When to Ask the NPS Question, and When Not To
NPS is a relationship metric. Where it belongs, where effort is the better instrument, and the cases where asking it actively misleads.
TL;DR
- NPS measures a relationship, not an interaction. Ask it about your brand, not about a single ticket.
- Its two defensible jobs: relationship tracking over time, and competitive benchmarking against companies asked the same question the same way.
- After a discrete transaction, effort is the better instrument. Use Net Easy Score there.
- NPS misleads in three predictable situations: low-choice categories, immediately after service recovery, and where the sample is not the decision maker.
What the question is actually measuring
"How likely is it that you would recommend [company] to a friend or colleague?" asks the respondent to make a social bet. They are being asked to put their own credibility behind you.
That framing is why the question works at relationship level. Answering it well requires the respondent to summarise everything they know about you: the product, the price, the last time something went wrong, whether the brand is one they want to be associated with. It is a whole-relationship judgement compressed into a number.
It is also why the question fails at interaction level. If you ask it thirty seconds after a customer closes a chat about a password reset, you are asking a whole-relationship question about a five-minute event. The respondent either answers about the relationship, in which case the score tells you nothing about the chat, or answers about the chat, in which case you have collected a badly worded effort score. Both outcomes are bad, and you cannot tell from the data which one you got.
The two jobs NPS does well
Relationship tracking. Ask a defined population on a fixed cadence, quarterly or twice yearly, sampled independently of recent activity. The value is the trend and the movement in the distribution, not the absolute level. This is the use case the metric was designed for in Reichheld's original 2003 Harvard Business Review article.
Competitive benchmarking. The reason NPS survives despite serious academic criticism is standardisation. Same question, same 0 to 10 scale, same grouping rule. That makes cross-company comparison possible in a way bespoke satisfaction indices never allow. The moment you change the wording or the scale to suit your business, you have kept the burden and lost the benefit. See why NPS uses an 11-point scale.
Worth being honest about the limits of the claim underneath both. Keiningham, Cooil, Andreassen and Aksoy (2007, Journal of Marketing), working with longitudinal data from 21 firms and over 15,500 interviews, could not replicate the assertion that Net Promoter predicts growth better than other loyalty measures. Use NPS as a comparable, trackable relationship indicator. Do not sell it internally as a proven growth predictor. That claim does not survive replication.
Most NPS programmes fail because the question was attached to a trigger rather than to a population. Once NPS fires after every order, every ticket and every delivery, the score becomes a function of who transacted this month. A promotion that pulls in bargain hunters moves the score. A billing cycle moves the score. Nothing about the relationship changed. The number moved anyway, and someone will be asked to explain it.
Where effort belongs instead
After a discrete, completed interaction, the useful question is not whether the customer would recommend you. It is whether you made the thing hard.
The effort framing comes from Dixon, Freeman and Toman's 2010 Harvard Business Review article "Stop Trying to Delight Your Customers", based on a study of more than 75,000 people interacting with contact centres and self-service channels. Their finding was that reducing effort predicted loyalty better than delighting customers did.
At Numr we run this as Net Easy Score, calculated on the net method: the percentage of respondents who found it easy minus the percentage who found it difficult. Neutral responses sit out of the calculation. That is deliberate. NES is not an average of a rating scale. Averaging buries the tail, and the difficult tail is the part that generates repeat contacts and churn. The net method keeps the two ends visible and makes the metric behave the way NPS behaves, which matters when both sit on the same dashboard.
The division is clean:
Question | Level | Ask after |
|---|---|---|
Recommendation (NPS) | Relationship | Fixed cadence, sampled from the base |
Net Easy Score | Interaction | A completed, discrete task |
Satisfaction (CSAT) | Interaction or attribute | A specific element you can change |
Three cases where NPS actively misleads
Low-choice categories. Utilities, regulated telecoms, statutory services, monopoly infrastructure. The recommendation question presumes the respondent has a real alternative to recommend against. Where switching is impractical or the category is compulsory, scores compress and the metric loses discriminating power. It will still produce a number. The number will not distinguish good operators from bad ones. Track effort and resolution instead, and use NPS only against your own history.
Immediately after service recovery. A well-handled complaint produces a short-lived gratitude spike. Fire the recommendation question inside that window and you capture relief, not loyalty. It reads as improvement and it decays. If you must measure recovery, measure whether the issue actually stayed fixed, on a delay, and measure effort at the point of contact.
Where the respondent is not the decision maker. Common in B2B. The daily user of a platform may have no influence over renewal, and the economic buyer may never log in. An NPS built from end users can stay high through a churn event. Segment by role and treat the buyer-side reading as the one connected to revenue.
A fourth, quieter failure: over-surveying. Once the same customer receives the question from four different triggers, response rates fall and the responding sample skews toward the most engaged and the most angry. See survey fatigue if that page is live in your set.
Amitayu Basu, CEO, NumrThe question I ask a client before agreeing to add an NPS trigger is simple. If this score drops four points next quarter, what will you do differently on Monday? If nobody can answer, the survey should not exist. Most of the triggers we inherit fail that test. They generate a number that gets reported and never acted on, and they burn the customer's willingness to answer the one survey that would have mattered.
A workable placement rule
- Pick the population, not the trigger. Define who the relationship survey goes to and sample them on a cadence, independent of recent activity.
- Cap contact frequency across the whole programme. One customer, one survey per window, whichever team owns it.
- Put effort at the interaction points. Post-contact, post-delivery, post-onboarding. Use the net method.
- Add one open-ended follow-up. The score tells you the direction. The verbatim tells you the reason. Without it you are reporting a number you cannot explain.
- Segment before you compare. New versus tenured, channel, market, and in B2B, role.
- State the caveats in the report. Category choice, response-style differences between markets, sample skew. A score presented without them will be over-read.
Frequently asked questions
Can I ask NPS after every transaction? You can, but it is not what the metric is for, and the resulting score will move with transaction mix rather than with the relationship. If you need per-interaction feedback, ask an effort question and keep relationship NPS on a separate, sampled cadence.
Is transactional NPS invalid then? Not invalid, but it should be treated as a distinct series with its own baseline and never averaged with or benchmarked against relationship NPS. They answer different questions about different populations.
How often should I ask the relationship question? Quarterly for most consumer businesses, twice yearly for long-cycle B2B. The constraint is whether enough can change between waves to make a move meaningful. Asking monthly usually produces noise you will be tempted to explain.
What should I ask instead after a support ticket? An effort question, scored as Net Easy Score: percent who found it easy minus percent who found it difficult. Add a free-text follow-up asking what would have made it easier.
Does NPS work in B2B? Yes, with segmentation by role. Score the economic buyer, the day-to-day user and the technical owner separately. A blended B2B NPS often hides exactly the signal you needed.
Is it wrong to report NPS as a percentage? Yes. NPS is a points figure running from -100 to +100. Write "an NPS of 31" or "31 points". The promoter and detractor shares are percentages; the net result is not.
Should NPS be tied to bonuses? Be careful. A score attached to compensation invites gaming at the point of collection, from coaching the customer to selective sending. If you do it, tie the incentive to the closed-loop follow-up rate and to verified issue resolution rather than to the score itself.
We operate in a regulated monopoly. Should we run NPS at all? Only against your own trend, and only alongside effort and resolution measures. Cross-company benchmarks in low-choice categories are not meaningful, because the question presumes a choice the respondent does not have.
Related guides
For related reading, see How to Design a CX Survey That Customers Actually Complete, Building the Anatomy of a High-Converting Survey: Mixing Close-Ended and Open-Ended Questions, How to Avoid Survey Bias in CX Research, How to Build a Survey Governance Process for Enterprise CX, Response Rate vs Completion Rate: What’s the Difference, and What Is a Likert Scale? Types, Examples, and How to Analyze Likert Data.
