Forty survey responses can produce four usable sentences. What decides whether you get them? It is settled long before you write your first question.
Your survey tool has been returning the same six complaints for two years, and your roadmap has moved once. Somewhere between the export and the planning meeting, the sentence a customer actually said got averaged into a five-point scale. Voice of customer research fixes that, and it starts with a decision almost nobody makes on purpose.
The Program Definition Is Already Everywhere
Most pages on this topic start with a definition of the program, and they are the same page in different clothes. A market research agency defines it, then offers to run the study. A customer experience platform defines it again, then lists the channels its software ingests: surveys, support tickets, reviews, call recordings. A training company teaches the program and sells the course.
Each sells one of three things: the study, the software that runs it, or the training, and each stops where the method starts. One page in the field does treat the method as its own subject, and a tool vendor owns that page. The method is not unclaimed. It is never the product.
A voice of the customer program is a system: an owner, an intake route, a review cadence, a loop that tells customers what changed. Research is a method: a sample, a conversation, a coding pass, a readout. You can run the method in ten days with no system, and you cannot run the system without it.
The machine has the same gap. The AI Overview opens with a definition and three goals: reduce churn, guide product innovation, improve copywriting and messaging. The AI Mode panel goes further and lists channels to collect from. Neither says how many conversations you need, who to recruit, what to ask, or what to do with a transcript. Three AI summaries later, you still do not know how to do voice of customer research. The rest of this page is the method.
Sampling: Who You Talk To Decides What You Hear
Why Your Happiest and Your Angriest Customers Are Both Wrong Samples
Start with your happiest customers and you get your own marketing back: they are the people the product worked for, describing it in the terms your sales deck already uses. Start with the angriest and you get a support incident. Both groups were picked on an outcome you already know.
Recruit for variation instead. Build the list on three axes: how the customer arrived (referral, paid, self-serve), how long they have stayed (inside ninety days, past one renewal, past two), and which job they hired you for. Nine slots across those axes teach you more than thirty names drawn from a satisfaction score.
How Many Interviews Before the Answers Repeat
This is the number everyone guesses at, and its published answers measure different things.
Study | What it measured | Interviews needed | What you get |
|---|---|---|---|
Hennink and Kaiser, 2022 | 23 saturation studies, reviewed | 9 to 17, median 12 to 13 | A defensible default for one narrow question and a similar group |
Hennink, Kaiser and Marconi, 2017 | Code saturation | 6 to 9, or 84 to 91 percent of codes | You have heard the topics |
Hennink, Kaiser and Marconi, 2017 | Meaning saturation | 16 to 24 | You understand what the topics mean |
Guest, Bunce and Johnson, 2006 | Appearance of codes | 94 percent of codes in the first 6 | Fast coverage, thin interpretation |
Griffin and Hauser, 1993 | Customer needs surfaced | 20 hour-long interviews, over 90 percent of needs | Where the 20 to 30 advice comes from |
The middle rows change how you plan. Hennink, Kaiser and Marconi put the difference in a sentence worth memorizing: "Code saturation indicates we have heard it all; meaning saturation indicates we understand it all." Hearing it all takes six to nine conversations. Understanding it takes roughly twice that. Most teams stop at the first threshold, produce a list of themes, then wonder why the themes match the ones they had before the study.
Set a stopping rule before you start. Run interviews in pairs, and when two consecutive conversations in the same band produce no new code, that band is done.
You will still meet the advice to run 20 to 30 interviews. Griffin and Hauser is behind it, and it measured conversations needed to surface needs, not how precisely a population can be described. Use twenty when you want width across a scattered customer base. Do not report percentages from it, because a sample that size carries a margin of error wide enough to make the number decorative.
The Interview
Five Questions That Produce Language Instead of Ratings
An interview asks the customer to produce the sentence you could not have written. Book forty minutes with one interviewer and a recording you have permission to keep, then use five questions, in this order.
"Walk me through the last time you went looking for something like this. Where were you, and what had just happened?" Specific moments produce detail; general questions produce opinions about your category.
"What did you try before us, and what made you stop?" This gives you the workaround, which is usually the real competitor. Listen for the verb.
"When you described this problem to a colleague, what did you call it?" The highest-value question in the set, because it hands you the customer's own noun, which is the noun your headlines should use.
"What did you expect this to do in the first week?" A customer who expected a report and got a dashboard is unhappy with a mismatch you can fix in copy.
"If it disappeared tomorrow, what would you do instead?" This measures dependence without asking how much they like you. The answer is a list of substitutes, your competitive landscape in the customer's own words.
Then stop talking. The most useful sentence usually arrives three seconds after the answer seems finished.
What Not to Ask, and the Tells That Give It Away
Leading questions supply the answer inside the question: "How frustrating is the setup process?" You have told the customer how to feel. Ask what happened, then let them describe it.
Double-barreled questions ask two things at once, and the methodology literature calls them the most common design fault, for a plain reason: the answer cannot be attributed to either half. The tell is the word "and" joining two verbs, as in "how do you find the pricing and the onboarding?" Ask twice.
Hypotheticals about the future produce fiction. "Would you pay for this?" measures politeness. "What did you pay for last time?" measures behavior.
The other tells are easy to hear: "would you say that", "don't you think", and any moment where the customer answers a question you did not ask.
Coding the Answers
Turning Transcripts Into a Phrase Bank
Coding sounds like a research specialty. It is bookkeeping one person can do in a week for a dozen interviews.
Transcribe verbatim, or keep the recording and type the quotes while it is fresh. Read the first transcript once without marking anything. On the second pass, highlight spans that answer one of three things: what the customer was doing, what got in the way, and what they said about it. Name each span with the customer's own noun. Group the names into five to eight themes and stop, because more than eight means you are naming situations, not patterns.
Here is the shape of it. The customer says: "I had it in a spreadsheet for about eight months. Every Monday I would go through the list and decide who to chase. The thing that finally made me look for something else was a Monday where I had forty rows and I just closed the tab."
Three codes: the ritual (a Monday review of the list), the workaround (a spreadsheet held together for eight months), and the trigger (a row count that turned into a closed tab). One phrase-bank entry: "I just closed the tab." Context: self-serve customer, eight months in, list-based workflow. Signal: recurring across that band.
That entry is a headline, a churn-risk segment and an onboarding warning in five words.
The Customer's Words Versus the Words You Wish They Had Used
Refusing to improve the quote is what separates a phrase bank from a summary. Write "streamline" where the customer said "stop chasing people" and you have thrown away the only asset the interview produced.
A test you can run in ten minutes: read your phrase bank next to your homepage. If the two sound like the same person wrote them, you have been editing. The phrases that survive usually feel slightly wrong inside a brand document, and those are the keepers. A phrase bank that contains nothing embarrassing is a summary of your own marketing, written by your customers on your behalf.
Acting on It
Where the Phrase Bank Goes: Messaging, Positioning, and the Objection List
A phrase bank has four destinations, each cheap to update.
Headlines and ad copy. Take the customer's noun and verb, put them in the first line, and test that against the version you wrote from your own vocabulary.
The objection list. The answers to "what did you try before us, and what made you stop" are your objection handling, in the order customers raise them. If three interviewees describe the same failed workaround, it belongs where you explain what you do.
Positioning. The jobs your interviewees describe decide which segment you are in, and that decision comes first. If you have not done that work, read segmentation, targeting and positioning next, because a phrase bank tells you what people say while positioning decides what you claim.
Advocacy. The customers who gave you the sharpest language are the ones who will speak on the record, so a customer advocacy program starts with a conversation they already enjoyed.
The One-Page Readout That Survives a Product Review
Research usually dies in the meeting. A long deck invites a debate about method, and that debate is a way to avoid a decision.
Write one page with five blocks. The decision you are asking for, in one sentence. The three sentences that carry it, quoted verbatim with the interview number attached. The sample in two lines: who, how many, which bands. The count, meaning how many interviews produced the sentence. And one line on what would change your mind.
That last block keeps the page honest. Forrester reports that only 12 percent of customer experience professionals rate the maturity of their program as high or very high. The bottleneck is rarely collection. It is the step between a transcript and a decision.
I run an agency, and this is where I have watched the work fail most often: the findings are real, and the readout is a fifty-slide deck that someone summarizes in one sentence. Five blocks on one page get argued with, which is progress. For the wider family of methods, see market research techniques.
When a Platform Is Actually Worth It
A platform earns its cost when the amount of incoming feedback passes what one person can read in the time available. Below that point the tool stores text nobody opens.
Count the unstructured customer text your team receives in a month: support tickets, review sites, call recordings, survey verbatims, sales notes. If that number is in the dozens, a spreadsheet and a monthly coding session will beat anything you buy. If it is in the hundreds or thousands, in quantities no one person can read, then ingestion, deduplication and tagging start paying for themselves. Software solves scale.
Either side of that line, the sampling decision is still yours, and the coding rule still decides whether you end with themes or a phrase bank. If you reach the point of choosing a system, the guide to market research tools is the next read rather than a detour.
Conclusion
Voice of customer research is where the finding gets made, and neither a platform nor an agency will hand you the sentence that reframes your homepage.
You already have the raw material: your last forty survey responses, your support queue, the recording of a call your team had last week. What you do not have yet is a sampling decision and a script. Pick two bands, book four conversations this week, and ask the five questions. Keep the sentence that makes you wince, put it on one page, and take it to the meeting where the roadmap gets decided.
Voice of Customer Research: Frequently Asked Questions
What is voice of customer research?
A method for collecting the language customers use about their own situation, then turning it into decisions about messaging, product and positioning. It differs from a satisfaction survey in what it produces: sentences rather than scores, from a sample small enough to read closely.
How many customers should you interview?
For one narrow question and a similar group, plan on 9 to 17 conversations, with 12 to 13 as a working median, based on the systematic review by Hennink and Kaiser. Expect to hear all the topics after six to nine and to understand them after 16 to 24. Covering several distinct segments means treating each as its own group.
Do you need software to run a VoC program?
No. You need a sample you can defend, five questions, a recording, and an afternoon a week of coding. Software becomes useful when incoming feedback passes the point where one person can read it, and not before.
Start With Five Conversations
You do not need a platform to run this. You need a sample you can defend, five questions, and one afternoon of coding. Allable keeps the transcripts, the phrase bank and the drafts in one project, so the next headline starts from a customer sentence instead of a summary.
