TL;DR
Compare the same administrative facts across languages, using questions reviewed by a qualified speaker. Preserve each answer and its sources separately. A fluent translation or a similar-looking answer does not establish that both versions identify the right office and appointment process.
Define a fact comparison, not a language ranking
An owner may want to know whether people asking an assistant in different languages receive consistent practice information. Start with a narrow task: identify the office offering a service and explain how to request an appointment.
This is a factual review, not a competition to see which language produces more brand mentions. Different questions and contexts can produce different answers. A small sample cannot establish a platform's overall quality for a language or a population.
Choose only public administrative information. Do not translate patient histories or personal health questions into the test set.
Build matched questions with a qualified reviewer
Begin with one approved practice fact record. List the office name, service availability, contact route and request-versus-confirmation process. Then ask a qualified language reviewer to create natural questions that seek the same information in each language.
Do not assume a literal machine translation preserves the task. A phrase translated as “book an appointment” may carry a stronger expectation than “request a callback.” Have the reviewer explain any difference that could affect the answer.
Use a worksheet with these fields:
| Field | What to preserve |
|---|---|
| Fact being tested | The exact approved administrative fact |
| Question versions | Natural wording reviewed for equivalent intent |
| Context | Platform, date and visible location/session details |
| Answer versions | Original-language answer, kept intact |
| Source links | Exact citations supplied with each answer |
| Reviewer finding | Correct, incorrect, ambiguous or not assessable |
Compare facts before style
Treat a change from “request a response” to “confirmed appointment” as a scheduling-state discrepancy to investigate, even when the translated answer reads smoothly. Compare each version with the approved process before deciding which wording is wrong.
If an answer cites an outdated translated page while another language version is current, investigate source maintenance. That observation alone does not establish a systematic language failure.
Google recommends distinct URLs for language versions of web pages. Where the practice has such versions, record the exact language-specific source rather than assuming every link refers to the same current content. Google's multilingual-site guidance.
Route the finding to the right owner
Use the practice-information correction process to preserve the observation and separate source repairs from answer rechecks. Send a stale translated source to the content and language reviewers. Send an incorrect operational fact to the practice owner of that fact. If the cited source is accurate but the answer changes its meaning, preserve the evidence and use the platform's documented feedback route where available.
Keep unresolved language judgments marked unassessable until a qualified reviewer can evaluate them. Do not let an English-only reviewer approve a foreign-language answer solely from an automated back-translation.
Repeat the method after corrections
Verify the source change first. Then repeat the matched task and record the new observations without erasing the old ones. Report each language's factual findings separately; do not average away an incorrect appointment instruction because another answer was accurate.
Frequently asked questions
Must every question be translated word for word?
No. The aim is equivalent administrative intent expressed naturally. Record meaningful wording choices so the method can be repeated.
Does this replace a translated-website review?
No. It examines observed AI answers and their sources. The practice still needs an appropriate process for approving and maintaining its own language versions.