CSAT vs CES vs NPS: three metrics, three different questions

The three survey metrics get treated as interchangeable satisfaction scores, which is how a team ends up sending an NPS survey after a password reset and drawing conclusions about loyalty from the result. They measure different things on different time horizons, and only one of them belongs on an individual ticket.

CSAT asks about the interaction that just happened — how satisfied were you with this support experience — usually on a five-point scale, reported as the percentage choosing the top two boxes. It is immediate, specific, and scoped to one ticket, which makes it the only one of the three that can be attributed to a single agent or category with a straight face.

CES, the customer effort score, asks how hard the person had to work: the company made it easy for me to handle my issue, agree to disagree, typically on a seven-point scale. It measures friction rather than feeling. That is why it tends to predict what someone does next — call back, escalate, give up — better than satisfaction does. Someone can be pleased with a friendly agent and still have spent four days getting a laptop.

NPS asks about the relationship: how likely are you to recommend us, zero to ten. Nines and tens are promoters, sevens and eights are passive, zero through six are detractors, and the score is promoters minus detractors expressed as a percentage. With 100 responses, 50 promoters and 20 detractors give +30. Note what that arithmetic does: it discards the 30 passives entirely and spans a range from −100 to +100, which is why small samples move it violently and why month-to-month NPS movement on a few dozen responses is noise wearing a decimal point.

The strongest position in this article: do not survey NPS after individual tickets. NPS is a relationship metric that moves on product, price, and accumulated experience over months. Asking it moments after a support interaction produces a number dominated by that interaction — CSAT wearing an eleven-point scale — and that contaminated number then gets reported upward as company-level loyalty and used to justify decisions that have nothing to do with support.

It is a particularly poor fit for internal helpdesks, where would you recommend the IT department is a question with no action attached. Your colleagues cannot switch providers. Measure whether the interaction worked and how much effort it cost them; leave relationship measurement to a periodic survey run by whoever actually owns the relationship.

All three share one structural flaw: the people who answer are not a random sample of the people you served. Survey response skews to the extremes, because delighted and furious both supply motivation and the neutral majority supplies none, and it skews toward people with spare time. If one person in six responds, your score describes a self-selected sixth of the queue and says nothing reliable about the other five.

Three habits keep this honest. Track response rate immediately next to the score, and treat any change in response rate as a candidate explanation for a change in score before you look for real causes. Read the distribution rather than the average — a simultaneous rise in top-box and bottom-box responses is a completely different story from a gentle drift, and averaging erases the difference. And never compare your score to another company’s unless the wording, timing, and trigger are identical, which they are not.

Timing deserves its own warning. Surveying at the moment of close measures the close, not the fix — the requester is answering about how the conversation felt, before they have found out whether the problem stayed solved. If the resolution needs to hold, wait until it has had the chance to fail.

Collect CSAT continuously but sparsely — on a sample of tickets, not all of them — and read it by category rather than in aggregate. The aggregate number is a mood ring that drifts a point or two forever. The per-category cut is what tells you which part of the service is failing and who should hear about it.

Collect CES where you suspect friction and have the means to remove it: a slow approval chain, a category with heavy back-and-forth, a self-service flow you just rebuilt. CES earns its place as a before-and-after measurement on a specific process, not as a permanent dashboard tile that nobody knows how to move.

Collect NPS at the relationship level, quarterly or twice a year, for external customers where the concept genuinely applies. For an internal service desk, an annual pulse asking whether people can get their work unblocked will teach you more than a rolling NPS that nobody can act on.

The most common internal support mistake is surveying every ticket, forever. It produces falling response rates, survey fatigue that contaminates every other internal survey your company runs, and a score too noisy to move a decision. Employees who get a satisfaction request after every password reset stop answering all of them, including the one that mattered.

Sample instead. Take a fraction of tickets, weighted toward the categories you are actively trying to improve, and add a targeted survey after anything that reopened or escalated, where the information is genuinely worth the interruption. And whatever you collect, close the loop visibly at least once a quarter by telling people what changed because of their answers. Response rate is largely a function of whether answering has ever produced anything.

Survey design is upstream work, but the number still hangs on a ticket. A CSAT score attached to a record that was really three tickets, or to a close that nobody can trace, does not support much analysis — when a category dips, you want to read what happened rather than reconstruct it.

FlowTux keeps that record coherent: intake from Slack, Teams, email, and WhatsApp arrives in one queue, semantic deduplication attaches a repeat contact to the original thread instead of spawning fresh volume, and every triage and resolution step — including allow-listed autonomous resolutions — is written to the ticket timeline as a full audit trail. Per-category suggest, approve, and autonomous modes mean that when satisfaction moves in one category, you can see exactly how those tickets were handled.