Voice of the customer, from feedback to action

Every company has some version of this: a survey after purchase, a score on a dashboard, a quarterly deck. The tooling is mature, the collection is solved, and most programmes still produce very little that changes a decision.

The reason is rarely the instrument. It is that the programme hears from a specific and unrepresentative slice of customers, and nobody has looked carefully at who is not in it.

Three sources, and their different biases

The customers most likely to answer are the delighted and the furious. Everybody in between is where the revenue actually lives.

Who never appears in your data

Two populations, and both are systematically absent from anything collected through your own channels.

The silently dissatisfied. Most unhappy customers never complain — they reduce usage, stop recommending, and leave. Because they never file anything, they are invisible in unsolicited feedback, and because they are disengaged, they do not answer surveys either. The single most valuable population in customer research is the hardest to hear from, and that is not a coincidence.

Churned customers. The moment somebody leaves, most systems stop surveying them, precisely when their answer is most informative. Exit surveys catch a fraction and are answered under conditions that produce polite, uninformative reasons. "Too expensive" is what people say when they do not want the conversation to continue.

There is a third that matters if you operate internationally: anyone in a market where you have no support presence and no survey in their language. Their absence looks identical to satisfaction on a dashboard.

NPS, honestly

The score is a weak instrument and the comments underneath it are the asset.

As a management number it compresses a distribution into one figure, moves for reasons that have nothing to do with the product, and is not comparable across markets — the same underlying sentiment produces different scores in different countries because scale use differs culturally. Treating a two-point movement as a signal is common and mostly wrong.

The verbatims are a different matter. Open-ended responses are the only part of most VoC programmes that contains anything new, and they are the part most often left unread because analysing them takes work that summarising a number does not.

If you keep one thing, keep the comments. If you have budget for one improvement, spend it on reading them properly rather than on collecting more of them.

A voice of customer program is the loop, not the tool

Collection is not a programme. The loop is: collect, analyse, decide, act, and tell the people who told you.

That final step is skipped almost universally and it is the one that determines whether the programme still works in year two. Customers who answer and see nothing change learn that answering is pointless, and they stop. Your response rate falls, your remaining respondents become less representative, and the data quietly degrades while the dashboard keeps reporting.

1. Fewer questions, asked closer to the event

   A two-question survey at the moment something happened beats a fifteen-question relationship survey nobody finishes. Length is bought with response rate, and response rate is bought with representativeness.

2. An owner for each theme

   A finding with no name attached to it does not become a change. This is an organisational design problem wearing a research costume.

3. A visible action

   Publish what changed because of feedback, even when it is small. This is the cheapest thing on this list and the one that keeps response rates from collapsing.

4. An incentive where you are asking for real time

   Goodwill covers a two-question survey and not a twenty-minute one. What to pay and why sets out the arithmetic and the forms that quietly exclude people.

5. A route for what cannot be fixed

   Some findings will not be actioned, and saying so honestly is better than silence. Silence is indistinguishable from not having read it.

The point where a survey stops working

Surveys reach people on your list who are willing to answer. When the question concerns anybody outside that description, the instrument has to change.

Four cases where recruitment is the right answer rather than another survey wave.

Lapsed and churned customers, reached independently of your own mailing list, with an incentive that acknowledges you are asking for time from somebody who has already left.

Non-customers who considered you and chose something else. The highest-value population in the field, and one that appears in no company's own data by definition.

Markets where you have no presence. Reaching people in a country you are considering entering is a research problem, not a customer-communication problem.

And specific professional segments, where the person who matters is a role rather than an account holder. How professional and market-specific recruitment actually works covers that mechanism, and where the profile is senior enough that the format becomes a scheduled conversation, what expert networks charge and screen for is the comparison worth making.

Reading verbatims so they produce decisions

Open-ended responses are where the information is and where most programmes stall, because reading them properly is work and counting them is not.

Code them against a scheme you wrote before reading, and revise the scheme deliberately rather than drifting. A code list that grows organically as somebody works through a backlog produces categories that overlap, which makes the resulting counts meaningless and the trend over time uninterpretable.

Separate the problem from the sentiment. "The export takes four clicks" and "I am fed up with this product" are different findings, and a scheme that only captures tone loses the first entirely — which is the one somebody could act on this week.

Count distinct customers rather than distinct mentions. One articulate person filing eleven comments about the same issue is one customer, and a tally by mention hands the roadmap to whoever writes the most.

And keep the raw text. Summaries answer the question that was current when they were written; the underlying responses can answer a different question in six months, and re-collecting is far more expensive than re-reading.

Markets, and what translation removes

A programme running in eleven countries and analysing everything in English has already lost most of what it collected.

Register does not survive machine translation. The difference between a mild complaint and a serious one, between politeness and resignation, between sarcasm and praise, lives in exactly the layer that translation flattens. A verbatim rendered into competent English reads as a neutral statement whether the original was furious or fond.

Scale use differs too, which is why a satisfaction number cannot be compared across markets without adjustment that almost nobody makes.

The practical version is to have open-ended responses read and coded by somebody native to the market, against the same coding scheme, and to compare within markets over time rather than between markets at a point in time. How each layer of targeting narrows the pool and moves the rate sets out what native-speaker eligibility costs, and it is less than the analysis it makes possible.

Machine-translating verbatims does not give you eleven markets' feedback. It gives you one market's reading of eleven.

What it costs, and where the money should go

The survey platform is the smallest line in a serious programme and the one that absorbs most of the attention.

The allocation worth arguing for is less collection and more of everything downstream of it. Most programmes have more data than they have read, and adding another wave to a backlog nobody has analysed is a way of appearing to invest without learning anything.

What to ask a vendor

Five questions that separate a platform from a programme.

Which populations can this reach, and which can it not? A platform that surveys your own list should say so plainly rather than describing itself as customer intelligence.

How are open-ended responses handled — coded by people, by a model, or handed over raw? All three are defensible; the answer changes what you should expect from the output.

How is multi-market handled: native-language coding, or translation into English first?

What is the evidence trail behind each response, so a finding can be re-examined later against a different question? What counts as evidence and how to specify it up front covers the formats.

And what does the vendor do about response bias, concretely? "We use a representative panel" is a claim; ask how representativeness is checked and what happens when a segment stops responding.

Where this sits

A VoC programme answers what customers say. It does not answer whether they can complete a task, which is usability testing, or whether a message was understood, which is creative testing. Those get conflated in planning documents, and a programme asked to do all three does none well.

What it uniquely provides is continuity — the same questions asked over time, so a change is visible as a change rather than as a one-off finding. That is worth protecting, which is the argument for keeping the instrument stable and spending the marginal budget on the populations it cannot reach. How a brief becomes reserved capacity and verified responses describes that mechanism, including the markets where recruiting is genuinely hard and we will say so before you commit.

Where a specific question needs answering rather than a continuous signal maintaining, which research method fits which question is the faster route to a design.

For the instrument itself rather than the programme around it, how a research survey should be designed and sampled covers the question wording and the sampling that decide whether the numbers mean anything.