A basic estimate divides self-service conversations without a follow-on ticket by total conversations. Treat that figure as directional: users can abandon after a poor answer, so reliable measurement also needs feedback or linked support outcomes.
The hard part is attribution. A user who got a wrong answer and gave up counts as deflected under naive measurement. Better measures combine conversation outcomes, user feedback, and whether the same user opened a ticket afterwards.
Deflection rate is a lagging metric; the leading inputs are answer accuracy and docs coverage. Unanswered questions help you investigate coverage, retrieval, or configuration problems. Review the conversation and relevant docs before deciding what to fix.