Voice of the Customer Questions: Surveys, Metrics and a Template
Voice of the customer survey and interview questions grouped by journey stage, with NPS, CSAT and CES formulas, wording mistakes to avoid, and a template.

Voice of the customer (VoC) questions are the questions you put to customers to learn what they need, what they think of you, and why. Below is a question bank grouped by journey stage, with the reason each question works. It also covers the three standard metric questions and their formulas, the wording mistakes that spoil answers, and a template you can copy.
Survey or interview
A survey tells you how many customers feel a certain way. An interview tells you why. Most VoC programmes need both, and the questions differ.
| Survey | Interview | |
|---|---|---|
| Answers | Scores and short text from many customers | Stories and reasons from a few customers |
| Question style | Fixed wording, mostly closed, one or two open questions | Open questions about specific past events, with follow-up probes |
| Use it to | Track a metric, compare segments, spot a change | Find causes, learn the customer's own words, test a new idea |
| Main risk | Wording and order shape the answers | Leading questions from the interviewer |
One approach from survey research combines them. Pew Research Center sometimes runs a pilot with open-ended questions to find the most common answers, then turns those answers into options for a closed question.[7]
The three metric questions
Most VoC surveys are built around one of three scored questions. Use the original wording and scale, because a changed scale gives a number you cannot compare with anyone else's.
| Metric | Question and scale | How to calculate it | Measures |
|---|---|---|---|
| Net Promoter Score (NPS) | "How likely are you to recommend us to a friend or colleague?" 0 to 10[1] | % promoters (9 or 10) minus % detractors (0 to 6). Passives (7 or 8) count in the total but not in the score.[1] | Loyalty to the company |
| Customer satisfaction (CSAT) | "How would you rate your overall satisfaction with the [product/service] you received?" 1 (very unsatisfied) to 5 (very satisfied)[2] | Number of 4 and 5 answers divided by all answers, times 100[2] | Satisfaction with one interaction or product |
| Customer Effort Score (CES) | "[Company] made it easy for me to handle my issue." 1 (strongly disagree) to 7 (strongly agree)[6] | Usually the average of all answers; higher means less effort | How easy a task or service contact was |
- NPS comes from Bain & Company, which calls the recommend question "the ultimate question".[1] Several of the pages that rank for this topic print it on a 1 to 10 scale. That scale has no zero, so the results do not match the standard score.
- CSAT has no single owner, and the top-two-box formula above is the common convention.[2] The American Customer Satisfaction Index (ACSI), run since 1994, uses a fuller method. It asks three questions on a 1 to 10 scale: overall satisfaction, how the experience compared with expectations, and how close it came to the ideal. The results are weighted by ACSI's model and reported on a 0 to 100 scale.[3]
- CES was introduced by Matthew Dixon, Karen Freeman and Nick Toman of CEB in the Harvard Business Review in 2010.[4] CEB's later research covered more than 125,000 customers. It found that 96% of customers who had a high-effort experience said they were more likely to be disloyal, against 6% after a low-effort one.[5] CEB published a new version of the score with that research in 2013.[5] The 1 to 7 agreement statement in the table is the form Medallia uses in its VoC questions.[6]
A score alone does not say what to fix. Follow every metric question with one open question: "What is the main reason for your score?" CustomerGauge and Checkbox both pair the NPS question with this kind of follow-up.[13][14] Ask for the main reason, not every reason, so the answer stays short enough to code.
Survey questions by journey stage
Ask each question at the moment the customer can answer it from fresh memory. The sets below are grouped by stage. Pick two to four from one stage for a short triggered survey, and do not run every set at once.
After a purchase or sign-up
- "Did you find everything you were looking for today?" (Yes / No, then "What was missing?" if No). It finds gaps in the range or the site that sales data cannot show, because a customer who left without buying leaves no record.
- "What almost stopped you from buying?" (open). It names the barrier while the customer still remembers it.
- "Which other options did you consider?" (list of competitors plus "Other" and "None"). It shows who you compete with in the customer's mind, which may not match your own list.
- "How did you first hear about us?" (list plus "Other"). It checks your attribution data against what customers remember.
Onboarding and first use
- "What were you hoping to get done when you signed up?" (open). Compare the answers with what customers actually use. A large gap means your marketing promises one job and your product does another.
- "How easy was it to get started?" (1 to 7, very difficult to very easy). It gives an effort score for one stage. Use the same scale each time so you can compare cohorts.
- "What, if anything, was confusing in your first week?" (open). "If anything" lets the customer say nothing was confusing without disagreeing with the question.
Product use
- "What do you mainly use [product] for?" (open, or a list built from earlier open answers). It tells you which jobs to protect when you change the product.
- "How would you feel if you could no longer use [product]?" (very disappointed / somewhat disappointed / not disappointed). It measures how much customers depend on the product, which a satisfaction score does not show.
- "What is the one thing we could change that would make [product] more useful to you?" (open). Asking for one thing gives you a ranked signal instead of a wish list.
Support and service
- "Did we solve your problem today?" (Yes / No). It measures first contact resolution from the customer's side.[6]
- The CSAT or CES question from the table above. Medallia recommends sending these immediately after the interaction.[6]
- "Before contacting us, did you try to solve this using our help articles?" (Yes / No). It shows whether your self-service content is found and whether it works.[6]
- "How many times did you contact us about this issue?" (Once / Twice / Three or more times). Repeat contact is a direct sign of effort.
Renewal, downgrade or cancellation
- The NPS question and its follow-up. Run it on a schedule, not after a single event, so it measures the relationship.
- "What is the main reason you are cancelling?" (list of likely reasons, plus "Other"). Build the list from earlier open answers and keep the options mutually exclusive.[8]
- "What would have made you stay?" (open). It separates problems you can fix from customers who were never a good fit.
- "May we contact you to talk about your answers?" (Yes / No). It recruits interviewees from the customers with the most to tell you.
Interview questions
An interview uses fewer questions and more follow-ups. Nielsen Norman Group recommends a few open questions that get people talking, and asking about specific events rather than general processes.[10] A customer who describes "the last time" gives you a real sequence of steps. A customer asked what they "usually" do gives you a general summary instead.
- "Tell me about the last time you [did the task your product helps with]."
- "What were you using before us, and what made you look for something else?"
- "Walk me through what happened when you first set it up."
- "What was the hardest part of that?"
- "If you had to explain to a colleague what we do, what would you say?"
- "What did you expect to happen that didn't?"
Use neutral probes such as "Tell me more about that" and "Why is that important to you?"[10] Nielsen Norman Group warns that many leading questions compromise the validity of the interview.[10] Our guide to understanding your ideal customer uses the same kind of interview to define who your best customers are.
Wording mistakes to avoid
Small wording changes move answers. In one Pew Research Center survey, 58% of respondents chose the economy as the most important issue when it was on a list, but only 35% named it when the question was open.[7]
| Mistake | Example | Better |
|---|---|---|
| Leading question | "How much did you enjoy our new checkout?" | "How easy or difficult was the new checkout to use?" |
| Double-barrelled question | "How satisfied are you with our product and support?" | Two questions, one about the product and one about support[7][8] |
| Agree/disagree statement where a choice works | "Our prices are fair. Agree / disagree" | "For what you get, are our prices too high, about right or too low?"[7] |
| Missing options | A list of reasons with no "Other" | Add "Other (please say)" and, where it fits, "Don't know"[8] |
| Scale with no zero for NPS | Recommend question on 1 to 10 | 0 to 10, as Bain defines it[1] |
| Specific question before a general one | Ask about delivery, then about overall satisfaction | Ask the general question first[8] |
- Agree/disagree items. Pew found that respondents who are less educated or less informed agree with statements more often, and the effect is stronger when an interviewer is present. Pew recommends offering a choice between alternative statements instead.[7] CES is an exception. Its agreement statement is the standard form, so keep it if you want comparable scores.
- Order effects. Earlier questions set the context for later ones. When closed questions come before an open one on the same topic, respondents mention the ideas from the closed questions more often.[7] Put your open "main reason" question straight after the score, before any list of reasons.
- Option order. In surveys people read themselves, respondents tend to pick items near the top of a list. Randomise lists of reasons, but keep rating scales in their natural order.[7]
- Changing the question. AAPOR's rule is "If you want to measure change, don't change the measure."[8] Once a tracked question is live, keep its wording, scale and position.
How to ask: length, timing, channel and sample size
- Length. Vendor guidance ranges from six questions or fewer (CustomerGauge) to six to ten questions that take about two minutes (Checkbox).[13][14] A transactional support survey can be two or three questions: the score, the reason and the resolution question.
- Timing. Send transactional surveys right after the event, such as a support contact or a delivery.[6] Run relationship surveys such as NPS on a fixed schedule, such as every quarter.[13] Checkbox suggests waiting 7 to 14 days after a customer starts using a feature before asking about it.[14]
- Channel. Ask in the channel the customer just used: in the app after an in-app task, by email after an order, in the chat window after a chat. Pew found that people on the phone tend to pick items near the end of a list, while people reading a list pick items near the top. Do not compare phone results with online results without checking.[7]
- Pretest. AAPOR calls pretesting important before fielding any survey. Ask a few people like your customers to answer it and say how they understood each question.[8]
- Response rate. If few people answer, you may miss whole types of customers. AAPOR suggests reminders sent in different ways and at different times, and incentives where they fit.[8]
How many responses you need
Precision depends on the number of responses, not on how many customers you have. A simple random sample of 1,067 responses gives a margin of error of about plus or minus 3 percentage points at 95% confidence.[9] A subgroup has fewer responses and a wider margin. In Pew's example, about 160 responses gave plus or minus 8 points.[9] So if you plan to compare regions or plans, each group needs enough answers on its own. The margin of error covers sampling only, and it says nothing about bias from who chose to answer.[9]
Analysing open-text answers
The open "main reason" answers tell you what to fix, but only once someone reads and counts them. AAPOR notes that open answers must be coded before you can count them.[8] A simple process:
- Read a sample first. Read 100 to 200 answers and write down the themes you see, in the customers' words.
- Write a coding frame. List each theme with a one-line definition and one example answer. Keep themes specific enough to act on: "delivery later than promised" rather than "delivery".
- Code every answer. One answer can carry more than one theme. Mark answers that say nothing, such as "fine" or "n/a", as their own code so they do not inflate a real theme.
- Double-code a sample. Have a second person code the same 100 answers and compare. Where you disagree, tighten the definitions.
- Count by score group. Count each theme separately for detractors, passives and promoters, or for low and high CSAT. A theme that appears mostly among detractors is a cause to fix.
Coding with an LLM
A large language model can apply your coding frame to thousands of answers. Give it the frame, the definitions and a few labelled examples, and ask it to return a theme code for each answer. Research shows it can come close to human coders, but not on every task:
- A 2026 study compared GPT-5.4 with five human coders on 903 open answers from a survey of PhD students. Agreement between the model and the humans was 0.61 for coding and 0.54 for generating themes. The humans agreed with each other at 0.68, and the model agreed with its own repeat runs at 0.76. Agreement varied widely between questions, and the authors limit the result to their own setting.[11]
- A 2025 study of German open answers tested several models. Only a fine-tuned model reached satisfactory accuracy. The others coded some categories much better than others, which changed how often each reason appeared in the results.[12]
So treat model coding as a first pass that a person checks. Code a sample by hand, compare it with the model's codes, and fix the frame or the prompt where they disagree. Check again whenever the frame or the model changes. Remove names, emails and order numbers before you send answers to an outside model. Our guide to AI customer support agents covers how to review AI output against human judgement in service work.
Voice of the customer survey template
This template is a relationship survey of five questions. Copy it and change the bracketed words. For a support survey, replace question 1 with the CSAT or CES question and question 5 with "Did we solve your problem today?"
| # | Question | Answer type | Why it is there |
|---|---|---|---|
| 1 | How likely are you to recommend [company] to a friend or colleague? | 0 to 10 | NPS, comparable over time and with other companies[1] |
| 2 | What is the main reason for your score? | Open text | Tells you what to fix or keep; coded into themes |
| 3 | What do you mainly use [product] for? | Open text, later a list | Lets you split scores by the job the customer hires you for |
| 4 | What is the one thing we could change that would make [product] more useful to you? | Open text | A ranked list of requests once coded |
| 5 | May we contact you to talk about your answers? | Yes / No | Recruits interviewees |
Before you send it:
- Pretest it with a few customers.[8]
- Decide who will read the open answers each week.
- Decide how you will tell customers what changed.
If the effort score is your main measure, see why a CX mindset measures effort, not just satisfaction.
Frequently asked questions
What questions should a voice of the customer survey ask?
Start with one score question that matches your goal: NPS for loyalty, CSAT for a single interaction, or CES for how easy a task was. Follow it with one open question that asks for the main reason behind the score. Add only stage-specific questions whose answers someone will act on.
How many questions should a voice of the customer survey have?
Short surveys work best. Vendor guidance ranges from six or fewer questions to six to ten questions that take about two minutes. A transactional survey after a support contact can be two or three questions.
What is the difference between VoC and NPS?
Voice of the customer covers all the questions you ask customers, in surveys and interviews, to learn what they need and why. NPS is one metric within it, based on a single 0 to 10 question about how likely the customer is to recommend you.
What are the main voice of the customer metrics?
The three common ones are Net Promoter Score (NPS), customer satisfaction (CSAT) and Customer Effort Score (CES). NPS is the percentage of promoters minus the percentage of detractors. CSAT is usually the share of 4 and 5 answers on a 1 to 5 scale, and CES is the average agreement on a 1 to 7 scale that the company made it easy to handle the issue.
What are good voice of the customer interview questions?
Ask about a specific past event instead of general habits, such as "Tell me about the last time you tried to solve this problem." Follow up with neutral probes like "Tell me more about that" and "Why is that important to you?", and avoid leading questions.
Can AI analyse open-ended survey answers?
Yes, with checks. In one 2026 study of 903 answers, GPT-5.4 agreed with five human coders at 0.61, against 0.68 between the humans themselves. Another study found that only a fine-tuned model reached satisfactory accuracy, so have a person code a sample and compare before you trust the results.
Sources
- Bain & Company. "Measuring Your Net Promoter Score". Net Promoter System. Checked on 1 November 2026.
- Qualtrics. "What is CSAT and how do you measure it?". Checked on 1 November 2026.
- Wikipedia. "American Customer Satisfaction Index", citing B. E. Bryant and C. Fornell, American Customer Satisfaction Index Methodology Report (2005). Checked on 1 November 2026.
- Dixon, M., Freeman, K. and Toman, N. (2010). "Stop Trying to Delight Your Customers". Harvard Business Review, July-August 2010. Checked on 1 November 2026.
- CEB. (2013). "Delighting Customers Doesn't Pay, Making Their Lives Easier Does" (press release, 16 October 2013). PR Newswire. Checked on 1 November 2026.
- Medallia. (2021). "The 5 Most Effective Voice of the Customer Survey Questions". Checked on 1 November 2026.
- Pew Research Center. "Writing Survey Questions". Checked on 1 November 2026.
- AAPOR. "Best Practices for Survey Research". American Association for Public Opinion Research. Checked on 1 November 2026.
- Mercer, A. (2016). "5 key things to know about the margin of error in election polls". Pew Research Center. Checked on 1 November 2026.
- Nielsen Norman Group. "User Interviews 101". Checked on 1 November 2026.
- Bergmann, L. et al. (2026). "LLMs for Survey Text Analysis: A Performance Comparison Between Humans and GPT-5 on Inductive Content Analysis". arXiv:2608.22417. Checked on 1 November 2026.
- von der Heyde, L., Haensch, A.-C., Weiß, B. and Daikeler, J. (2025). "AIn't Nothing But a Survey? Using Large Language Models for Coding German Open-Ended Survey Responses on Survey Motivation". arXiv:2506.14634. Checked on 1 November 2026.
- CustomerGauge. (2026). "Voice of Customer Survey Questions: Examples". Checked on 1 November 2026.
- Checkbox. (2025). "57 Voice of the Customer Survey Questions to Get You High-Impact Insights". Checked on 1 November 2026.