CSAT vs NPS: which survey to send, when, and 14 question templates

Twenty customers answer two questions. Seventeen of them rate their last support conversation four or five stars, so your CSAT is 85%. Asked whether they would recommend you to a friend or colleague, the same twenty produce a Net Promoter Score of minus 5.

Both numbers are correct. They describe different things, which is the whole story of CSAT vs NPS. This guide gives you the definitions and formulas side by side, one worked example you can check by hand, a rule for which survey to send and when, 14 question templates you can copy, the sample sizes you need before a number means anything, and the research critique of NPS with its sources. Talvo is the worked example for the setup, and the advice holds for any tool.

CSAT vs NPS vs CES, side by side

NPS comes from a December 2003 Harvard Business Review article by Frederick Reichheld, "The One Number You Need to Grow". Its scoring is fixed: promoters are 9 and 10, passives are 7 and 8, detractors are 0 to 6, and the score is the percentage of promoters minus the percentage of detractors (Bain's Net Promoter System). Passives count in the total but not in the score. Net Promoter, NPS and Net Promoter Score are registered trademarks or service marks of Bain & Company, NICE Systems and Fred Reichheld.

CSAT has no such owner, and that matters, as the next section shows.

One set of 20 customers, three CSAT scores and one NPS

These twenty pairs of answers are made up to show the arithmetic. Each customer rated their last support conversation from 1 to 5 stars and answered the recommend question from 0 to 10.

Now the scores.

Three observations.

First, the same raw ratings give you 4.30, 85% or 60%, depending on the convention. CSAT has no standard formula. Many tools report the percentage of 4s and 5s; some report the average; some count only the top box. A 4.3 from one tool and an 85% from another are the same data. Never compare scores across tools without checking the formula.

Second, the customers who scored you well on the conversation are not the ones willing to recommend you. Of the 17 who were satisfied, at most 5 can be promoters, because only 5 customers answered 9 or 10. A conversation can be handled perfectly and the customer can still feel the product is too expensive or too hard to use.

Third, this is not a defect in either measure. CSAT tells you whether your support team did its job. NPS tells you something about the whole relationship, only some of which support controls.

Which to send, and when

The decision rule is short. Send CSAT right after a conversation ends, because it asks about that conversation and fades from memory fast. Send NPS on a schedule, never after every chat, because it asks about everything and a bad afternoon in support shouldn't swing a relationship score.

If you send both, keep them apart in time. A customer who rated their chat five stars on Tuesday shouldn't be asked the recommend question on Wednesday. A sensible small-team rhythm, once you have cut the repeat questions and are hitting your first response time targets, is a CSAT prompt on every resolved conversation, an NPS survey once a quarter to customers who have been with you more than a month, and a short CES question after anything that tends to be painful, such as a refund or a data import.

Question templates you can copy

Fourteen templates, grouped by what they measure. Keep each survey to one scored question plus one optional comment. In our view, a third question is where response quality tends to fall; SurveyMonkey's guidance also warns about survey length.

CSAT templates

NPS templates

Open-ended and effort templates

Neutral wording matters more than clever wording. "How great was our support?" pushes people toward a high answer. "How satisfied were you" doesn't. Keep the scale direction the same every time, label the ends, and don't change the scale partway through a trend line.

How many responses before a number means anything

A percentage from twenty answers is closer to an anecdote than a measurement. For a score like 85% satisfied, the rough 95% margin of error is about plus or minus 16 points at 20 responses, 10 points at 50, 7 points at 100 and 3.5 points at 400. NPS is noisier because it combines two percentages: for the split in our example, the margin is roughly plus or minus 32 points at 20 responses, 21 at 50, 14.5 at 100 and 7 at 400.

These come from the standard normal approximation, which is rough at small counts, and they assume the people who answered are a fair sample of everyone. They usually aren't. People with a strong opinion answer more often. So treat the margins as the best case.

What this means in practice: a small team sending NPS to 60 customers a quarter should expect the score to jump around by 10 or more points for no reason. Look at the comments, the trend across several quarters, and the distribution. Don't celebrate a four-point gain or panic over a six-point drop.

What counts as a good score

Honestly: your own trend beats any benchmark. Still, these are the sourced numbers, each with its limit.

You will often read that a CSAT above some percentage in the 70s or 80s is "good". We looked for a primary source for that rule and didn't find one, so we don't repeat it. A software company with a median NPS of 30 and a CSAT of 4.3 stars isn't failing; a company with a CSAT that fell by half a star in a quarter has a problem worth finding.

Split CSAT by who answered

If you run an AI agent, the most useful cut of your satisfaction data is AI-resolved conversations against human-resolved ones. An AI that closes more conversations while lowering satisfaction on the ones it closes is moving work around, not helping. A billed resolution is not proof that the customer was happy: see AI resolution pricing for what vendors count as a resolution, and AI resolution rate for how to measure the share your AI really closes.

This is the number to check before you turn automation up. The measurement side of that is covered in the knowledge base chatbot guide. In Talvo, the CSAT on a conversation shows whether the AI or a person handled it, CSAT averages appear in Analytics next to volume and response times, split between AI-resolved and human-resolved conversations, and the AI performance dashboard on the Scale plan ($299 a month) charts that split alongside the hand-off funnel.

The honest critique of NPS

NPS became popular on a strong claim: that this one question predicts a company's growth better than other measures. Reichheld's 2003 article is where that began. The follow-up research is less kind.

Keiningham, Cooil, Andreassen and Aksoy examined 21 firms using more than 15,500 interviews from the Norwegian Customer Satisfaction Barometer. Their 2007 paper in the Journal of Marketing, "A Longitudinal Examination of Net Promoter and Firm Revenue Growth", failed to replicate the claim that NPS was clearly superior to other measures such as the ACSI. It won the 2007 MSI/H. Paul Root Award (record). Separately, a MeasuringU re-examination found that Reichheld's original data correlated NPS with past growth rather than future growth.

In May 2021, Gartner predicted that by 2025 more than 75% of organizations would have abandoned NPS as a measure of success for customer service and support (Gartner). That didn't happen, at least not among survey-tool customers: Survicate reported that 67.7% of its paying customers still ran NPS surveys in 2025 (Survicate). That is vendor data, but it shows the habit outlasted the forecast.

So what do you take from this? Use NPS for what it reliably does, which is give you one consistent question to trend and a prompt for comments. Don't use it as proof that you'll grow, and don't set a support team's bonus on it. For judging support, CSAT, the AI-versus-human split and the comments tell you more.

Running them in Talvo

Talvo has two surveys. The first is automatic: when a conversation is marked resolved, the customer sees a short star-rating prompt in the messenger with an optional comment. It appears once per conversation, and the rating lands on the thread, so you can see the score beside the conversation that earned it. The second is standalone surveys you build in Settings → Automation, with question types for NPS, rating, choice and text. Each gets its own public page you can link from an email, a chat or your site. The post-resolution star rating is on every plan, including Free. Standalone NPS, rating, choice and text surveys are included from the Starter plan ($49 a month, or $39 billed annually).

1. 1Resolve conversations only when finishedThe rating prompt appears automatically once per resolved conversation, with an optional comment, so closing a thread to clear the queue produces a rating request for it. See CSAT and NPS.
2. 2Build an NPS surveyIn Settings → Automation, add an NPS question and the matching follow-up from template 9, then share the page link in an email or on your site.
3. 3Map the other templatesUse rating for CSAT-style questions, choice for the yes/partly/no questions, and text for the open-ended ones.
4. 4Read the resultsResponses are listed individually, so a number you don't like leads back to the specific answers behind it.
5. 5Pair scores with themesCompare satisfaction with what customers are asking about; on the Scale plan, the Voice-of-Customer digest groups resolved conversations by theme and sentiment.

Where Talvo falls short: the help doesn't describe a built-in customer effort score, so template 13 would use a rating question on a standalone survey. The star-rating prompt shows an average rather than a percentage satisfied, and the help doesn't document triggering a standalone survey automatically at a point in the customer journey, so you send the link yourself. And Talvo is new and in beta: it has no customer benchmark data of its own, and none appears in this guide.