1 minute readReference note
How do you test brand information in voice assistants?
In brief
Test the spoken question and spoken answer as they were actually heard. Speech recognition, name pronunciation, location context, response length and follow-up conversation can change the research task before any brand fact is evaluated.
Retain the recognized wording when the product exposes it. Distinguish a misheard company name from a wrong fact about the correct company. Ask realistic availability, comparison and current-status questions, including relevant accents or language variants when those matter to the audience. Compare the spoken output with any on-screen text, sources and full-length reference.
Correct the diagnosed layer. Clear names and aliases can help factual references distinguish entities; a transcript can expose a missing qualification in a quoted statement; an outdated public record needs a source update. Where the product has used web search, inspect that context rather than assuming the voice interface represents a separate source universe. OpenAI documents location and conversation-related context in ChatGPT search. Retest in fresh and follow-up conversations, preserving which conditions changed and whether the listener now receives the correct meaning.