// CASE 01 · AIRBNB × TRANSLATED
Redesigning translation trust, one listing at a time.
About one host in five wrote their listing mixing two languages. Guests got confused, and the new machine-translation pipeline built with Translated had nothing clean to work on. We redesigned the listing flow so hosts would write in one language, and trust what the machine did with it.
fig 01 · The redesigned localization surface, host side.
fig 01 · The redesigned localization surface, host side.
5.6M
listings on the marketplace in early 2021
~1.2M
listings affected by mixed-language descriptions
16
hosts tested, two rounds plus shadowing
2021
shipped globally in the Winter Release
fig 02 · Activities and deliverables: the research plan, end to end.
01
Executive summary
Airbnb had just signed with Translated to localize the whole marketplace through a mix of machine translation and professional translators. One thing broke the pipeline before it started: hosts writing the same description in Italian and English, or in a language different from their account. The machine could not tell which voice to keep. We ran an expert review and two rounds of usability testing with hosts, and gave the product team a redesigned listing flow plus a set of interface changes to stop the problem at the source.
02
Four phases, one question
➔ Desk research and hypothesis framing on why hosts mix languages
➔ Data-driven recruitment of hosts through internal Airbnb queries
➔ Expert review of the listing creation journey
➔ Usability testing with 16 hosts in two rounds, plus shadowing sessions
Tools & tactics
Mixed-methods UX research
Remote moderated testing
Heuristic analysis
Microcopy and flow redesign recommendations
03
Issues discovered
➔ Localization features had no onboarding: most hosts did not know translation existed.
➔ Nothing warned a host writing in a language different from the one set on the account.
➔ No inline guidance or reminder about which language to write in.
➔ Prices were shown with a 1:1 currency conversion (100 USD read as 100 EUR).
➔ No preview of the translated listing, so hosts did not trust the machine and wrote both versions by hand.
Five design principles, one per issue
Put hosts in control
Guide and reassure
Explain, then ask for trust
Prevent the error before it happens
Show, don't just tell
04
Results
Shipped in the Winter Release 2021, across the whole marketplace. Mixed-language listings dropped, machine translation became the default hosts accepted, and the localization team had evidence instead of assumptions to prioritise the backlog. The lift in UX and revenue is an internal Airbnb figure I cannot publish, so it stays qualitative here. What I would do differently: measure post-booking signals (complaints, no-shows, reviews) from day one, not only listing conversion. Today, guest-host chat on Airbnb is translated in real time on the same foundations, and the same question, "does the machine keep the human voice?", is the one I ask of every AI feature I ship.
05
Lessons learned
➔ Localization is not a feature, it is an experience
➔ When people understand the process, they trust it
➔ Microcopy can shape macro outcomes
➔ Working next to data science cuts research time in half
