How AI translation is reshaping QR menus across Türkiye
Two years ago, a multilingual menu was something boutique hotels had. In 2026 it's table stakes for a neighbourhood lokanta in Istanbul's Beyoğlu. Here's why that shifted, and the numbers behind our 14-language rollout.
Translating a menu is not swapping words. Render "kuru fasulye" as "dried beans" and you are technically correct while the guest still has no idea what they would be ordering. Menu translation is a context problem: portion size, cooking method, allergens and cultural reference all have to survive the trip.
Our first attempt at the AI translation layer was plain machine translation. It did not survive contact with the field. On the second attempt we fed the model the product category, ingredient list and price band alongside the string — and accuracy jumped sharply.
Tourist mix redrew the map
The visitor profile arriving in Türkiye has diversified over the past three years. Russian and German still dominate the Antalya coast; but demand for Japanese and Korean in Cappadocia, and for Arabic and Farsi in Istanbul, has climbed fast. A single English menu no longer covers the average.
- In coastal resorts, 58% of menu views open in a language other than Turkish.
- In Istanbul's historic peninsula that figure reaches 64%.
- In Anatolian city centres it is 9% — so the feature does not deliver the same value everywhere.
Why we kept a human in the loop
Every translation sits in an "awaiting approval" state in the dashboard before it goes live. That costs speed; it is non-negotiable on accuracy. A mistranslated allergen label isn't a poor user experience — it's a health risk.
When we opened the menu up to 11 languages, the first thing we noticed wasn't order volume — it was how many fewer questions guests asked the server. Service got faster on its own.
Where to start
- 1Look at your last three months of QR menu language stats — measure, don't guess.
- 2Translate your top 20 most-viewed items first; the long tail can wait.
- 3Put allergen and ingredient fields through a separate approval round.
- 4After going live, watch basket average by language for two weeks.
