What Is CSAT? Customer Satisfaction Score Formula and Benchmarks

CSAT is the support metric most people meet first: the stars under a chat, the faces in a follow-up email, the "How did we do?" after a call. The arithmetic is simple, yet two teams can report "77%" and "4.08 out of 5" from exactly the same answers, and both are correct.

This guide is for support leads, founders and students who need the definition, the formula and a fair way to judge a score: worked calculations you can copy, published benchmarks with their methods, survey wording, CSAT next to NPS and CES, CSAT for AI agents, and the mistakes that make the number mislead.

What is CSAT?

CSAT, short for customer satisfaction score, is a metric that asks customers one question right after an interaction, such as "How satisfied were you with the help you received?", and reports the share who answered at the satisfied end of the scale. It describes one moment, not the whole relationship, and is usually given as a percentage.

What does CSAT mean, and where did it come from?

CSAT means customer satisfaction score, and no single person or company invented it: businesses have surveyed satisfaction for decades, and national satisfaction indexes followed, Sweden's in 1989 and the American Customer Satisfaction Index in 1994.

According to ACSI's history page, the American Customer Satisfaction Index was started in 1994 by researchers at the University of Michigan with the American Society for Quality and CFI Group, and was first published in October 1994. Its model came from the Swedish Customer Satisfaction Barometer of 1989, and Claes Fornell developed both.

For a formal process, the International Organization for Standardization publishes ISO 10004:2018, guidelines for monitoring and measuring customer satisfaction, last reviewed and confirmed in 2023. There is still no official CSAT question: each company or software tool chooses its own wording and scale, which is the main reason CSAT figures from different places rarely compare. Net Promoter Score is different: it has a named origin, Fred Reichheld's "The One Number You Need to Grow" in Harvard Business Review in December 2003, and one fixed question.

How do you calculate CSAT?

You calculate CSAT by dividing the number of satisfied responses by the total number of responses and multiplying by 100.

CSAT (%) = (satisfied responses ÷ total responses) × 100

Satisfied responses ("top two box"):
  1 to 5 scale  ->  answers of 4 or 5
  1 to 7 scale  ->  answers of 6 or 7
  thumbs        ->  thumbs-up

Top-two-box is the market-research name for counting only the two most positive points on a scale as satisfied. On a 1 to 5 scale that means 4 (satisfied) and 5 (very satisfied). Zendesk's CSAT documentation uses the same split for its five-point surveys: 1 to 3 is bad and 4 to 5 is good.

Two details change the result. The denominator is responses, not surveys sent, so a customer who ignored the survey is not in it. And a neutral 3 stays in the denominator but not in the numerator, so it lowers CSAT without counting as a complaint.

Which answers count as satisfied?

The scale decides what "satisfied" means, so write the rule down next to every score you report.

CSAT calculator: how do you work out your score?

To work out a CSAT score, count the satisfied answers, count all answers, divide the first by the second and multiply by 100. The examples cover a five-point survey, thumbs ratings and a small sample, then the average-score method.

Example 1: a 1 to 5 survey after support tickets

A team emails a one-question survey after each solved ticket and receives 200 answers in a month.

Answers: 5 = 96, 4 = 58, 3 = 22, 2 = 14, 1 = 10      (total 200)

Step 1  Satisfied answers   = 96 + 58          = 154
Step 2  Divide by all       = 154 ÷ 200        = 0.77
Step 3  Multiply by 100     = 0.77 × 100       = 77%

The team's CSAT is 77%. The single number hides that 22 neutral answers (11%) pulled it down, while only 24 people, 12%, chose 1 or 2.

Example 2: thumbs ratings in a chat widget

A website chat shows a thumbs-up and a thumbs-down when each conversation ends. In a month, 2,600 conversations end and visitors leave 400 votes: 312 up and 88 down.

CSAT           = 312 ÷ (312 + 88) × 100 = 312 ÷ 400 × 100 = 78%
Response rate  = 400 ÷ 2,600 × 100                        = 15.4%
95% margin     ≈ 1.96 × √(0.78 × 0.22 ÷ 400)              ≈ ±4.1 points

Report it as "78% from 400 votes, 15.4% of conversations". The true figure is probably within about four points of 78%, so next month's 76% on a similar number of votes is not evidence of a decline.

Example 3: a small sample, where one vote moves the score

A small business gets 11 ratings in a month: 9 thumbs-up and 2 thumbs-down.

CSAT                          = 9 ÷ 11 × 100  = 81.8%
With one more thumbs-down     = 9 ÷ 12 × 100  = 75.0%
95% range for 9 of 11 (Wilson method)         ≈ 52% to 95%

One vote moved the score by almost seven points. With a dozen ratings the honest summary is "9 of 11 positive", not a percentage with a decimal point. Pool several months, or read the conversations themselves, before acting on a change.

The average-score method, and why it gives a different number

Some tools report the mean rating instead of the share satisfied. The 200 answers in Example 1 average 4.08 out of 5:

Average = (5 × 96 + 4 × 58 + 3 × 22 + 2 × 14 + 1 × 10) ÷ 200
        = 816 ÷ 200 = 4.08 out of 5
As a share of the maximum: 4.08 ÷ 5 × 100 = 81.6%

So the same answers can be reported as "CSAT 77%", "4.08 out of 5" or "81.6%", and the methods can even move in opposite directions. Suppose a second team's 200 answers are 130 fives, 10 fours, 50 threes, 5 twos and 5 ones. Its top-two-box CSAT is 140 ÷ 200 = 70%, below the first team's 77%, yet its average is 855 ÷ 200 = 4.275, above 4.08. The average rewards a crowd of fives and hides a block of neutral threes.

Neither method is wrong; the UK Customer Satisfaction Index, for one, takes the mean of customers' ratings and multiplies it by ten. Pick one, label it, and never compare it with a number calculated the other way.

What is a good CSAT score?

There is no universal good CSAT score: published figures range from 64.2% for rated live chats (LiveChat, 2024 data) to a national index of 78.3 out of 100 in the UK (July 2026), and they are measured too differently to compare directly. A good score is one that holds or rises against your own history on the same question, scale and channel.

Figures from each source's own pages, checked 12 November 2026.

The closest comparison for a chat widget is LiveChat's Customer Service Report: 64.2% across rated chats and 64.7% for chats handled by chatbots. It covers one vendor's customers and does not describe the rating scale, so treat it as a rough reference, not a target.

The American Customer Satisfaction Index publishes a national score for the US every quarter and industry scores through the year, all on a 0 to 100 scale. Several of its industry reports also score support itself, such as call centre satisfaction for airlines and internet providers and the helpfulness of customer support for online retailers. ACSI limits how its figures may be reused, so read the current ones on its own site instead of a copy here.

The UK Customer Satisfaction Index, published every six months by the Institute of Customer Service, stood at 78.3 out of 100 in July 2026, one point higher than a year earlier. For satisfaction and expectation figures about AI service in particular, our AI customer service statistics roundup lists each number with its source and year.

How do you run a CSAT survey? When to send it and what to ask

Run a CSAT survey by asking one question as soon as the interaction ends, in the same channel, on a scale you never change, and store each answer with its date, channel and topic.

A customer satisfaction survey is any questionnaire that asks customers how satisfied they are; a CSAT survey is its shortest form, one rating plus an optional comment box. Longer customer feedback surveys belong in research projects, not after every ticket.

When to send it

• Chat: in the chat window as soon as the conversation ends, while the customer is still there.
• Email and tickets: when the ticket is marked solved, so the customer rates an experience they still remember.
• Phone: an automated question at the end of the call, or a text message shortly after.
• Orders: after delivery rather than at checkout, unless the checkout is what you want to measure.

Ask once per interaction, and limit how often one customer is asked.

How to word the question

Use neutral wording, label both ends of the scale, ask about one thing and keep the same words in every channel. These five questions cover most situations:

1. How satisfied were you with the help you received today?
   (1 = Very dissatisfied, 5 = Very satisfied)
2. How was this conversation?  (thumbs up / thumbs down)
3. Overall, how satisfied are you with how we handled your request?
   (1 = Very dissatisfied, 5 = Very satisfied)
4. How satisfied are you with your order from [store name]?
   (1 = Very dissatisfied, 5 = Very satisfied)
5. How satisfied were you with the support you received from [agent name]?
   (1 = Very dissatisfied, 5 = Very satisfied)

Optional follow-up, shown after the rating:
What is one thing we could have done better?

Avoid leading words ("How great was..."), double questions ("How fast and friendly was...") and any request to "give us a 5". The optional follow-up is where the reasons live.

What is a good CSAT response rate, and why does response bias matter?

Most customers never answer a CSAT survey, so the score describes the people who chose to rate: in LiveChat's 2024 platform data, only 6.7% of chats received a rating.

Response rate is the share of customers asked who actually answered: ratings divided by surveys offered, times 100, as in Example 2 (400 ÷ 2,600 = 15.4%). There is no standard good response rate, so track your own and treat a sudden change as a change in who is answering.

Response bias, more precisely non-response bias, is the distortion that appears when the customers who answer differ from those who do not. People with a strong experience, good or bad, have more reason to answer than people who were simply fine, and customers who gave up halfway may never see the question. The bias can push the score either way, and arithmetic cannot remove it.

What helps: ask every eligible customer, keep it to one tap, report the response rate beside the score, and compare the topics of rated conversations with all conversations. If the rated ones are mostly easy questions, your CSAT flatters you.

CSAT vs NPS vs CES: what is the difference?

CSAT measures satisfaction with one interaction, NPS measures how likely someone is to recommend you overall, and CES measures how easy it was to get something done.

Bain's Net Promoter System page defines NPS as the percentage of promoters (9 or 10) minus the percentage of detractors (0 to 6), so it runs from minus 100 to plus 100.

The customer effort score (CES) comes from "Stop Trying to Delight Your Customers" by Matthew Dixon, Karen Freeman and Nick Toman in Harvard Business Review, July to August 2010. Their study of more than 75,000 people dealing with contact centres or self-service found that over-the-top service made little difference, and they argued that effort predicts loyalty better than satisfaction measures or NPS. Use CSAT for the interaction, CES for the process and NPS for the relationship, and resolution rate and response time to explain why any of them moved.

How does CSAT work for chatbots and AI agents?

For a chatbot or AI agent, CSAT is usually a one-tap rating of the whole conversation shown when the chat ends, and it needs extra care because fewer people rate, the conversations differ from a human team's, and a handoff splits the credit.

The rating covers the whole conversation

A visitor who got a good AI answer and then waited a day for a person rates the day, not the answer. Report CSAT separately for conversations the AI finished alone and for handed-off ones, or you cannot see which part needs work. In PepoChat, for example, the rating prompt appears once a conversation is resolved, whether the AI support agent or a teammate resolved it, so a vote after a human handoff covers both.

Fewer people rate, and the wrong people skip it

Chat visitors leave the moment they have what they need, and a rating shown at the end misses everyone who gave up halfway, often the least happy. Read CSAT for an AI agent beside resolution rate under a strict definition; what counts as a chatbot resolution compares eight vendors' definitions.

Why a bot's CSAT and a human team's are not comparable

An AI agent answers the easy, repetitive questions; people receive the escalations, which are harder and often angrier, so a human team's CSAT can look worse for reasons that have nothing to do with skill. A correct but unwelcome answer, such as "returns close after 30 days", earns a thumbs-down whoever delivers it. LiveChat's 64.7% for chatbot-handled chats and 64.2% for all rated chats come from different conversations, so they are not a head-to-head test either. Compare the AI with itself over time, topic by topic, and use the other measures in AI chatbot metrics that matter to see what the rating cannot.

How do you improve your CSAT score?

You improve CSAT by fixing the reasons behind negative ratings, mostly unsolved problems, slow replies, repeated effort and unclear policies, not by asking more nicely for good scores.

1. Read the conversations behind every negative rating, grouped by topic. A monthly transcript review turns the top causes into content fixes.
2. Fix wrong or missing answers at the source. An unclear help page spreads to every channel that quotes it.
3. Solve it the first time. Reducing repeat contacts is the first tactic the HBR effort authors recommend.
4. Shorten the wait where it is long. First response time benchmarks show what each channel's customers expect.
5. Make reaching a person easy, and say when they will reply. "We reply by email within one business day" beats a silent queue.
6. Explain unwelcome outcomes with the reason and the next option.
7. Follow up on low ratings. A short reply often fixes the problem and tells you what the rating did not.
8. Do not chase the number. Asking for fives or surveying only easy tickets raises the score and destroys its meaning.

What are the weaknesses of CSAT?

CSAT's main weaknesses are that it describes only the people who answer, only the moment they were asked and only on the scale you chose, so it is easy to misread and hard to compare.

• It is self-selected. Few customers answer, and they are not a random sample.
• It measures a moment, not loyalty. A customer can rate a chat 5 and still leave over price; the HBR effort research argues effort predicts loyalty better.
• It punishes correct "no" answers. Policy outcomes collect low scores even when the service was good.
• It has a ceiling. Near 90%, small changes look meaningful when they are noise.
• It is easy to game. Whoever chooses who gets surveyed can move the score.
• It does not travel. An email survey, a chat thumbs-up and a national index all produce "satisfaction" numbers that cannot be lined up.

That makes CSAT a trend line and a pointer to conversations worth reading, not a grade.

What are the most common CSAT mistakes?

The most common CSAT mistakes are averaging percentages from periods of different sizes, counting neutral answers as satisfied, changing the question or scale mid-year and comparing your number with a benchmark measured another way.

The first one is the hardest to spot. Here is a month with one quiet week and one busy week:

Week 1:  9 satisfied of 10 responses    = 90%
Week 2:  133 satisfied of 190 responses = 70%

Wrong: average of the two percentages = (90 + 70) ÷ 2 = 80%
Right: pool the responses = (9 + 133) ÷ (10 + 190) = 142 ÷ 200 = 71%

The quiet week counted as much as the busy one, and the month looked nine points better than it was. Add up the satisfied answers and all answers, then divide once; the same applies to combining agents, channels or regions.

Other mistakes to avoid:

• Counting 3 out of 5 as satisfied. It inflates CSAT and breaks comparison with your history.
• Changing the wording or scale. Start a new series instead.
• Mixing channels in one number. Chat, email and phone customers are asked at different moments.
• Dropping the response count. A percentage without its count invites action on noise.
• Comparing with ACSI or UKCSI. They are indexes, not survey percentages.
• Treating CSAT as sentiment. Ratings and AI-inferred mood measure different things.

How PepoChat uses CSAT

PepoChat is an AI support agent for websites, and it collects a CSAT-style rating on conversations that reach the end. When a conversation is resolved, by the agent after the visitor says they are done or by a teammate in the team inbox, the widget asks "How was this conversation?" with a thumbs-up and a thumbs-down.

The vote appears as a "Rated good" or "Rated poor" badge on the conversation in the team inbox, so you can open the transcript behind any thumbs-down. Analytics shows "Positive ratings", the thumbs-up share of all votes with both counts, for 7, 30 or 90 days, plus a 90-day Sentiments card; test chats are left out. The REST API returns each conversation's rating as up, down or null, so you can calculate CSAT by topic or by handoff yourself.

The limits are real. PepoChat has no survey builder, no 1 to 5 scale, no comment box, no email survey, no NPS or CES, and no AI sentiment scoring. A visitor who leaves before the conversation is resolved is never asked, and the CSV export has no rating column. PepoChat publishes no CSAT figures of its own.

Ratings and analytics are on every plan, including the free plan with 100 AI replies a month; create a free workspace to see your own numbers, or compare the paid plans on the pricing page.

What to do next

• This week, pull the last 90 days of ratings, calculate CSAT with its response count and response rate, and read the ten most recent negative conversations.
• If you also want AI-inferred mood, compare the options in customer sentiment analysis for support chats.
• See exactly what PepoChat's ratings cards show in the Analytics documentation.
• Start a free PepoChat workspace and watch thumbs ratings arrive on your own conversations; every chat feature is on the free plan.
• Related explainers: What is NPS?, What is an SLA? and What is a help desk?.