What is CSAT?

CSAT (Customer Satisfaction Score) is the percentage of survey respondents who rated a specific interaction as satisfied. On the standard 5-point scale that means the 4s and the 5s: CSAT % = (4s + 5s) ÷ total responses × 100.

A worked example with real numbers. You resolve 400 conversations in a week and 92 customers answer the survey: 51 rated 5, 23 rated 4, 9 rated 3, 5 rated 2, and 4 rated 1. Satisfied responses = 51 + 23 = 74. CSAT = 74 ÷ 92 × 100 = 80.4%. The neutral 3s never enter the numerator but always stay in the denominator, so indifferent customers drag the score down exactly as hard as angry ones.

Two conventions decide what number comes out, and neither is a footnote:

• Top-two box vs. top-one box. Count only the 5s and the same 92 responses give 51 ÷ 92 = 55.4% — a 25-point swing from identical data. Zendesk's CSAT guidance uses the top-two-box convention, as does virtually every published benchmark, so switching silently makes your score incomparable to everyone else's and to your own history.
• Percentage vs. composite average. Averaging the raw ratings gives (51×5 + 23×4 + 9×3 + 5×2 + 4×1) ÷ 92 = 4.22 out of 5. Useful for spotting small internal shifts, useless for benchmarking, because industry figures are published as percentages.

What separates CSAT from its two neighbours is scope, not scale: CSAT is interaction-scoped, NPS is relationship-scoped, and CES is effort-scoped. The full arithmetic, including 10-point scales, is in our guide to calculating a CSAT score.

A CSAT number is only meaningful next to the right benchmark and with an honest read on how the survey itself bent the result. Two anchors worth memorising. The American Customer Satisfaction Index put US cross-industry satisfaction at 76.7 on its 100-point scale in Q1 2026 — down 0.3% year on year and identical to where it sat in 2013 (ACSI, 2026). Zendesk's CSAT guidance, last updated May 2026, treats 75–85% as good and above 90% as exemplary. If you are running 82%, you are ordinary, not outstanding.

The distortions that make a CSAT score lie, roughly in the order they bite small teams:

• Non-response bias. Zendesk lists this as a documented drawback of CSAT: when only a slice of customers answers, that slice skews toward the loyal and the enraged, and the quietly disappointed middle is simply absent from the denominator.
• Agent-selected sampling. If agents choose which conversations get a survey, the score measures agent confidence, not customer satisfaction. This is the single most common way a support team accidentally manufactures a 95%.
• Timing. A survey fired the instant a ticket is marked resolved measures relief; one fired after the customer has lived with the fix measures whether it actually worked. The two produce different numbers from the same team.
• Cultural response style. Zendesk also flags this: respondents from some cultures avoid the extremes of a scale entirely, so a multi-market score mixes rating habits with actual sentiment.
• Denominator drift. Dividing by surveys sent instead of surveys completed quietly reports a response rate wearing a CSAT label.

There is also a trap in reading a flat score as a healthy one. ACSI's Q1 2026 release recorded customer complaints surging 16% in the quarter while customer retention simultaneously rose — what it calls pent-up defection, where customers stay because switching is hard rather than because they are happy. A stable CSAT sitting on top of rising complaint volume is not a plateau; it is a warning that the metric has stopped tracking the thing you care about.

Design choices in how a survey is triggered change the number more than most improvement tactics do. Three that matter, drawn from how CSAT is implemented in Converge:

• Agent-triggered, not timer-triggered. The survey is offered in a confirmation dialog when an agent moves a conversation to resolved, and it is delivered as a message on the same channel the customer already used — the WhatsApp thread, the Messenger thread, the widget — rather than as a separate email that has to survive a spam filter. The trade-off is honest: this deliberately hands sampling control to agents, which is why agent-level response rates need watching alongside agent-level scores.
• One pending survey per customer. A second survey request while an earlier one is still unanswered is rejected outright. Without that guard, a customer with four tickets in a week gets four prompts, answers the one that annoyed them most, and your denominator quietly fills with your most contacted customers.
• A 48-hour embargo on individual responses. Aggregate scores are live, but individual ratings and comments stay hidden from the dashboard for 48 hours after submission, and are visible to owners and admins only. The reason is behavioural rather than technical: when an agent can tie a 2-star rating to a conversation they closed an hour ago, the incentive to nudge customers toward a rating starts to outweigh the incentive to resolve the ticket. The 48-hour figure comes from stacking the two industry norms — Zendesk sends surveys roughly a day after resolution, and customers typically take about another day to reply.

None of these raise a score by themselves. What they do is stop the score from being an artefact of the survey mechanism, which is a prerequisite for the number being worth acting on at all.