Voice of customer analysis without surveys
Voice of customer analysis turns what buyers actually said into decisions somebody can act on. Surveys are one input and the weakest one, because the sample is assembled from people willing to answer a survey. The conversations your team already records are larger and mostly unused. Spellit ClientTruth reads those recordings and counts every share against the group the question is about, so a figure about objections never gets divided by every call.
Almost every guide to voice of customer analysis opens with a survey program: pick a metric, pick a cadence, send the form, tag the answers, close the loop. The steps are sound and the order is backwards. The form is the last place a buyer tells you anything, and it is the only place where somebody had to volunteer before you heard them.
The research on who volunteers is not encouraging. In 2017 Pew Research Center matched 40,182 telephone numbers from one of its own surveys against a national voter file and compared the people who answered with the people who never did. On civic measures the gap was large enough to change the story: 8 percent of people reported working with neighbors in high response government surveys, against 46 percent in the telephone poll, an overstatement of 38 points. The people who answer surveys are joiners, and Pew spells out what that does to an estimate. The same report puts the response rate for a typical telephone poll at around 9 percent.
Business surveys are not political polls, and the correction weights that rescue a national poll are not available to a company sending a form to its own customer list. What survives the comparison is the mechanism: you hear from the people who chose to be heard, and nothing inside the result tells you who stayed quiet.
Your survey sample is assembled from the people who answer surveys
A survey measures the opinions of respondents. It becomes a measurement of your customers only if respondents resemble customers, and nothing inside the survey can establish that they do. The two ends of the feeling scale reply first: somebody who just had a problem fixed, and somebody who just had one created.
That is not a reason to stop running surveys. It is a reason to stop treating a survey as the spine of a voice of customer program, because the spine has to be something you can count honestly, and a self selected sample cannot be counted honestly no matter how clean the tagging is afterward.
Each input carries a different blind spot, including the one this article argues for.
| Input | What it reads well | What it cannot tell you |
|---|---|---|
| Survey with a score | one number tracked over time | who stayed silent and what they would have said |
| Public reviews | the language of the two extremes | anything about the middle of your base |
| Support tickets | problems in the product as it shipped | what the buyer wanted before any money moved |
| Chat logs | small frictions nobody files a ticket about | intent, since most of it arrives after the purchase |
| Recorded sales calls | objections, questions and requests at the moment of deciding | what happens once the customer is live |
The row that matters for a program is the last one, and it is the row that every vendor guide lists and nobody builds on.
The conversations already happened and nobody agreed to be studied
Nothing has to be fielded. The calls were recorded for other reasons, the buyer talked because talking was the point of the meeting, and no incentive, reminder or form stood between what they thought and what you have on file. There is no response rate because there was no invitation.
A recorded call is not a neutral sample of the market either. It is a sample of people who agreed to take a meeting with your sales team, which is its own filter and a strong one. What it is free of is the second filter, the one surveys add on top: among the people you already reached, you hear from everybody rather than from the fraction who felt strongly enough to reply.
For most companies the volume difference is not close either. A quarterly survey with a few hundred responses sits beside several thousand recorded conversations from the same period that nobody has read. Teams that already run a structured call review process are usually reading those recordings for coaching, which is a different question asked of the same material. The rep side of a call and the buyer side of a call are two readings of one recording, and most organizations only ever do the first.
The second reading is the one that has never been run over your own archive. The buyer side of a month you have already read for coaching.
A share is honest only when its base is the group the question is about
Divide by the group the question concerns. The share of calls where an objection was handled well belongs over calls that contained an objection, never over every call in the period. Put the wrong base underneath and the number moves in a flattering direction, which is exactly why nobody catches it.
This is not a rounding problem. A base that includes conversations where the topic never came up will drag any share toward whatever the quiet majority looks like, and the resulting figure looks stable and is quietly wrong. The same discipline is what makes a metric like an agreed next step worth reporting at all.
The time window is a base too, and that is where we got it wrong. Unanswered requests, the moments where a buyer asks for something and nothing comes back, were generated week by week like everything else on the screen. Week by week the count was almost always zero. It read as good news for months. It was a sample size artifact: a single week rarely holds enough conversations for a rare event to appear even once, so a real and persistent gap rendered as an empty block.
The repair was to stop slicing that one signal by week. Unanswered requests are now generated over the whole history as standing items, because a product gap does not reset on Monday, and a week is simply too short a window for something that turns up this rarely. Everything else on the screen is still weekly. One signal needed a different clock, and finding that out took longer than it should have because the weekly number was never wrong, only empty.
A conclusion without the buyer's own words is an opinion wearing a number
A finding is finished when somebody can read the line the customer said, the turns on either side of it, and open the recording at that point. Until then it is a label, which is somebody's reading of what was said rather than the thing that was said.
The first version of the client card in ClientTruth did not do this. It named the main pain, named the motivator, and put nothing underneath either of them. Everything was defensible and nothing was checkable. A separate step was built afterward whose only job is to attach the exact client line plus two messages on each side and the call it came from, and the card was not trustworthy until that step existed.
The selection rule inside that step is worth copying even if you build none of the rest. The representative conversation for a topic is chosen as the one where the client said the most about it, measured by the length of the client's own turn rather than the rep's. A long rep explanation about cost is not evidence that cost is a problem; it is evidence that the rep talks about cost. Sorting by total call length or by how much was said in the room gets you the second thing while looking like the first.
Free tagging has a matching trap. When the label vocabulary is open, a long tail of one-off labels appears beside the handful of themes that recur, and the screen has to collapse everything past the top group into an Other bucket so the real distribution stays visible. Anyone building a taxonomy from conversation data runs into this, which is why a stable label set is worth the argument it costs. The argument is worth having once, in a room, before anyone counts anything, because a label set that keeps shifting makes every week incomparable with the one before it and the trend line stops meaning what it appears to mean. Our own objection taxonomy and how it was cut down came out of exactly that fight.
Pick the objection your team already disagrees about internally, and the recordings will settle it faster than another round of opinion. A month of calls, and the disagreement has an answer.
Most of the output should be a sentence, not a chart
A voice of customer program earns its place when the screen ends in something a person says or writes, not in a percentage. The failure mode of the whole category is well documented by the people who sell into it: a finding is discovered, discussed and forgotten, and the next quarter rediscovers it.
Three of the seven screens in ClientTruth end in an action rather than a figure, and the conversions are deliberately boring. One of them turns a question nobody answered on the call into a page the website is missing.
| What showed up in the conversation | What it becomes |
|---|---|
| An objection that keeps recurring | two sentences a rep can actually say, ready to drop into the script |
| A question buyers ask before anyone pitches them | a section on a page people already land on |
| A request that nobody answered | a line on the product backlog with the conversations attached |
Script lines pulled from objections that actually recur are also the cheapest material for a new rep onboarding plan, since they arrive already phrased the way buyers phrase the problem.
The middle row is the one marketing teams underuse. Questions asked unprompted in a sales call are the exact wording of the thing a buyer wanted to know, before your positioning had a chance to reframe it, and that wording belongs on the page rather than in a research deck. The same material feeds the rep facing side of the picture, which is what the screens that read these calls from the seller's angle are for.
The deal status attached to a customer here is inferred from the conversations rather than read out of a CRM, and the product says so on screen instead of letting the reader assume otherwise, because an early call that sounds enthusiastic is an artifact of a first conversation and not a forecast.
Where the standard guides are right, and where they stop
They are right about the pipeline. Collect from many channels, tag consistently, route findings to an owner, close the loop with the customer. Nothing in this article replaces that. The disagreement is narrower and it is about which channel carries the weight, and about whether a program can report its own reliability.
It cannot, and the strongest version of that point comes from people defending surveys rather than attacking them. Groves and Peytcheva assembled 59 methodological studies in which nonresponse bias could be measured directly against known values, and found that the response rate and the size of the bias barely track each other. A low rate does not prove an estimate is wrong, and a high rate does not prove it is right. That is uncomfortable for anyone presenting survey results to a leadership team, because the response rate on the slide says nothing about how much to trust the rest of the slide.
Recorded conversations do not escape every problem. They tell you nothing about buyers who never took a meeting, and nothing about the churn reasons that only surface a year into the contract. What they remove is the specific failure of asking: the gap between the population you contacted and the population that answered. On a recorded call that gap is zero, because everyone who was in the conversation is in the data.
What to do next
Take last month of recordings and count one thing: how many of those conversations contain a request from the buyer that never received an answer on the call. Not an objection, not a question that got handled, a request that went unmet. Then read the five longest buyer turns inside those calls.
If the count is zero, widen the window to the full quarter before concluding anything, because that is the exact shape the sample size trap takes. If the count is not zero, you have the first standing item of a voice of customer program, and it arrived without a form or a thank you email.
Then do the same with questions. Every question a buyer asked before anyone pitched them is a page you have not written, in wording you did not have to invent. That is a week of work for a marketing team and it is drawn entirely from conversations that already sat in storage.
Nothing to fill in and nothing to send your customers.
- A response rate tells you how many people replied. It does not tell you how wrong the answer is.
- Recorded sales calls carry no response rate at all, because nobody opted in and nobody opted out.
- Every share needs the base the question is about: objections over calls that contained an objection, not over all calls.
- A finding is finished when it reaches a verbatim line, the turns around it, and a place in the recording.
- Most of the output should be a sentence somebody says out loud next week, not a chart.
A live demo of our products and a real conversation about the growth problem you need to solve.
Book a demo