Glotari

WORKFLOW

Why an AI without domain knowledge translates your shop worse

Eight real errors from 19 languages and the misconceptions behind them. All of them come from our own shop, most from runs with a current model and a careful prompt.

· 7 min read

A current language model translates a product text fluently and almost always grammatically correct. That is exactly what makes its errors dangerous: they sound right. What is missing is not language but knowledge of your catalogue. Which words are names, which are technical terms, which fields must not be translated at all. We translate naturbohne.de, a shop for coffee and espresso machines, into 19 languages. All of the following errors come from machine translations in that shop, most of them from our own runs.

Eight errors that sounded fluent
SourceResultWhat was missing
Mondgold DecafMondgold Descafeinado (Spanish)knowing that this is a product name
Eureka Mignon TurboЕурека Миньон Турбо (Bulgarian)brand and model stay in Latin script
Superkop, ElektraSuperhead, Electrabrand names are not words
Siebträgersieve carrier (51 products)technical term: portafilter
Abschlagboxtee boxtechnical term: knock box
Siebträger (Bulgarian)10 different wordsone term for the whole shop
bei Venedignear Venedig (all 19 languages)a rule without context: growing region or city
Handfilter (Swedish)Handfiltera German leftover that looks like a technical term

Machine translations on naturbohne.de, found between 07.09. and 27.09.2026. Evidence in the rule files of the translation runs and in the review of 27.09.2026.

Misconception 1: the AI recognises names by itself

It recognises names it knows. "Mondgold Decaf" is a coffee from our own roastery, not a term from the training data. To the model, "Decaf" was a word, so it translated it, and in Spanish the product name became "Mondgold Descafeinado". Anyone in the Spanish shop searching for the name on the bag no longer finds it.

With well-known brands the opposite happens, and it is no better. "Eureka Mignon Turbo" came back in Bulgarian in Cyrillic script, in Greek half of it. "Superkop" became "Superhead", "Elektra" became "Electra". A customer looking for a grinder searches for the manufacturer's name as it appears on the machine.

Misconception 2: a good model knows the technical terms

It knows the words, not your field. A Siebträger is a "portafilter" in English. Machine-translated, 51 products said "sieve carrier", the literal translation no barista uses. The Abschlagbox became a "tee box", a term from the golf course. Both are grammatically flawless and technically wrong.

For SEO this is not cosmetic. Nobody searches for "sieve carrier". A page that uses the word your customers do not search for will not be found for the search they actually make.

Misconception 3: each translation on its own is good enough

When products are translated one by one, the machine decides anew every time. In Bulgarian our shop uses ten different words for "Siebträger". Taken one at a time, none of them is completely wrong. Together they break filters, because the same type of machine runs under ten names, and they make the shop look as if nobody checked.

Misconception 4: a rule in the prompt solves it

Rules help, but only with the knowledge of what they are for. Our instruction "names of regions, farms and places stay unchanged" was meant for growing regions, so that a cooperative in Peru does not turn into a translated fantasy name. The model also applied it to "Venedig". Result: "near Venedig" in all 19 languages, because the German name of the city is itself already a translation.

The same goes for checks. On one day our search for German leftovers triggered three correction runs for "French Press", "Cold Brew" and "V60", which are correct as they are in every language. Conversely, "Handfilter" in a Swedish text looks like an international technical term and is still a German word. A rule without domain knowledge finds the wrong thing and misses the right one.

Misconception 5: everything that is text gets translated

A shop is more than descriptions. Metafields hold units, identifiers, product feed codes and markup. Translating them costs money, and for Google Shopping fields the value may no longer be readable afterwards. So the question is not only how well something is translated, but also what must not be translated at all.

What domain knowledge means in practice

For us it comes down to four things, none of them magic:

  1. A never-translate list for brands, models and your own product names, as phrases rather than single words, so "Turbo" stays free as a word but "Mignon Turbo" does not.
  2. A glossary with fixed term pairs per language: Siebträger is portafilter, Abschlagbox is knock box. Decided once, the same everywhere.
  3. Rules with a reason and a limit: growing regions stay, cities with their own name get translated.
  4. A second check with the same knowledge. Otherwise it flags French Press as an error and lets Handfilter through.
A model translates language. It only knows your catalogue once you tell it.

Evidence

  1. Glotari Magazine: German text on the Norwegian page, case study of 27.09.2026
  2. Glotari Magazine: What a Translation Score measures, and what it does not