The purpose of a sandbox is to make mistakes cheaply. When you first train a bot on your website content, some of its answers will be wrong, vague, or accidentally revealing. Discovering that in front of a customer is expensive. Discovering it in a playground costs you ten minutes and a correction. This is why the sensible launch sequence is always build, test, correct, then publish.
What to test there: your top twenty real customer questions, taken from your actual inbox rather than invented; the awkward questions about price, availability and guarantees; questions your bot should refuse, to confirm the guardrails hold; and deliberately ambiguous phrasing, because real visitors are not careful typists. Also test questions you have no content for, to see whether the bot admits uncertainty and hands off rather than improvising.
Clerkzo gives you a playground for exactly this from the moment your 14-day trial starts, so the sandbox stage costs nothing and there is no reason to skip it. Once you are satisfied with the answers, going live is a single script tag and takes about ten minutes, which means the testing is the part worth spending time on.
Related guides
Related terms
Browse all 104 terms- Answer CorrectionAn answer correction is the act of identifying a wrong or unhelpful chatbot reply and fixing it at the source, so the same question gets the right answer the next time it is asked.
- AI GuardrailAn AI guardrail is a constraint that defines what a chatbot will and will not say or do, keeping it inside safe, accurate and on-brand territory regardless of what a visitor asks.
- Knowledge SourceA knowledge source is any body of content a chatbot uses to answer questions, such as your website pages, product descriptions, policy documents or a list of frequently asked questions.
- Conversation RatingA conversation rating is a simple signal, usually a thumbs up or thumbs down, that a visitor gives to a chatbot reply to indicate whether it was helpful. It is the cheapest quality feedback a chat widget can collect.