Last updated October 2026
Quick answer
To run a voice of customer (VoC) survey, decide whether you need to discover needs or measure experience, then pick the matching instrument: interviews for needs, a short frozen rating survey for experience. Field it to a population you can name and count. Then check who did not answer before reporting what respondents said. Most programs skip that last step.
Key takeaways
- "Voice of customer" means two different things: a 1993 interview method for finding needs, and a software category for measuring and acting on experience.
- Size interviews by saturation (about 10 to 30) and surveys by precision (often several hundred to several thousand responses).
- Report the response rate next to every score. A big sample with a low response rate is still a biased sample.
- Compare who answered with who you invited before you report any satisfaction number.
- Plan the follow-up workflow (owners, deadlines, resolution) in the system that already tracks customer issues.
What is a voice of customer survey?
A voice of customer survey is a repeatable questionnaire that measures how a defined group of customers experiences something you already offer. The term also has an older meaning, and mixing up the two is the most common source of bad VoC programs.
The 1993 research method. Griffin and Hauser's "The Voice of the Customer" (Marketing Science, 1993) introduced VoC as a way to uncover needs for product development. It relies on one-on-one interviews of about an hour each, usually with 10 to 30 customers. The output is a ranked hierarchy of needs, not a score.
The 2026 software category. Industry analysts now use VoC to describe platforms that collect customer input, analyze it, and act on it. Gartner's market page for Voice of the Customer Platforms lists three mandatory capability areas: data collection (direct and indirect), analysis and insight, and action.
How to choose. Pick the definition by the decision you face, not by preference:
- Use the 1993 method to learn what customers need from something you have not built yet. A roadmap question is usually a needs question.
- Use the 2026 approach to measure how existing customers experience what you ship now. A retention or service-quality question is usually a measurement question.
The two have opposite sample logic. The interview method needs a few long conversations. The measurement approach needs a defined population, a stable questionnaire, and a response rate you can state.
When should you run a VoC survey, and when should you not?
Run a VoC survey when you need a repeatable, population-level read on a defined group of customers and you plan to ask the same questions again in about three months. It is a measurement tool, not a discovery tool.
You are ready to field when you can:
- Name the population you want to describe.
- Count it.
- State how respondents will be selected from it.
- Route a flagged response to a named owner who can resolve it.
If you cannot do all four, you are collecting comments, not measuring.
Use a different method when your question is one of these:
The evidence is well documented. Kohavi et al. (2013) found only one third of ideas tested at Microsoft improved their target metric. Murphy et al. (2005) found stated willingness to pay runs about 1.35 times actual. Von Hippel (1986) showed users of very novel products lack the experience to answer accurately. Chandon, Morwitz and Reinartz (2005) found surveying people changes how their intentions relate to behavior.
The non-response row is the defining limitation of the method. In a study of 173,886 live chat sessions, Park, Cha and Rhim (2018) found that only 16.2% of customers answered the satisfaction survey. Raters were 79.7% positive or neutral, while the inferred figure for non-raters was 45.5%.
Which design decisions shape a VoC survey?
Settle three design decisions in writing before you write a question. Together they decide whether your numbers can be compared across periods.
1. Rating questions or open text. Rating questions carry the measurement; open text explains it. In a job-satisfaction study by Wijngaards, Burger and van Exel (PLOS ONE, 2019), scores built from semi-open questions agreed with human coding far better than scores from fully open questions (average r = .774 vs. .508). The open-text scores were also unstable over time, with test-retest reliability averaging .255. Do not make a text-derived score your tracked metric.
2. Surveys you send or comments customers volunteer. Both are biased, in opposite ways. Surveys skew positive because unhappy customers answer less often. Volunteered reviews skew because of who chooses to post. Hu, Pavlou and Zhang (MIS Quarterly, 2017) show that these self-selection biases make the average online rating a biased estimate of quality. Report each source separately and say which population it covers.
3. Fixed schedule or event trigger. A relational survey runs on a fixed calendar to a defined customer base. A transactional survey is triggered by a completed interaction, such as a support case. Run both if you need both, but never merge them into one trendline. A transactional score is conditioned on the interaction having happened, so it is not comparable to a relational baseline.
The combination that holds up: a relational schedule, a frozen set of rating questions, and a few semi-open follow-up questions. The worst is a transactional trigger with a text-derived score, where both the denominator and the metric change every period.
Which VoC approach fits your question?
Match your question to an instrument. Stop at the first row that fits.
| Question you are trying to answer | Approach | Instrument | How to size it |
|---|
| What do customers need from a product that does not exist yet? | 1993 method | One-on-one interviews coded into a needs hierarchy | Saturation: about 9 to 30 interviews per similar segment |
| How satisfied is a defined population right now? | 2026 approach | Frozen rating battery, relational schedule | Precision (see the margin of error table) |
| Did satisfaction change between two periods? | 2026 approach | The same frozen battery, same population | Power (see the detectable change table) |
| Why did a specific interaction go badly? | 2026 approach | Rating battery plus semi-open questions, triggered by the interaction | Precision on the triggered group only |
| Which fix should we build first? | Neither alone | Conjoint or discrete choice | Depends on the design |
| Did our change cause the improvement? | Neither | Online controlled experiment | Power on the primary metric |
How do you write a VoC questionnaire?
A VoC questionnaire has two parts. A short set of rating questions produces the number you report and track. Two semi-open questions produce the reasons behind it and are never scored. Once the first period is fielded, freeze the wording, scale, labels and order; any change starts a new baseline.
Keep the rating set to four to six questions (a judgment call; each extra question costs completions). Change only the bracketed text below.
| Item | Type | Wording | Scale |
|---|
| S1 | Rating, primary | Overall, how satisfied are you with [PRODUCT OR SERVICE] over the last 30 days? | 5 points, Very dissatisfied to Very satisfied, every point labeled |
| D1 | Rating, driver | How satisfied are you with how easy [PRODUCT] is to use? | Same scale and labels |
| D2 | Rating, driver | How satisfied are you with the speed of [PRODUCT]? | Same scale and labels |
| D3 | Rating, driver | How satisfied are you with the support you receive for [PRODUCT]? | Same scale and labels |
| P1 | Semi-open | What made you choose that rating? | Open text, shown to everyone |
| P2 | Semi-open | What is the one thing we could change that would matter most to you? | Open text, shown to everyone |
Rules that protect your data:
- Show P1 to everyone, not just unhappy respondents. A diagnosis drawn only from detractors only describes detractors.
- Leave out a "Prefer not to answer" option on rating questions, so a blank clearly means the person skipped it.
- Label every scale point. An unlabeled "3" can mean neutral, adequate, or "I didn't want to think about it."
- Put the time frame in the question ("over the last 30 days") and keep it fixed.
- Ask about a named product or interaction, not "the company," and never ask about two things in one question.
What to report. Lead with the average (mean) satisfaction score and show the top-two-box percentage (the share choosing the top two points) beside it. Morgan and Rego (Marketing Science, 2006) found that "average satisfaction scores have the greatest value in predicting future business performance" and that "Top 2 Box satisfaction scores also have good predictive value."
How Sprig helps: AI follow-ups. The Field Agent can ask a conversational follow-up question about an open-text answer while the respondent is still in the study. It cannot reach people who never answered, so it deepens answers without fixing bias.
How many responses does a VoC survey need?
A VoC survey needs enough responses for the precision your decision requires. For a single snapshot, that is often 300 to 600. To detect a small change between periods, it can take several thousand per period. The famous "20 to 30 customers" figure applies only to interviews, not surveys.
Where 20 to 30 comes from. In Griffin and Hauser's study, 20 interviews surfaced over 90% of the needs found across 30 customers. Hennink and Kaiser's 2022 review confirmed that interview studies with similar participants typically reach saturation within 9 to 17 interviews. Saturation measures coverage of ideas, not the precision of a percentage.
Margin of error for one survey. These figures assume 95% confidence, a simple random sample, a top-two-box share near 70%, and a standard deviation of 1.0 on a 5-point scale. The "critical difference" is how far two equal-size periods must differ before the change is real.
| Responses | Margin of error (top-two-box %) | Margin of error (mean, 5-point scale) | Critical difference between two periods |
|---|
| 100 | ±9.0 points | ±0.20 | 12.7 points |
| 200 | ±6.4 points | ±0.14 | 9.0 points |
| 400 | ±4.5 points | ±0.10 | 6.4 points |
| 1,000 | ±2.8 points | ±0.06 | 4.0 points |
| 2,000 | ±2.0 points | ±0.04 | 2.8 points |
Example: with 200 responses per quarter, a top-two-box score moving from 68% to 74% has not changed in any statistical sense, because the critical difference is about 9 points.
Responses needed to detect a real change (80% power, 95% confidence, 70% baseline):
| Change to detect | Responses per period |
|---|
| 3 points | about 3,550 |
| 5 points | about 1,250 |
| 10 points | about 290 |
| 15 points | about 120 |
Response rate matters more than response count. Adding responses narrows the margin of error around a biased estimate; it does not remove the bias. The U.S. Office of Management and Budget's Standards and Guidelines for Statistical Surveys (2006) says that when a survey's unit response rate is below 80 percent, you should conduct an analysis of nonresponse bias, and below 70 percent on any single question, an item nonresponse analysis. Nearly every VoC survey falls below 80%.
Who should you survey, and how should you field it?
Survey the population your conclusion will be about, on every channel that reaches it, and record exactly how many people you invited. Define the population (the sampling frame) before you set any audience filter.
Write the frame as a sentence with a count. "All accounts with at least one active seat in the last 30 days: 41,200 accounts as of September 1" is a frame. "Engaged users" is not, because nobody can reproduce it next quarter. A sample that only reaches customers who are currently active will miss the customers most likely to leave.
Know the difference between screening and quotas. Screening removes people outside the frame. Quotas hold the mix of people inside it. On most survey platforms, quotas count people who have already responded, so they cannot shape who gets invited.
Record segment details on every response. Attach each segment you plan to analyze (plan tier, region, tenure) to the response at collection. Only use segments you also hold for the whole frame, because the non-response analysis compares respondents against everyone.
Fielding checklist:
- Use at least two channels if your frame includes inactive customers. Email reaches the whole contact list; panels reach people outside your customer base.
- Send one invitation and one reminder. Compare scores from first-wave and reminder respondents before pooling them; a gap tells you which way your non-response bias runs.
- Record five counts at fielding time: invitations sent, delivered, shown, started, and completed. Without them you cannot compute a response rate.
- Before launch, confirm the questionnaire is identical to last period, the period label is attached to every response, and bot filtering is on for link and panel distribution.
Choose a cadence you can sustain. Quarterly is common, but pick a schedule you can keep with a frozen questionnaire for at least four periods. Over-surveying also suppresses response: Porter, Whitcomb and Weitzer (2004) saw student response rates fall from 68% to 58% to 46% as prior surveys rose from zero to two. Those are student figures, so treat them as proof that fatigue is real, not as customer benchmarks.
How Sprig helps: reaching the frame. Sprig fields one questionnaire across email with a custom sending domain, shareable links, QR codes and research panels, and supports 300+ targeting attributes that can be written onto each response. Two limits apply: attributes filter who is eligible but do not hold the mix of responses constant, and there is no native SMS, phone, contact-center or offline capture.
How do you analyze VoC survey results?
Compute the response rate and compare respondents with the full population before you compute any satisfaction score. That order is the whole method. Reversing it is how teams end up defending a number they cannot support.
Step 1: response rates. The unit response rate is completed responses divided by eligible invitations. The item response rate is answers to a question divided by the people who reached it. Check both against the 80% and 70% thresholds above.
Step 2: the non-response worksheet. This five-step check turns a response rate into an estimate of bias.
StepWhat you computeWhat it tells you1Unit response rateWhether you clear 80% (you usually will not)2Respondent mix on each attribute you hold for everyoneWho answered3Population mix on the same attributesWho you invited4Respondents minus population, per attributeWhich groups are over- or under-represented, and by how much5Mean satisfaction within each over- and under-represented groupWhether the mix gap pushes the headline score up or down
If an over-represented group scores above average, your headline is biased upward. Late (reminder) respondents tend to resemble non-respondents, so a drop between waves suggests the people you never reached score lower still.
Step 3: the satisfaction read. Report the mean and top-two-box together, each with its margin of error and the number of completed responses. Exclude blank answers from each question's base. Never recode a blank as a neutral midpoint; that hides an item non-response problem inside the score.
Step 4: themes from open text. Use themes to explain the score, not to replace it. A theme count is only trackable if the coding list is frozen like the questionnaire. Most automated theming rebuilds the theme list as new responses arrive, so this period's counts are not comparable to last period's. For a full method, see coding open-ended responses into themes and comparing results across customer segments.
Cleaning rules. Set removal rules in advance (for example, straightlined responses with empty open text) and report the removed count next to the analyzed count.
How do you read and act on a VoC result?
Read every VoC result as three numbers: the score, its margin of error, and the response rate behind it. A score reported without the other two cannot be interpreted.
Check any period-over-period change against the critical difference before calling it a change; at typical sample sizes, most quarterly movement is noise. Be wary of a rising score with a falling response rate, which usually means respondents are more self-selected, not happier.
The closing-the-loop operating spec. Build this in the ticketing or customer relationship management (CRM) system that already tracks ownership and resolution, not in a spreadsheet.
| Element | What you define | A sensible default |
|---|
| Trigger | Which responses enter the loop | Bottom-two-box on S1, or open text matching a named rule |
| Owner | Who receives the case, by role | The account or service owner already responsible for that customer |
| Response time | How quickly first contact happens | One business day for bottom-box; five for everything else |
| Resolution states | The fixed list a case can be in | Open, contacted, resolved, resolved with product change, closed unresolved |
| Escalation | What goes to a wider team | Cases past the target, and themes recurring across accounts |
| Customer update | What the customer hears, and when | An acknowledgment at first contact and the outcome at resolution |
Publish the count of cases resolved alongside the satisfaction score. A score alone reads as a vanity metric; a resolution count reads as work delivered.
Can you analyze VoC data with Claude or ChatGPT?
Yes. A language model connected to your response export can compute the response rate, the respondent-versus-population comparison, and the satisfaction read in one pass. It cannot decide the order, so your prompt must. For a fuller treatment, see how to analyze survey data with AI without sacrificing rigor.
Data: [RESPONSE EXPORT CSV] and [POPULATION FRAME CSV], joined on [userId].
Use code to calculate everything. Write and run the code, and show it.
1. Compute the unit response rate and each question's item response rate.
2. For every attribute in both files, compare respondent share with
population share. Name over- and under-represented segments.
3. Only then compute the mean and top-two-box for [S1], overall and by
segment, with 95% margins of error and n. Exclude blanks; never recode
them to a midpoint. Report segments under [30] as "n = X".
4. Recompute the mean a second way. If the two disagree, stop and report both.
5. Flag only segment differences larger than the critical difference.
Four risks to design around:
- Agreement can look better than it is. Hill et al. (PLOS Digital Health, 2026) found 93.5% agreement between a model and human coders, but a kappa of only 0.34, because most codes were rare. Build the theme list once yourself, then have the model apply that frozen list.
- Coding errors are not random. Ashwin, Chhabra and Rao (2025) found that model coding errors vary with who the respondent is, so automated theming can under-report some segments' concerns.
- Outputs vary between runs. Run any number you plan to report twice and compare.
- Models tend to agree with you. Never ask a model to confirm a theme you already suspect.
How Sprig helps: MCP. Sprig's Model Context Protocol connection gives Claude, ChatGPT, Gemini and Copilot access to study data, scoped to the user's role, with response data not used for training. On Enterprise plans (rolling out as of October 2026), agents can also create, edit, translate, target, and launch studies, MCP-created studies carry an MCP label, and admins can disable MCP at any time. Each call returns up to 1,000 responses, so a full program pull takes several calls.
What are the limits of a VoC program?
The evidence for customer satisfaction as a measure is strong. The evidence for VoC programs as most teams run them is weak.
- The flaw is old. Griffin and Hauser's 1993 paper already documented "a self-selection bias in satisfaction measures."
- Measuring changes customers. Dholakia and Morwitz (2002) found that measuring satisfaction changed purchase and defection behavior, with effects that persisted a year later.
- Program impact is unproven. No peer-reviewed study was found showing that closed-loop VoC programs improve retention or revenue.
- The underlying measure works. A meta-analysis of 535 correlations by Mittal et al. (2023) links customer satisfaction to retention (r = .60) and to firm financial outcomes.
The fix is in how the program is run: write down the frame, analyze non-response, and build the follow-up workflow before buying a dashboard.
What are the most common VoC mistakes?
Most failed VoC programs come from eight avoidable mistakes, and most are made before a single response arrives.
- Sizing a survey at 20 to 30 responses based on an interview saturation finding.
- Changing the questionnaire between periods while reporting one trendline.
- Blending relational and transactional results into one series.
- Treating a sample of active users as a sample of all customers.
- Reporting a period-over-period change without checking the critical difference.
- Recoding skipped questions to a scale midpoint instead of excluding them.
- Reporting a theme count as a metric when the theme list regenerates.
- Reporting a satisfaction score without the response rate that produced it.
After the survey, the common failure is routing either everything or nothing. A named trigger with a response-time target prevents both.
How does Sprig support a VoC program?
Sprig covers the collection and analysis parts of a VoC program. It does not cover the action layer, so plan case management in the system you already use.
What Sprig covers: distribution across email, links, QR codes and research panels; 300+ targeting attributes written onto each response; AI follow-up questions; open-text theming traceable to individual responses; and an MCP connection with published governance.
What Sprig does not cover:
- Case management, owner assignment or resolution tracking for responses
- A published weighting or statistical-power method
- Quotas on attributes you already hold (response quotas are capped at three screener questions)
- Native SMS, phone, contact-center or offline capture
- A theming accuracy figure or minimum response count (themes also regenerate, so counts are not a tracked metric)
Sprig is not among the twelve vendors evaluated in the 2026 Gartner® Magic Quadrant™ for Voice of the Customer Platforms. Teams evaluating the platform category rather than the method can start with how Sprig compares to Medallia and Sprig's experience measurement approach.
Gartner® and Magic Quadrant™ are trademarks of Gartner, Inc. and its affiliates. Gartner does not endorse any vendor, product or service depicted in its research publications and does not advise technology users to select only those vendors with the highest ratings or other designation. Gartner research publications consist of the opinions of Gartner's research organization and should not be construed as statements of fact. Gartner disclaims all warranties with respect to this research.
Which methods work alongside a VoC survey?
Several methods answer questions a VoC survey cannot. Choose by what the decision requires.
- Customer satisfaction (CSAT): the rating core of this method. See how to run a CSAT survey.
- Customer effort score (CES): a narrower read on a single interaction. See how to run a CES survey.
- Customer needs assessment: the modern form of the 1993 branch. See how to run a customer needs assessment.
- Experiments, conjoint and behavioral analytics: for cause, trade-offs, and actual behavior.
- Brand tracking: measures perception in a market that mostly does not use your product. Never put it on the same chart as a VoC score.
Frequently asked questions
Short answers to the questions teams ask most often about VoC surveys.
What is a voice of customer survey?
A voice of customer survey is a repeatable questionnaire that measures how a defined group of customers experiences a product or service. The term has two meanings: a 1993 research method built on 10 to 30 one-on-one interviews, and a modern software category covering data collection, analysis and action. Most teams mean the measurement version.
How many responses does a voice of customer survey need?
For a single snapshot, 300 to 600 responses is a practical range. At 400 responses, a top-two-box score near 70% has a margin of error of about 4.5 points. Detecting a 3-point change between periods takes about 3,550 responses per period.
Is 20 to 30 customers enough for a VoC survey?
No. The 20 to 30 figure comes from Griffin and Hauser's 1993 interview study, where 20 hour-long interviews found over 90% of customer needs. It measures interview saturation, not survey precision. A survey of 30 people has a margin of error of roughly 16 to 18 points.
What is a good response rate for a VoC survey?
No published cross-industry benchmark discloses its sample and method, so this guide sets no floor. The best-documented peer-reviewed figure is 16.2%, from 173,886 chat sessions. The more useful standard is federal: below an 80% response rate, OMB guidelines call for a non-response bias analysis.
How is VoC different from customer satisfaction measurement?
Customer satisfaction measurement is one instrument inside a VoC program. A satisfaction survey produces a score for a defined population. A VoC program adds other input sources, analysis of open text, and an action layer that assigns responses to owners and tracks resolution.
Should a VoC survey be relational or transactional?
Run both if you need both, and report them separately. A relational survey measures a defined customer base on a fixed schedule. A transactional survey measures people who just completed an interaction. Because transactional results depend on the interaction, never merge them into a relational trendline.
Can AI analyze VoC open-text responses?
Yes, with limits. Models apply an existing coding framework about as well as human coders, but high agreement can hide low reliability: one 2026 study found 93.5% agreement with a kappa of only 0.34. Build the theme list yourself, freeze it, and have the model apply it each period.
What is the smallest VoC program worth running?
One relational survey per quarter: four rating questions and two semi-open questions, sent to a written-down population, with the five denominator counts recorded and one named owner for flagged responses. Anything smaller is collecting comments, not measuring.
The bottom line
If you need to know what customers want from something you have not built, run 10 to 30 one-on-one interviews and code them into a needs hierarchy. If you need to know how customers experience what you already ship, run a frozen rating survey with semi-open follow-ups, to a written-down population, with the response rate recorded.
Either way, analyze who did not answer before you report what respondents said. Then build the follow-up workflow in the system that already tracks ownership and resolution. The best next step is a properly sampled relational survey with the non-response worksheet filled in. Start from the customer satisfaction survey template.
