On this page
Churn research means interviewing users who have left or are leaving, starting from the last day they used the product and reconstructing what happened, rather than asking "Why did you stop using it?" Behavior data tells you at which step users leave; interviews tell you why. After reading this guide, you'll be able to select participants by churn type, run 10 to 20 interviews with a ready-made interview guide, and turn your findings into changes you can ship and validate.
When to use it and how to set the research goal
Churn is not a single situation. Each type calls for a different research focus and different questions, so sort out which one you're studying before you start.
| Churn type | Typical signal | Research focus |
|---|---|---|
| Early churn | Doesn't come back within a week of signing up or first use | Where they got stuck the first time, and the gap between what they expected and what they got |
| Lapsed users | Used to use it regularly, recent activity has dropped sharply | When and why their usage habits changed, and what replaced the product |
| Subscription cancellations | Turned off auto-renewal or asked to cancel | What happened before and after canceling: not enough value, a pricing issue, or the need went away |
Write the research goal as a question that can actually be answered, ideally focused on one specific churn point. "Understand why users churn" is too broad; "What did new players who stopped around level 3 in their first week run into before leaving?" can directly shape your guide and your sample. Focus each study on one churn type. If you want to understand all three, split them into three studies.
Define churn in your study plan, for example how many consecutive inactive days count as churned. Otherwise your sample selection and your interpretation of the findings won't use the same definition.
Who to interview: select by behavior
Select participants from behavior data, not from survey answers where people say they "don't use it much anymore." The criteria below are examples; adjust them to how often your product is used:
- Early churn: users whose last activity was within 7 days of signing up and who stopped around the same key point.
- Lapsed users: users who used the product regularly in at least 2 of the past 3 months and whose activity dropped by more than half in the last 30 days.
- Cancellations: users who canceled in the last 2 to 4 weeks. If they left too long ago, they won't remember the details.
- Comparison group: users who joined in the same period and are still active. This group can be smaller; it tells you whether a problem is unique to churned users.
Start with 10 to 15 interviews per type. Guest, Bunce & Johnson (2006) found that with a homogeneous group, new themes become noticeably rarer after about 12 interviews; when groups differ a lot, cover each group separately. Users who have already left are less likely to respond, so state the interview length and incentive in your invitation and make clear this is not a win-back call.
Interview guide
The core approach is to start from the last time the participant used the product, work backward in time, have them describe what actually happened, and probe at the key moments. You can copy the guide below as is; replace the text in parentheses with your product.
- [Opening] Thanks for making the time. Today I'd like to hear about your real experience using (product). There are no right or wrong answers, and the things that didn't work for you are the most helpful for us.
- [Opening] We won't try to sell you anything or talk you into coming back. The whole conversation takes about 15 minutes.
- [Core topic: last use] Do you remember the last time you opened (product)? What were you doing?
- [Core topic: last use] After that session, what happened that kept you from opening it again?
- [Core topic: original expectations] Why did you start using it in the first place? What were you hoping it would help you with?
- [Core topic: original expectations] Once you'd used it for a while, what turned out the way you expected, and what didn't?
- [Core topic: the letdown] Was there a moment when you felt you didn't want to keep going? Can you walk me through what happened?
- [Core topic: alternatives] Since you stopped using it, how do you handle this now? What do you use instead?
- [Core topic: cancellations only] On the day you decided to cancel, what was on your mind? How long had you been thinking about it?
- [Probes] Ask for a concrete example: when was the last time this happened?
- [Probes] Turn vague answers (hard to use, too expensive, no time) into a specific situation: which step exactly? Expensive compared with what?
- [Probes] Ask about actions: what did you do at that point? Did you try to get help?
- [Probes] Check how much it weighs: if this problem were fixed, would you come back? Were there any other reasons?
- [Closing] If you could change just one thing, what would it be?
- [Closing] Is there anything I didn't ask about that you think matters? Thank you.
Don't open with "Why did you stop using it?" Participants will give you the easiest answer, such as "I didn't have time," and it becomes hard to take the conversation further.
Sample dialogue: from "it's not fun" to a specific moment
This is a fictional example. A vague "it's not fun" is traced to a specific game mechanic, and you also learn something in the other direction: the participant doesn't dislike the game.
How to synthesize findings
- Start by building a timeline for each participant: when they started using the product, key events, their last use, and what they switched to after leaving.
- Group the reasons for leaving into themes, and separate the direct cause (the event that triggered leaving) from background causes (ongoing frustrations that weren't deal-breakers on their own).
- For each finding, state how many participants mentioned it, include 2 to 3 quotes and at least one counterexample. Example: 7 of 18 participants said they left after getting stuck on a level, while 2 said getting stuck was not a problem.
- Break results down by segment. The same reason can account for a very different share among new and long-time users.
- Tie each finding to one change you can ship and one metric to validate it.
If you interview only churned users, the share of each reason reflects only how reasons are distributed among churned users. It can't be generalized to all users. State this limitation in your findings.
Common mistakes
- Asking "Why did you churn?" directly: you mostly get reasons constructed after the fact. Start from what happened during the last use instead.
- Taking the first reason at face value: "I didn't have time" and "It's too expensive" are often polite answers. Ask about the specific situation and what they are comparing against.
- Mixing all three types of churn in one study: new users, lapsed users and cancellations leave for very different reasons, and when you combine them you can't tell whose problem is whose.
- Inviting users who left long ago: people who left six months ago don't remember the details and can only answer from general impressions.
- No comparison group: users who stayed may have hit the same problem without leaving, in which case it isn't the main reason.
- Turning the interview into a win-back attempt: if you offer discounts or try to persuade people to come back, participants switch into polite mode and you can't trust the rest of their answers.
Validate findings with behavior data
Findings from interviews are hypotheses that need to be confirmed with behavior data. You can do this in three steps:
- Before release, compare against the data: check the points participants mentioned. If participants said they got stuck on a certain level, see whether churned users failed that level noticeably more often than users who stayed.
- After release, validate: make changes based on the findings and set the metrics and observation window in advance. After reworking a tutorial, track that level's pass rate and first-week retention; for a cancellation study, track renewal and cancellation rates.
- Go back when words and behavior don't match: if many people mention a reason but the data shows no difference, the reason may just be easy to say, and the real one hasn't come up yet.
Change one thing at a time, or run an A/B test. Otherwise you can't tell which change moved retention.
How to do it in heari
With heari, you can select users in ThinkingAI Agentic Engine by behavior data, such as users who churned in their first week, whose activity dropped, or who just canceled a subscription, and AI drafts the study plan and interview guide. After the research lead reviews it, AI runs many interviews at once, starting from the last use and asking follow-up questions based on each answer. The insight report groups findings by theme, and each finding comes with quotes, counterexamples and participant counts. Once changes ship, you validate the findings against behavior data.
FAQ
How many churn interviews do you need?
Plan 10 to 15 interviews per churn type. Guest, Bunce & Johnson (2006) found that with a homogeneous group, new themes drop off noticeably after about 12 interviews. If participants come from very different groups, run that many interviews for each group.
What if churned users don't want to be interviewed?
In the invitation, state the interview length and incentive, and make clear this isn't a sales or win-back call; you just want to hear about their real experience. Prioritize users who left in the last 2 to 4 weeks. They remember what happened more clearly and give more specific answers.
Can a survey replace churn research?
Surveys are useful for measuring how common each reason is once you already know the reasons, not for discovering them. Open-ended answers in churn surveys tend to be vague, like "hard to use" or "too expensive," and you need interview follow-up questions to get to the specific situation and step.
Why should churn interviews start from the last time someone used the product?
Specific events are more reliable than general opinions. When participants recall their last session, they naturally describe what they did, how they felt and what went wrong. That information is more concrete than the answer to a direct "why" question, and it is easier to match against behavior data.
