Why evidence disclosure helps users
- Users report that seeing citations increases their confidence in an AI answer. However, users rarely actually click through to check them. The value of the disclosure is in the offer, not the click-through rate.
- Reasoning content is qualitatively different from output content—it may contain errors, false starts, or dead ends, and should be visually distinguished from the final answer rather than presented with the same confidence.
Types of evidence disclosure
Evidence disclosure falls into two main categories:- Reasoning trace/extended thinking—This typically appears above an agent response with a label like “Thought for [x] seconds” and a chevron that allows the user to click into the reasoning that led the agent to the answer it gave. It always has different styling from the chat itself to make it clear it’s categorically different from the styling of the agent responses themselves.
- Citations—These are pills that appear with an agent response that cite the agent’s sources.