Voice of Customer (VoC)
 - Arfadia.jpg "Voice of Customer (VoC) - Arfadia")
Voice of Customer (VoC) is a structured program for collecting, analyzing, and acting on what customers need, expect, and experience. It draws on interviews, surveys, complaints, reviews, and behavioral data, then turns that evidence into prioritized decisions.
The term comes from product development. In their 1993 paper in Marketing Science, Abbie Griffin and John Hauser described the voice of the customer as a hierarchy of customer needs, each with a priority attached, used inside Quality Function Deployment to connect what customers want with engineering and design decisions. Marketing teams now use the phrase more loosely. That looseness causes real trouble, because a dashboard of scores is not a voice and a pile of comments is not a program.
What Is Voice of Customer (VoC)?
At its core, VoC answers one question: what do customers need, and how much does each need matter to them? Everything else in a program, from survey design to escalation rules, exists to answer that question with less bias and to act on the answer sooner.
Customer Needs, Not Requested Solutions
Griffin and Hauser drew a sharp line here. A customer need, in their definition, is a description in the customer's own words of the benefit a product or service should deliver. It is not a feature request and not a technical specification. A customer who asks for a bigger monitor may really need text that is easy to read, and screen size is only one of several ways to meet that need.
The same paper separates basic needs, which customers simply assume, from articulated needs and exciting needs. Basic needs are the dangerous ones. ISO 10004:2018, the international guideline for monitoring and measuring customer satisfaction, makes a similar point from the quality side: customers do not always state every aspect they expect, and items they take for granted may never be mentioned at all.
A VoC Program Versus a Single Feedback Metric
Scores such as Net Promoter Score (NPS) or satisfaction ratings can feed a VoC program. They are not the program. A score tells you how a group answered one question at one moment, while VoC also covers which needs sit behind the answer, how important each need is, who owns the fix, and whether the fix worked. The same boundary applies to customer experience (CX): CX is the experience being managed, and VoC is the evidence base used to manage it.
| Aspect | Single feedback score (for example NPS or CSAT) | VoC program |
|---|---|---|
| What it captures | How a group answered one question at one moment | Customer needs, how important each one is, and the reasons behind them |
| Main sources | One survey item | Direct, indirect, and behavioral and operational evidence |
| Typical output | A number and its trend | Prioritized needs with owners and decisions |
| Who acts on it | Not defined by the score itself | An owner for each prioritized need |
| How success is judged | Movement in the score | Whether the targeted problem actually changed |
ISO 10004:2018 frames the work as a cycle rather than a survey. Organizations identify customer expectations, gather data, analyze it, communicate the results, and keep monitoring the actions taken. The standard does not present VoC as a named program, but the structure maps closely onto one.
Where VoC Comes From: Direct, Indirect, and Behavioral Sources
ISO 10004:2018 sorts customer satisfaction information into two groups: direct measures, gathered from customers themselves, and indirect indicators, drawn from records that already exist. This entry adds a third working category, behavioral and operational data. That third label is an editorial grouping, not an ISO term.
Direct Sources: Interviews and Surveys
Direct sources ask customers to answer. ISO describes qualitative surveys as a way to uncover which characteristics matter to customers, and quantitative surveys as a way to measure satisfaction, benchmark, and track change over time. Interviews sit on the qualitative side, alongside other qualitative research methods.
Format matters less than many teams assume. In an office equipment study reported by Griffin and Hauser, two one-hour, one-on-one interviews surfaced about 51% of the identified customer needs, roughly the same as a single two-hour focus group at 50%. Surveys carry their own risks. The American Association for Public Opinion Research (AAPOR) advises researchers to ask whether a survey alone can answer the question, or whether other data, such as administrative records, will also be needed.
Indirect Sources: Complaints, Reviews, and Support Contacts
Indirect sources are records customers create for other reasons. ISO 10004:2018 lists complaint and dispute trends, calls for assistance, compliments, product returns and repairs, service failures, consumer group and media reports, regulator publications, and comments on social media as examples. ISO 10002:2018 adds the complaints-handling view, covering how an organization stays open to feedback, resolves complaints, and analyzes them to improve products and services. Public conversation tracked through social listening belongs in this group too.
These sources are cheap and abundant. They are also highly selective, though. Each one records only the people who chose to write, call, return, or post.
Behavioral and Operational Data
Purchase records, cancellations, repeat contacts, delivery delays, and on-site paths show what customers did rather than what they said. Laid against the stages of a customer journey map, this data shows where friction concentrates. It cannot explain motive on its own. An abandoned checkout is a behavior that needs explaining, and an interview or a short survey usually supplies the explanation.
| Source type | Examples | What it shows well | Main limitation |
|---|---|---|---|
| Direct | Interviews, surveys, usability sessions | Needs, priorities, and reasons in the customer's words | Covers only people who were invited and agreed to take part |
| Indirect | Complaints, reviews, support contacts, returns, social posts | Specific failures, customer language, emerging issues | Self-selected, so volume does not equal prevalence |
| Behavioral and operational | Transactions, cancellations, delivery and service records | What customers actually did, at scale | Shows behavior without the reason behind it |
The Customers Your VoC Program Does Not Hear
Most VoC programs hear the customers who chose to speak. That sounds obvious at first glance. Its consequences are not, because every source described above filters out some customers before a single comment reaches an analyst, and the filters tend to remove the same kinds of people again and again.
Silent Dissatisfaction
The UK's Consumer Detriment Survey 2024, carried out by the National Centre for Social Research for the Competition and Markets Authority, offers a rare view of what happens after something goes wrong. Drawing on 6,371 UK adults and detailed data on 9,957 problem incidents in the 12 months to April or May 2024, it found that consumers took no action at all in 22% of incidents. The most common reasons were that the problem did not seem serious enough (27% of unactioned incidents) or that a good outcome seemed unlikely (21%).
When consumers did act, 75% of those incidents involved contacting the seller or provider directly. That is the channel a VoC program usually sees. The survey also found that consumers aged 18 to 39 and those who described their finances as difficult were consistently more likely to experience problems and less likely to act on them, and that problems with products bought through social media platforms were less likely to be actioned. A complaint log, in other words, under-represents some of the customers most likely to be harmed.
Regulators apply the same caution to their own data. The US Consumer Financial Protection Bureau states that its Consumer Complaint Database is not a statistical sample of consumers' experiences, and that a low number of complaints does not necessarily mean little or no consumer harm.
Who Speaks First, and Loudest
Self-selection shapes what gets said, not just who says it. Xinxin Li and Lorin Hitt, writing in Information Systems Research in 2008, modeled how early buyers whose preferences differ from later buyers can skew online product ratings, and tested the model on book reviews posted on Amazon. Griffin and Hauser saw a related effect in satisfaction scores. When customers rated only the brand they had chosen, a small niche brand outscored the long-time category leader, because each brand was judged by people already inclined to like it.
Why Synthetic Respondents Cannot Fill the Gap
Generating AI respondents to stand in for missing customers is a tempting shortcut. The available direct evidence is poor. In a September 2026 study, Pew Research Center asked an AI model to answer three US survey waves as digital twins of real panel members. The synthetic results differed from the human results by about 12 percentage points on average, and the errors were especially large for groups including Republicans, Black adults, people who had not attended college, and people who rarely or never use the internet.
One example makes the pattern concrete. In a March 2026 survey, 42% of real Hispanic adults in the US said they were at least somewhat likely to follow the World Cup, while 97% of the simulated Hispanic respondents said so. Pew measured public opinion, not customer feedback, so applying the finding to VoC is an inference. The UK complaint-data findings and the Pew findings cover different populations and different groups, so they stand as separate cautions rather than one combined claim.
How to Analyze VoC Without Losing the Customer's Words
Separate Observations From Findings
The UK Government Digital Service manual on how to analyse a research session sets out a simple discipline. Write down exactly what people said or did, sort those observations into themes, write findings as short statements, and then decide what to change or research next. Keeping observation apart from interpretation protects the customer's actual words. It also makes disagreement visible early.
Structure and Prioritize Needs
Griffin and Hauser recommend arranging needs into a hierarchy of primary, secondary, and tertiary needs and attaching an importance weight to each. Two of their findings are worth keeping close. Frequency of mention did not work as a stand-in for importance, since the needs customers rated as important were no more likely to be mentioned than needs in general. And a hierarchy built from customer card sorts tended to reflect how customers use the product, while one built by the internal team tended to reflect how the company builds it.
Segments need their own priorities. A need that matters greatly to one buyer persona may barely register for another, and averaging the two can hide both.
Using AI for Coding and Summaries
Language models can speed up coding, clustering, and summarizing open text, including first-pass sentiment analysis. A Journal of Marketing study by Neeraj Arora, Ishita Chakraborty, and Yohei Nishimura found that expert judges rated human and LLM hybrids above either humans or LLMs working alone, and the authors stress that a human supervisor remains necessary because models can be wrong, biased, or invent content. The NIST Generative AI Profile (NIST AI 600-1, July 2024) calls that last risk confabulation and suggests reviewing and verifying the sources and citations in model outputs. For VoC, every theme and every quoted customer line should trace back to a real record.
Acting on VoC: Closing the Loop
Resolving Cases Versus Fixing Causes
Closing the loop happens at two levels. One is answering the individual customer. The other is changing whatever caused the problem, and ISO 10002:2018 covers both, from resolving complaints to analyzing them so that products and services improve.
The individual level matters more than it may seem. In the UK survey, 52% of problem incidents ended with a positive resolution. Where consumers asked for something and received nothing, 48% of those incidents had a negative effect on the person's mental health, compared with 15% where they received what they asked for. These are self-reported, incident-level figures from the UK, so they show an association rather than proof that a reply alone repairs harm.
Keeping Review Feedback Honest
Public reviews are a VoC source, so the way they are collected shapes data quality. Two sets of rules are relevant, one platform-wide and one specific to the United States. Google's Maps content policy does not allow merchants to discourage negative reviews, selectively ask only satisfied customers, or tell staff to collect a set number of reviews. In the US, the Federal Trade Commission's final rule on consumer reviews and testimonials, announced in August 2024, bans incentives conditioned on a positive or negative sentiment and prohibits certain forms of review suppression. A filtered review stream is more than a compliance risk. It is a distorted voice.
Measuring a VoC Program Honestly
Program Health Versus Outcome Claims
ISO 10004:2018 asks organizations to monitor the actions taken in response to customer satisfaction information and to assess whether those actions worked. That points to tracking program health separately from business outcomes:
- Coverage: which segments, channels, and journey stages produced evidence this period, and which did not.
- Follow-through: how many prioritized needs have an owner, a decision, and a completed action.
- Effect: whether the specific problem an action targeted actually changed.
| Measure | Question it answers | What it does not show on its own |
|---|---|---|
| Coverage | Which segments, channels, and journey stages produced evidence this period | How many customers inside a covered segment stayed silent |
| Follow-through | How many prioritized needs have an owner, a decision, and a completed action | Whether the action fixed the problem |
| Effect | Whether the specific problem an action targeted changed | That the change caused gains in revenue or retention |
| Complaint volume | How many complaints arrived, and in which areas | Whether harm fell, since an absence of complaints does not mean high satisfaction |
Claims about revenue or customer retention need their own evidence. A rise in satisfaction after a change does not show that the change caused it, and AAPOR warns that altering question wording, context, or survey mode can create apparent shifts that are not real. No widely accepted benchmark for VoC return on investment turned up in the official, academic, and government sources reviewed for this entry.
Why Falling Complaint Volume Is Ambiguous
Fewer complaints can mean fewer problems. They can also mean a harder complaint process, fewer customers, or customers who gave up. The definition of customer satisfaction used in ISO 10004:2018 notes that complaints are a common indicator of low satisfaction, but that their absence does not necessarily mean high satisfaction. Read a drop in complaints next to direct measures and behavioral data before calling it a win.
Frequently Asked Questions
What is Voice of Customer (VoC) in marketing?
VoC is a structured program for collecting, analyzing, and acting on customer needs, expectations, and experiences. Marketing teams use it to shape messaging, offers, and research priorities, and the same evidence also informs product and service changes.
Is NPS the same as Voice of Customer?
No. NPS is one survey score that can feed a VoC program. VoC also covers why customers answered as they did, which needs matter most, and what the organization changed in response.
What are examples of VoC data?
Common examples include interviews, surveys, complaints, support conversations, product reviews, returns, and social media comments, plus behavioral and operational records such as cancellations and service failures.
How many customer interviews does VoC research need?
There is no universal number. In a 1993 product study, Griffin and Hauser found that 20 interviews identified over 90% of the needs found across 30 interviews, and they suggested 20 to 30 interviews for a fairly homogeneous segment as a working range. Different products, segments, and research goals can change that number.
Can AI replace customer interviews?
In a US opinion-survey study, they were not a reliable substitute. Pew Research Center found that AI-generated survey answers missed real US public opinion by about 12 percentage points on average. That is one US opinion-survey study, so it does not settle every interview workflow, and research on human and AI hybrids still treats human review as necessary.
