In practice
After closing a support ticket you ask "How satisfied were you with the help you received?" 80 people answer and 62 pick 4 or 5. Your CSAT for support is 62 ÷ 80 = 77.5%.
It isn't net promoter score, which asks how likely someone is to recommend you. It isn't customer effort score, which asks how hard something was.
How to measure it
- Question: "How satisfied were you with [the interaction]?", asked right after it.
- Scale: 1 Very unsatisfied, 2 Unsatisfied, 3 Neutral, 4 Satisfied, 5 Very satisfied.
- CSAT = (number of 4 and 5 answers ÷ total answers) × 100.
Common mistakes
- Mixing methods: an average of 1 to 5 scores and a share of 4s and 5s are different numbers. Pick one and keep it.
- Asking days later, when the moment is gone.
- Comparing CSAT from different touchpoints as if it were one number.
What changes in AI-first products
CSAT tells you which moment went wrong, close to where it happened. That's useful to an agent only with the comment attached. Annsa's in-product surveys and CSV imports keep each answer in the customer's words next to the priority it supports, so Claude Code, Cursor, Codex or Grok can read why a step scored low before changing it. You decide what changes.
See how Annsa does it: Surveys →