Listening programmes, not surveys
A survey is an instrument; a listening programme is a system. The difference shows in what happens after the answers arrive. A business with surveys has scores. A business with a programme has decisions that depend on those scores, people who own them, rules for who gets asked what and when, and a habit of telling customers what changed.
A programme draws on several sources at once. Solicited feedback comes from surveys the business sends. Unsolicited feedback arrives anyway: support contacts, complaints, reviews, social posts, sales and account notes. Each source has a bias. Surveys hear from people willing to answer; reviews over-represent the very pleased and the very annoyed; support hears only from people with problems. Reading them together corrects for each one's blind spots. Scores say how many customers feel something; for questions of direction, the guide to running a customer advisory board adds a standing group of customers who debate plans in progress. Remarks that staff hear across a counter or over the phone, when they are not requests for help, are a further source of unsolicited feedback, and recording what customers tell your staff is the tutorial for writing them down so that they can be read beside the others.
The design starts from decisions, as the measurement hub argues. For each survey, write down which team will act on it and what they would change if the score moved. A survey with no answer to that question should not be sent. Applying the KPI tree framework helps here: the retention outcome sits at the top, and each perception measure should connect to a driver beneath it. A relationship score that cannot be placed on the tree is measuring something nobody manages.
Ethics and data protection apply throughout. ESOMAR's professional codes for research set out principles of honesty with respondents, voluntary participation and care with personal data. Where customers are in the European Union, the GDPR governs how survey responses linked to identifiable customers can be collected, used and kept. Feedback that is used to follow up with individuals, or joined to their account, is personal data, and customers should be told how it will be used. This is general direction rather than legal advice.
NPS, CSAT, CES: what each measures
Three survey measures dominate customer experience work. They look similar but measure different things.
| Measure | The question | What it captures | When to ask | Strength | Weakness |
|---|---|---|---|---|---|
| Net Promoter Score (NPS) | How likely are you to recommend us to a friend or colleague? | Overall attitude to the relationship | Periodically, not after a single interaction | Simple; widely understood; tracks the relationship over time | Coarse; hard to act on alone; sensitive to culture and sampling |
| Customer Satisfaction (CSAT) | How satisfied were you with this interaction or purchase? | Satisfaction with a specific moment | Straight after the interaction | Specific and actionable; easy to tie to a team | Says little about the overall relationship; ceiling effects |
| Customer Effort Score (CES) | How easy was it to get what you needed? | Effort a task or interaction required | After support contacts, onboarding or self-service tasks | Strongly linked to friction the business can remove | Narrow; less meaningful for experiences that are meant to be enjoyable |
NPS is calculated by grouping respondents into promoters at the top of the scale, detractors at the bottom and passives in between, then subtracting the share of detractors from the share of promoters. The calculation throws away detail, which is one reason the score moves erratically on small samples. Read the distribution, not just the net figure.
CSAT is usually reported as the share of respondents choosing the top satisfaction options. Because most customers who bother to answer a transactional survey are satisfied, CSAT tends to sit high, and its value lies in the low scores and what they say.
CES asks about ease rather than delight. It is particularly useful for support and self-service, where the goal is to remove effort rather than to impress. Many businesses find that reducing effort does more for retention than adding moments of delight.
Relationship vs transactional surveys
Surveys fall into two families, and confusing them is a common source of misleading results.
Relationship surveys ask about the customer's overall view of the business. They are sent on a schedule to a sample of the customer base, independent of any particular interaction. NPS is the usual measure, often with a few supporting questions about the main parts of the experience. They show the trend in how customers feel and, read alongside behaviour, which customers are at risk. Their weakness is that they are hard to act on: a falling relationship score says something is wrong, rarely what.
Transactional surveys ask about a specific event: a purchase, a delivery, a support contact, an onboarding step. They are triggered by the event and sent soon after it, while the experience is fresh. CSAT and CES are the usual measures. They point at a team and a process, which makes them actionable, but they tell you only about the moments you chose to survey.
A sound programme uses both: a small number of transactional surveys at the moments that matter most, and a periodic relationship survey that shows the overall direction. Questionnaire craft, from wording to scale design and question order, belongs to the survey and questionnaire design guide in the marketing research root. For research into why visitors do not convert on a website, as opposed to how existing customers feel, the conversion research methods guide covers on-site polls and user testing aimed at CVRShare of visits or clicks that convert.. The tutorial on on-page polls and post-purchase surveys borrows this guide's open-text coding for a single question asked at one step of the purchase, where the answers feed a conversion backlog rather than a score.
Timing, sampling and survey fatigue
When and whom a business asks shapes the answer as much as the question does.
Timing. Transactional surveys should arrive soon after the event, but not before the outcome is known: a delivery survey sent before the parcel arrives measures hope, and a support survey sent before the problem is fixed measures the first reply. Relationship surveys should avoid moments that distort them, such as just after a price rise or during a known outage, unless that is what the business wants to measure.
Sampling. A survey sent to everyone and answered by the willing is not a sample of customers; it is a sample of people who answer surveys. Compare respondents with the customer base on tenure, value and segment, and weight or report separately where they differ. Small segments need enough responses to read; reporting a score for a group of a handful of people invites conclusions the data cannot support. For a relationship survey drawn from the customer base, the tutorial on drawing a survey sample from a customer list stratifies the list by tenure or value and applies the contact rules before anyone is invited.
Fatigue. Customers asked too often stop answering, and those who keep answering become less representative. Contact rules prevent this: a customer surveyed recently is not surveyed again for a set period, regardless of which team wants to ask. A contact ruleA business-wide rule that a customer surveyed or interviewed recently is not asked to take part again for a set period, whichever team wants to ask. only holds when every team's invitations go into one record of who was asked and when. Nielsen Norman Group's guidance on survey design makes a related point about length: every extra question costs completion, so ask only what the business will use.
Survey changes can be tested. Sending two versions of a survey invitation or question order to random halves of the sample is an A/B TestControlled comparison of two variants., and it shows whether a change affects response rates or scores before it is rolled out.
Reading results and open text
Scores are the start of the reading, not the end. Three habits make results useful.
Read trends and distributions, not single numbers. A month-to-month movement within the normal range of noise means nothing. Track the score over a longer period, show how much it varies, and look at how the distribution moves: a fall in the share of top scores is a different problem from a rise in the share of bottom ones. A control chart draws that normal range from the score's own history, so a movement is flagged only when it falls outside what the score usually does. When a relationship score has fallen for several waves, the diagnosis for a relationship survey score that is falling separates that noise from a shifted sample, a distorting send moment, rising effort and service failures before the score is read as a verdict.
Segment before concluding. An overall score can hold steady while new customers grow unhappier and long-standing ones grow happier. Split by tenure, product, channel and value. High-value customers deserve their own view, because a dissatisfied customer with high LTVTotal value a customer brings over time. represents more at risk.
Read the open text. The comment box is the most valuable part of most surveys. Code comments into themes, either by hand for small volumes or with text analysis for large ones, and check automated themes by reading samples. The tutorial on coding open-text answers by hand builds the codeframeThe defined set of codes open-text answers are tagged with, built from a sample, applied consistently, checked by a second coder and counted against a base., checks it with a second coder and counts respondents per theme against the base. Where an AI tool does the coding, the guide to synthesising feedback with AI shows how to keep quotes exact and counts taken from the coded table rather than the tool's summary. Count themes over time and link them to scores: the themes most common among low scorers are the causes worth fixing. Quote customers directly in reports, because a sentence in a customer's words persuades a product team faster than a chart.
The final reading is against behaviour. Join survey responses to what the same customers did afterwards, and check whether low scorers leave more often. The churn and retention metrics guide covers the behavioural side; if perception does not predict behaviour, the perception measure needs rethinking. A dashboard that sets scores, themes and behaviour side by side, built along the lines of the dashboard design guide, keeps the reading honest.
Closing the loop: inner and outer
Feedback that changes nothing teaches customers to stop giving it. Closing the loop has two halves, and both need a named owner and a time limit.
The inner loop goes back to the individual customer. Low scores and worrying comments trigger a follow-up from a person, within a defined time, who acknowledges the feedback and, where possible, fixes the problem. The rule for which responses trigger a follow-up should be written down, and the follow-up recorded so the business can see whether it happened. Customers must have agreed to be contacted about their feedback, and anonymous surveys must stay anonymous. The promise a survey makes about names is settled before it is sent, and the guide to choosing that promise sets out which follow-ups and links to account records an anonymous, a confidential and an attributable survey each allow. For a repeated survey, the customer survey wave SOP makes the follow-ups one step of every wave, with an owner and a record beside the draw, the coding and the chart.
The outer loop goes to the teams. Themes from open text and the causes behind low scores are reviewed on a fixed cadence by the people who own the parts of the experience concerned. Each review produces actions with owners and dates, and the next review starts by checking them. Track the outer loop with its own KPIA metric tied directly to an objective.: actions agreed, actions completed, and whether the related theme declined afterwards.
Telling customers what changed completes the loop. A short note to people who raised an issue, or a public "you told us, we did" update, raises future response rates and shows that answering was worth the effort.
An insurer finds the real problem behind a flat score
Brightmoor Insurance, a home insurer with about 120,000 policyholders, sends a relationship NPS survey every quarter and has scored between +18 and +22 for two years. Nobody can say what would move it. The team adds a CES survey after claims and a CSAT survey after renewals, both capped at one survey per customer per quarter. Open-text coding of claims surveys shows that 35% of comments from claimants who found the process difficult mention sending the same documents more than once. The claims team introduces a single upload link that stays open for the life of the claim. Within two quarters the share of claimants rating the process easy rises from 52% to 71%, and customers who made a claim renew at a rate about six points higher than the year before.
Satisfaction scores rarely explain why a customer left; Jobs-to-be-Done for retention complements them with switch interviews that separate jobs never done from jobs the product stopped doing.
Common mistakes
- Surveys without decisions. A score that no team owns and no decision depends on is noise.
- Surveying too often. Fatigue lowers response and skews who answers; set contact rules across teams.
- Mixing relationship and transactional measures. An NPS question after every support chat measures the chat, not the relationship.
- Reacting to noise. Small month-to-month movements on small samples rarely mean anything.
- Ignoring open text. The comment box explains the score; skipping it leaves the business guessing.
- No inner loop. Customers who give low scores and hear nothing learn that feedback is pointless.
- Never checking against behaviour. A perception score that does not predict retention is measuring the wrong thing.
- Careless use of personal data. Linking responses to accounts or following up without consent breaches trust and, in many places, the law.
