A synthetic customer is a polyglot by design, but refer to this Google page for the list of languages supported by the underlying Gemini model.

LLMs have a strong bias towards English for instructions and internal reasoning, so we recommend that you create the persona as follows:

• Name and Description: Enter these in English. Using English for this data leverages the LLM's core logic and instruction-following strengths.
• Traits: Here again, enter these in English for the same reasons. Importantly, provide the language-related instructions, for example, "This persona is a native of Madrid and speaks Castilian Spanish with local idioms."
• Typical responses: Provide these examples in the actual, target language (e.g., Castilian Spanish) to set the correct tone, syntax, and cultural "flavor."

Traits are instructions to the LLM (how it should behave), while typical responses are examples of output (what it should sound like). These operate at different layers of the LLM's processing, thus the different language recommendations.