Confidence and retrieval
Interpret confidence bands and retrieval-mode comparisons without treating them as a quality guarantee.
Confidence is the score of the best matching passage the agent retrieved for that message. It reads as a percentage: 92 means the strongest passage very probably answers the question. It does not count how many passages matched, so a source that states a fact once scores the same as a source that repeats it on ten pages.
The calculation changed on 19 September 2026. Before that date, confidence was the average of the five passages the agent kept. That average fell when a source stated a fact only once, because the unused slots held unrelated passages. Messages from before the change therefore read lower than the same message reads today. Compare a date range that crosses this change with care, and prefer a range on one side of it.
Analytics groups recorded assistant confidence into High (70 or more), Medium (50–69), and Low (below 50). Average confidence and the distribution include only assistant messages that have a stored confidence value. A message has no stored value when the agent retrieved nothing, when the agent skipped the search because the message needed none, or when the scoring step did not complete. Confidence is a retrieval-related signal, not a promise that an answer is correct, complete, safe, or approved for a customer.
When low-confidence answers are present, the alert can take you to Activity filtered for low confidence. When you dismiss the alert, the browser remembers this choice for a short time only. The dismissal does not retrain the agent or change any stored score. Review the underlying conversations, their sources, and the visitor’s question. Then decide whether to improve the sources, the instructions, or a workflow.
The Performance tab also groups responses by recorded retrieval mode: Standard, Hybrid, and Deep Search. It shows counts and average confidence by mode. Not every response has a mode or a confidence value. For this reason, do not compare these charts as if they show all traffic. Use them to form a hypothesis, then validate it through Activity and controlled tests.