A menu is three kinds of text
Treating a menu as one document is why translated menus read badly. It is three different kinds of text with three different rules.
| Text | Treatment | Why |
|---|---|---|
| Dish name | Keep the original, explain underneath | It is what the guest will say to the server |
| Description | Translate properly | This is what the guest reads to decide |
| Allergen and dietary terms | Translate, then verify with a speaker | An error here is a safety issue |
Section names, category headings, and operational notes — "served after 6pm", "minimum two people" — belong with the descriptions. They are ordinary prose and machine translation handles them well.
Where machine translation is genuinely good now
It is worth being fair about this, because the useful version of the advice is not "do not use it".
- Ordinary descriptive prose. Ingredients in a list, preparation described in a sentence, portion notes.
- Consistency across a long menu. A human translating a hundred items over two days will render the same phrase three ways; a machine will not.
- Speed against a changing menu. A weekly specials board cannot wait for an agency turnaround, and an untranslated specials board is worse than a machine-drafted one.
- Structural fidelity. It will not silently drop a line or reorder a list, which a hurried human copying between documents will.
Those four cover most of the words on a menu. The remaining few are where the damage happens.
The failure gallery
The errors are consistent enough to anticipate, which is what makes reviewing them fast.
Named dishes translated literally
A dish with a name is being named, not described. Rendering it literally produces something that is both unappetising and unorderable — the guest cannot say it to a server, and the server would not recognise it if they did. Keep the original and put the explanation underneath.
Cuts of meat
Butchery traditions divide the animal differently in different countries, so a great many cuts have no true equivalent. Machine translation reaches for the nearest common word and produces a cut that either does not exist in the target market or is a different piece of meat entirely. This is the single most common substantive error on translated menus.
Regional ingredients
A local variety of cheese, pepper, or grain gets replaced with the nearest supermarket equivalent. The guest orders expecting the thing they were told about. If an ingredient is regional enough to be a selling point, it is regional enough to keep in the original with a short gloss.
Cooking methods collapse
Braised, stewed, pot-roasted, and slow-cooked frequently arrive in the target language as one word. The distinctions carry real information about texture and richness, and losing them flattens the menu into a list of ingredients.
Protected and brand names
A protected designation of origin is a legal name, not a description. Translating it is wrong twice: it misleads the guest and it discards the reason you paid more for the ingredient. The same applies to any supplier brand you name deliberately.
Tone that does not travel
Jokes, dialect, and deliberately over-written menu copy arrive as nonsense or, worse, as something unintentionally rude. If your menu has a voice, accept that the translated version will have less of it, and prefer clear over clever in the target language.
Allergen terms need a person, once
This is the one category where a translation error is not an embarrassment but a hazard.
- "Nuts" and "peanuts" are one word in some languages and two in others, and a peanut is not a tree nut.
- "Shellfish" covers one category in US usage and splits into crustaceans and molluscs in EU labelling — a distinction that must survive translation intact.
- "Gluten-free" is a regulated claim with a threshold in most markets, not a synonym for "no wheat".
- "May contain" has an established legal reading in some languages and none in others, so a literal rendering can promise more or less than intended.
The good news is that this is a fixed, short vocabulary. Have a native speaker review the allergen and dietary terms once per language, lock them, and reuse them across every item. It is an hour of work per language and then it is done. The structuring side is covered in how to put allergen information on a digital menu.
Draft, review, approve
The workflow that makes this safe is not complicated, and it is the entire difference between a machine-translated menu that works and one that embarrasses you.
- Draft every item automatically from the source language. Do not translate item by item as you go — batch it, so the reviewer sees the whole menu and can enforce consistency.
- Lock the dish names before review, so the reviewer is not re-deciding that question a hundred times.
- Review in one pass against the failure list above, rather than reading for general quality. A reviewer looking for five specific error types is much faster than one asked to check everything.
- Verify the allergen and dietary vocabulary separately, once per language, and reuse it.
- Approve explicitly. Publication should be an act, not a default.
- On a menu change, re-draft only what changed. Do not regenerate the whole menu, or you will discard corrections a reviewer already made.
That last step matters more than it sounds. A system that re-translates everything on every edit quietly reverts human corrections, and the reviewer stops trusting their own work.
Who should review, and what to give them
The reviewer does not need to be a professional translator. They need to be a fluent speaker who understands food, and they need the right brief.
Realistic sources: a member of staff who speaks the language, a regular supplier, a hotel concierge you work with, or a freelancer for an hour. What matters is the brief, not the credential.
- Give them the source menu alongside the draft, not the draft alone. Without the original they cannot see what was lost.
- Tell them dish names are fixed and not up for discussion.
- Give them the five failure types to look for, so the task is checking rather than reading.
- Ask them to flag anything that sounds unappetising, not just anything incorrect. A technically accurate description that reads badly still costs you the order.
For choosing which languages to publish in the first place, and why the original dish name should stay visible next to the translation, see multilingual restaurant menus.