Last updated: September 12, 2026
What a CSAT score is
A CSAT score is the number of survey responses in the top boxes divided by the total responses received, multiplied by 100. The arithmetic is exact. What it is exact about is the group of callers who chose to answer, which on most floors is a small and self selected slice of the people who called you.
Say 2,000 people reach your floor in a month and 90 answer the survey. Sixty eight pick a 4 or a 5, so CSAT lands at 75.6%. That figure describes 90 callers, and the other 1,910 appear in your reporting as nothing at all.
Somebody fills out a survey when the call left a mark, and a caller who got what they needed in 40 seconds hangs up and goes back to work. That tilts the sample toward people with something to report, and no measurement of the tilt held up when we went looking for one with a publisher and a population attached, so treat it as a reason to read the number carefully rather than a correction you can apply.
Response rate belongs on the dashboard beside the score, in the same size type. A month that fell from 6% to 3% changed its own sample, and the score moved with it.
The formula, and the part people argue about
CSAT is responses in the top boxes, divided by total responses, times 100. Top box counts the highest rating alone. A top two box version takes the highest two, which is the usual choice on a five point card and the reason most published CSAT figures read generously.
Neutral answers are where the arguing happens. Take those same 90 responses: 68 at a 4 or a 5, 12 at a 3, and 10 at a 1 or a 2. Leaving the neutrals in the denominator gives 68 divided by 90, or 75.6%. Dropping them gives 68 divided by 78, or 87.2%. Same callers, same month, 11.6 points apart, and nobody reading the report can tell which version they were handed.
Write that rule down next to the number, since the QA score formula carries the same trap in a criterion marked not applicable.
Why the scale moves the number
Five point cards are the most common, with 1 to 10 cards and plain thumbs up or down surveys behind them. The top two boxes on a five point card cover the top 40% of the range, and on a ten point card they cover the top 20%. Identical goodwill from identical callers produces a lower percentage, so a program that switched scales last quarter and reported a drop has mostly measured its own switch.
Question wording does the same thing more quietly. "How satisfied were you with this call" and "did we resolve your issue" get filed under the same acronym, and the second one sits closer to first call resolution than to satisfaction.
The response rate, and the calls nobody surveyed
Response rate is responses divided by surveys delivered. The ratio worth tracking beside it is surveys delivered divided by calls handled, which is usually the bigger hole and the one nobody looks at. An IVR survey only reaches a caller who got to an agent and stayed on the line to the end. Whoever gave up in queue at four minutes never sees it.
Your survey comes back at 4.5% one month and 2.1% the next, with the score sitting still at 78%. Steadiness is the reassuring part and the part you should not act on. Half the responses disappeared, and whatever lean was in the first sample got stronger in the second.
Two habits make the number readable, and neither one removes the bias. Publish the response rate beside every score, broken out by delivery method and call type, since an email survey on a billing call and an IVR survey on a 40 second dispatch call reach different people. Then refuse to compare two months whose response rates sit far apart, the same way you would refuse to read a QA average whose coverage collapsed that week, which is the argument the QA metrics page makes about coverage rate.
What does not depend on the caller volunteering is the call itself, and scoring calls instead of sampling them reaches the callers a survey never touched. A transcript tells you what happened rather than how the caller felt an hour later, which is the thing the survey exists for, so keep both.
When you ask changes what you get
An IVR survey at the end of the call catches people while the interaction is fresh, and only the ones who stayed on the line. Email the next morning reaches more of them and collects the whole experience, including the parts your agent never touched.
Say an answering service takes a call for an HVAC account at 2am. The operator classifies it under that account's urgency definitions, follows the on call schedule in force at that hour, and dispatches to whoever is up. That survey fires at 7am, and what the caller scores is whether somebody came out. Reading the callback number back digit by digit is invisible to them, and it decided whether the tech had a number to call at all.
A per call survey on work that finishes hours later measures the dispatch as much as the call, which is a reason to score the call separately rather than a reason to drop the survey.
What a good score looks like when the sample leans
CSAT has no row on the call center benchmarks page, which carries every figure traced back to a publisher, a date, and the population it was measured on. Figures for this metric circulate widely, and the ones we chased arrived with no population attached.
Your own trailing months are the usable target, held to the same scale, wording, timing, and delivery method, with the response rate printed next to each one. A jump that lines up with a survey change is a survey change. Read the spread rather than the average, which is the argument a good QA score rests on too.
Transactional and relationship measurement
A transactional survey fires off one interaction and traces back to a specific call, agent, and timestamp, which is what makes it useful for coaching. You can set it beside the recording and hear what the caller was reacting to.
Relationship surveys ask about the whole arrangement on a schedule of their own, usually quarterly, with no call attached. NPS lives on that side, since "would you recommend us" asks about the company rather than about the 40 seconds somebody just spent on the phone. Putting the two beside each other and calling the gap a trend compares different objects.
An answering service or a BPO has a third opinion in the room, since the client account paying you renews the contract. CallSniper reports dashboards by agent and by client account for that reason, rather than one blended floor average.
When CSAT and the QA score disagree
Both numbers describe the same calls, and they diverge often. That divergence is the most useful thing on either report, as long as nobody reconciles it by quietly adjusting one of them.
A high QA score with low satisfaction usually means the card is measuring process compliance and missing whatever the caller cared about. The agent hit the greeting, captured the callback number, stayed inside the account's do not say list, and the caller still had to ring back twice or waited nine minutes before anybody picked up. Neither shows up on the scorecard. Low scores with contented callers point the other way, at a card penalizing something callers do not mind, or at one reviewer being stricter than the rest.
The investigation is ordinary work. Check calibration variance first, since reviewer drift produces the same symptom on its own. Then pull the ten calls with the widest gap between the two numbers and read those transcripts end to end, because reading criterion by criterion is how a card's blind spot stays invisible. Every complaint a caller raised that no criterion covers goes on a list, and that list is the next revision of your card, with the agent conversations behind it belonging in coaching.
One limit before you go looking. Machine scoring holds up on rule based criteria and gets shakier on tone and empathy, which Aircall and StackAI both note, so adding an empathy criterion to close a CSAT gap produces arguable scores more often than answers. Every scored line in CallSniper carries a transcript timestamp and a confidence score, so a supervisor knows which flags deserve their time.
A call review scores a batch of your own recent recordings against the scorecard you already use, so you can set those scores beside the surveys you already collected and find the calls where the two disagree.
Frequently asked questions
How do you calculate a CSAT score?
Divide the responses in your top boxes by the total responses received, then multiply by 100. Most programs count the top two boxes on a five point card and leave neutral answers in the denominator, and whichever rule you pick belongs on the report.
What is a good CSAT score?
No figure for it traces back to a named publisher with a stated population, so this site carries none. Build the target out of your own trailing three months on the same scale and the same question wording.
What is the difference between CSAT and NPS?
CSAT asks about one interaction and ties back to a specific call and agent. NPS asks how likely somebody is to recommend the company, which covers the relationship over time, so a gap between the two describes the survey design more than the month.
Our QA scores are high and our CSAT is low. What does that mean?
Usually that the scorecard measures compliance with a process and misses something the caller cared about, such as a callback that never came. Pull the calls with the widest gap between the two numbers and read those transcripts in full.
Sources
- Aircall and StackAI on rule based criteria scoring more reliably than tone. Third party, unverified.
- Trinsic feature sheet and call notes, 2026-08-27 and 2026-09-03. Verified internally.
- The 2,000 call example, the neutral response arithmetic, and the scale comparison. Illustrations built for this page, not measurements.
- CSAT benchmarks. No figure with a named publisher and a stated population is carried on this site as of 2026-09-12.