Skip to main content

Understand Feedback, Consent, and Evidence

The evidence model can include question text, assistant response, message-level context snapshot, bounded context journey, thumbs rating, structured failure reason, follow-up behaviour, unanswered state, and representative evidence in Insights.

Structured negative-feedback reasons​

After a downvote, a visitor can choose:

  • Unclear
  • Outdated
  • Didn't answer question
  • Missing instructions
  • Need a human
  • Other

These are user-reported observations, not verified root causes. Undoing a rating removes its reason. “Other” is a category, not an invitation to place personal data in context.

Conversation logging requires a non-Free plan, enabled widget logging, and visitor opt-in. Until consent is known, eligible messages may be held briefly in the browser; opting out drops them. An absence of records can therefore mean no consent, no eligibility, no traffic, a deployment failure, or narrow filters—not an absence of user difficulty.

Eligible records can contain user-entered text and bounded attribution such as host, path, locale, UTM values, widget version, and runtime context. Do not put personal data in context, launcher IDs, or campaign parameters.

Interpretation limits​

Context shows where evidence arose. A journey shows bounded movement through a conversation. Failure reasons show what a visitor selected. Follow-up and unanswered states help triage. Theme grouping selects representative evidence. None independently proves causality.

Validate a theme against original Captured Questions, the affected product surface, and current source material before acting.