the market is telling you what it calls your product.

That is the real economics of it. A glossary turns a one-off correction into a permanent rule. Without one, every review finds the same problems again, and every new translation run reintroduces them.

Both the WordPress plugin and the PrestaShop module take a glossary, and it is the first thing worth setting up, before the first run rather than after the third. If you are about to review a translated site, the glossary is also the cheapest check available: searching for terms is something you can do in a language you do not read, which is the whole point of reviewing a translation you cannot read.