Accuracy policy

If your bot states something as fact that your own published content does not support, that month is credited back. Not a partial refund and not service credit toward a future month — the month you were charged for.

Answers are generated only from content retrieved out of your own pages, and every answer is scored for confidence before it is sent. Below the threshold the bot does not attempt an answer at all — it says it does not know and offers to pass the question to you. Refusing is the designed behaviour, not a failure mode.

Every night a separate model re-reads the answers your bot gave real visitors — the visitor's question, the bot's reply, and the exact pages the bot retrieved to write it — and classifies each answer as supported, unsupported, or a refusal. Unsupported answers appear in your inbox with the transcript so you can correct them once, and the counts are on your dashboard. Before a bot has any traffic, an on-demand quiz drawn from your own pages plays the same role. We publish the same numbers to you that we use ourselves — there is no internal figure you cannot see.

A refusal is not an error. Neither is an answer that is correct according to your pages but your pages are out of date — the bot can only be as accurate as what you publish, which is why the dashboard shows you what it read and when it last re-read it.

Reply to any HelpBrain email with the conversation link. We look at the transcript and the retrieved sources for that turn. If the answer was not supported by your content, the credit is applied without further argument.