On this page
The reasons behind low NPS scores only come out when you ask follow-up questions in an interview: start from the moment the participant gave the score, get to the specific step, compare expectations with what actually happened, and then ask what would change their score. After reading this guide, you'll be able to select participants, copy the interview guide to run your follow-ups, and synthesize the answers into findings grouped by cause, split into what you can fix now and what you can't, and ready to validate with behavior data.
When to use it
NPS (Net Promoter Score) was introduced by Fred Reichheld in a 2003 Harvard Business Review article. It asks a single question: How likely are you to recommend us to a friend or colleague, on a scale of 0 to 10? Scores of 9 to 10 are promoters, 7 to 8 are passives, and 0 to 6 are detractors. NPS is the percentage of promoters minus the percentage of detractors. Its strength is that it's simple and easy to track over time. Its weakness comes from the same place: a single number can't explain why.
- NPS keeps falling, or scores are clearly lower for a particular channel or type of user
- Open-ended survey responses mostly say vague things like "the experience was average" or "the service was bad"
- Scores fluctuate after a redesign or price change, and you need to know whether the change is the cause
- You need to explain to leadership why the score changed, not just report the number
Surveys can't get at the cause because the people giving the score often can't articulate it themselves. What they remember is a feeling, and someone needs to follow up on their answer to get to a specific experience.
How to set the research goal
Don't write the goal as "Understand why users give low scores." It's too broad, and the interviews will wander. Write it as a question that can be answered and that tells you what to do next once it is.
| Weak | Better |
|---|---|
| Understand user satisfaction | For new users who scored 0 to 6 in the past month, which steps the low scores come from |
| Improve NPS | Determine whether last quarter's price increase is the main reason detractors went up |
| Collect user feedback | Find the step in the application flow that frustrates detractors most, and hand it to the product team to schedule |
Each study should answer only one or two questions. The more questions you have, the shallower the follow-up on each one.
Who to interview: select by score and behavior
Interviewing only detractors makes the findings one-sided. Talk to all three groups, weighted toward detractors. The shares in the table below are examples; adjust them to your research goal.
| Segment | Selection criteria | Suggested share (example) |
|---|---|---|
| Detractors (0 to 6) | Scored in the last 30 days and has a complete usage history | About half |
| Passives (7 to 8) | Same as above; prioritize people whose usage is declining | About 30% |
| Promoters (9 to 10) | Same as above; prioritize people who recently had a similar experience | About 20% |
Behavioral criteria narrow things down better than scores alone. For example, "gave a low score and abandoned an application partway through" and "gave a low score but still uses it weekly" usually have different causes: the first group most likely got stuck in the flow, while the second is more likely bothered by a small, recurring annoyance. Guest, Bunce & Johnson (2006) found that with a homogeneous group, new themes drop off noticeably after about 12 interviews, so plan about 12 interviews per key group.
What to ask each of the three groups
- Detractors: which experience frustrated them, where they got stuck, how it was resolved, and what would change their score
- Passives: what's keeping them from a 9, whether they've considered switching, and what keeps them around
- Promoters: what they actually value, how they describe it when recommending it, and whether there was a time they almost didn't recommend it
Try to interview within one to two weeks of the score. The longer you wait, the less participants remember about what happened.
Four steps to follow up
- Start from the moment they scored: what were they doing, and what had just happened
- Find the specific point: which step, page or conversation let them down
- Compare expectation with reality: what they expected, and what actually happened
- Ask what would change the score: what would it take to rate higher
- [Opening] Thanks for making the time. Today I'd like to hear about your recent experience. There are no right answers; feel free to share the good and the bad.
- [Opening] When did you last use our product? What were you trying to get done?
- [Core topic: back to the moment of scoring] A few days ago you gave us a score of X. When you gave that score, which experience were you thinking of?
- [Core topic: the specific step] Which step did things start going wrong? What was happening on the screen or on the phone call at that point?
- [Core topic: expectations vs. reality] What did you expect to happen? What actually happened?
- [Core topic: consequences] How did it get resolved in the end? How much time did it take, and did you have to ask anyone for help?
- [Core topic: passives] What's keeping it from being a 9? Was there ever a time you thought about switching to someone else?
- [Core topic: promoters] Have you ever recommended us to anyone? What did you say?
- [Probes] You said it was "not very convenient." What exactly was inconvenient?
- [Probes] How many times has this happened? When was the most recent time?
- [Probes] What did you do at that point? (Ask about what they did, not hypotheticals.)
- [Probes] If this step doesn't change, what will you do next?
- [Closing] If you could change just one thing so you'd give us a higher score, what would it be?
- [Closing] Is there anything I didn't ask about that you think matters? Thank you for your time.
Example: following up
The first answer, "I waited several days," doesn't tell you what to fix. After follow-up questions, the problem narrows to "the notice asking for more documents wasn't clear," and probing further surfaces what the participant expected. That's something a specific team can act on. Notice that the follow-up questions never include anything like "Do you think our review is slow?" A question like that pushes the participant toward a preset answer.
How to synthesize findings
- For each participant, mark the specific steps they mentioned, then merge similar ones into categories, such as unclear document requirements, no visibility into progress, and price. Group by what participants describe, not by company department.
- For each category, note how many people mentioned it; one person can mention several categories. Example: 9 of 20 detractors mentioned unclear document requirements.
- Include two or three representative quotes for each category, plus at least one counterexample.
- Separate causes you can fix from those you can't fix for now. See the table below for how to handle each.
- List the participants who disagreed, so you don't overgeneralize.
| Type | Criteria | How to handle it |
|---|---|---|
| Fix now | Copy, notifications, step order; can be done within one iteration | Schedule it directly, naming the owning team and the validation metric |
| Fixable with investment | Requires system changes or cross-team work | Estimate how many people are affected and the cost, then add it to the roadmap |
| Can't fix for now | Constrained by regulation, contracts or cost, for example required identity verification | Keep the rule, change the explanation: tell users the reason and how long it will take in advance |
| Not recommended | Mentioned by very few people, or runs against what most users need | Log it and keep watching in the next round of research |
How to validate with behavior data
Interviews explain the cause; behavior data shows how many people the cause affects and whether things changed after a fix.
- Before interviews: look at the abandonment rate at each step of the flow to see whether low scorers cluster at one step, and use that to set the interview focus
- After interviews: compare what participants described with system records; for someone who said they waited several days, check whether the records show the case actually sat for several days
- After changes: set one validation metric per finding. For example, for unclear document requirements, track the completion rate of the document resubmission step and the number of related support calls.
- NPS itself moves slowly and is affected by many factors, so don't use it as the only metric. In the next survey round, check whether the share of people mentioning the same cause has gone down.
Common mistakes
Common mistakes
- Looking at the score without the reasons: the score drops two points and a meeting is called to discuss fixes. First find out which users and which step are pulling the score down.
- Interviewing only detractors: it's easy to mistake a few people's strong dissatisfaction for a widespread problem. Interview passives and promoters as a comparison.
- Asking directly why they gave a low score: participants will only give you general reasons. Start from one specific experience.
- Putting words in the participant's mouth: for example, asking "Was it because it was too slow?" Let participants describe it in their own words.
- Waiting a month or two after the score to interview: the details have faded and answers turn into general impressions. Reach out as soon as you can.
- Sending every cause to the product team: some causes belong to customer support, operations or policy. Assign each one to the team that owns it.
How to do it in heari
In heari, you can select participants with behavior data from ThinkingAI Agentic Engine, for example users who scored 0 to 6 in the last 30 days and abandoned an application partway through. AI drafts the study plan and interview guide, the research lead reviews and launches it, and AI asks follow-up questions in real time based on each answer, running many interviews at once. The report groups findings by cause and shows participant counts, each finding comes with quotes and counterexamples, and once changes ship you validate the findings with behavior data.
FAQ
Who created NPS, and how is it calculated?
NPS was introduced by Fred Reichheld in 2003. It asks users to rate how likely they are to recommend a product on a scale of 0 to 10. Scores of 9 to 10 are promoters, 7 to 8 are passives, and 0 to 6 are detractors. NPS equals the percentage of promoters minus the percentage of detractors, and ranges from -100 to 100.
How many people should you interview to follow up on low NPS scores?
As a reference, Guest, Bunce & Johnson (2006) found that after about 12 interviews with one type of person, few new themes emerge. If you want to analyze several groups, such as new and long-time users, about 12 interviews per group is safer. The more groups, the more interviews in total.
What if users who gave low scores don't want to be interviewed?
In the invitation, explain that the goal is to fix specific problems, not to sell or win them back, and keep it to 10 to 15 minutes. Text interviews are easier for people who don't want to talk on the phone. A thank-you gift is fine, but keep it modest so you don't attract people who only come for the gift.
What if interview findings disagree with NPS open-ended responses?
Answers to open-ended questions tend to be general, while interview follow-up questions get to the specific step, so the two are often different layers of the same issue. If they truly conflict, check the behavior data first, for example the abandonment rate at the relevant step, before deciding which to trust.
