Do bad Intercom ratings predict churn? How to check
How to test whether Intercom conversation ratings predict churn in your data: export ratings, join them to billing, compare churn rates and avoid traps.
By Florian
It seems obvious that a customer who rates a support conversation badly is more likely to cancel. Sometimes it is true. Sometimes the unhappy raters are your most engaged customers, the ones who care enough to complain, and the quiet ones are the ones who leave. You only find out by checking against your own billing data.
This guide shows how to run that check with Intercom’s conversation ratings: where the data lives, how to join it to churn, and how to read the result without fooling yourself. It sticks to Intercom’s help center on conversation ratings, Surveyed CSAT reporting, reporting metrics and attributes and the Teammate changes conversation state trigger, plus the API’s conversation model (all accessed November 2026).
How Intercom ratings work
How the survey is sent
Intercom sends its CSAT survey through Workflows. The usual setup is the “Ask for conversation rating” action right after the Teammate changes conversation state trigger, so the customer is asked when a conversation is closed. You can also add the action to other workflow paths. The default survey is a five-option emoji scale. If you want a different scale, Intercom suggests building your own with reply buttons, but those custom ratings are not included in the built-in CSAT report.
What is stored
Each rating is attached to the conversation. Intercom’s API documents it as an integer from 1 to 5 with an optional remark (the customer’s comment). Two details matter for analysis:
- Once a customer adds a remark, the rating can no longer be changed.
- You cannot create or edit ratings through the API; it is read-only for this data.
How Intercom computes CSAT
Intercom’s reporting definitions count the two highest ratings as positive. The CSAT score is positive ratings divided by all ratings, across teammates, Fin AI Agent and chatbots. The Surveyed CSAT report template shows overall, teammate and Fin scores, CSAT over time, remarks grouped by rating, and the survey request and response rates. Building custom CSAT charts requires the Advanced or Expert plan.
That score answers “how is our support doing?” It does not answer “which customers are about to leave?” For that, you need ratings per customer, joined to billing.
Getting the data out
You have two options.
Dataset export. In Reports, the dataset export includes a “Conversation Ratings” dataset with conversation-level rating data. This is the fastest path for a one-off analysis.
API. Retrieve conversations through the conversations API; each one carries its rating, remark and contacts. Intercom notes that older API versions can return a null rating for ratings visible in the app, so use a recent API version.
For each rating, keep: conversation ID, contact ID and email, rating (1 to 5), remark, rated-at date, and whether a teammate or Fin handled the conversation.
Joining ratings to churn
1. Build a customer table from billing
From Stripe or your billing system, list every customer who was active at some point in your window, with:
- start date;
- cancellation date (or empty if still active);
- whether the cancellation was voluntary or due to a failed payment.
Keep involuntary churn separate. A customer whose card expired did not leave because of a bad support chat. Mixing the two hides the effect you are looking for; see involuntary churn.
2. Match Intercom contacts to customers
Match by email, or by the Stripe ID if you use Intercom’s Stripe app, which syncs stripe_id onto Intercom users. Review the doubtful cases by hand: shared inboxes, aliases, a colleague writing on behalf of the account owner. For B2B products, decide whether a rating from any team member counts for the account.
3. Pick a fixed window
For each customer, take ratings from a defined period (for example the 90 days before a reference date), and check whether they cancelled in the following 90 days. Without a fixed window, long-lived customers simply accumulate more ratings, and you end up measuring tenure.
4. Group and compare
A simple first cut:
| Group | Definition |
|---|---|
| Rated low | At least one rating of 1 or 2 in the window |
| Rated high only | Only ratings of 4 or 5 |
| Rated middle only | Only 3s, or a mix without lows |
| Contacted, no rating | Had conversations but did not answer the survey |
| No contact | No conversation in the window |
Compute the voluntary churn rate in the next period for each group. If you need a refresher on the formula, see how to calculate churn rate.
5. Look at the remarks
The numbers tell you whether there is a link; remarks tell you why. Read every remark attached to a low rating for customers who later cancelled. Group them by theme: billing confusion, a bug, a missing feature, slow answers. That list is more useful than the correlation itself.
Reading the result honestly
A few traps catch almost everyone:
- Small samples. If only a small share of conversations get rated, the “rated low” group may be a few dozen customers. Report the count next to every rate, and do not act on a difference you could get by chance.
- Response bias. People who answer surveys differ from people who do not. The “contacted, no rating” group is often the most interesting one.
- Rating the agent vs rating the product. A 1 can mean “your agent was rude” or “your product cannot do this”. Only the second is likely to drive churn, and only the remark tells you which.
- Fin vs teammates. Ratings on AI-handled conversations and on human-handled ones may behave differently. Compare them separately before merging.
- Correlation is not causation. A low rating can be a symptom of a customer already on their way out. That still makes it a useful signal; it does not mean better support alone would have kept them.
What to do if ratings do predict churn
If low-rated customers churn noticeably more, act on the signal:
- Give low ratings an owner. Intercom’s reports can filter by rating and remark; have someone review the low-rated conversations every day and follow up within a day.
- Follow up on the remark, not the score. Reply to what the customer said.
- Check billing at the same time. A low rating right after a failed payment or a surprise renewal needs a billing fix, not an apology.
- Re-run the analysis each quarter with the same windows, so you can see whether the gap narrows.
If ratings do not predict churn in your data, that is a result too. Look at other support signals instead: conversation volume, topics, or time to first reply. We cover a wider approach in how to find what predicts churn with your own customer data.
Doing this with NoChurn
NoChurn’s Intercom connector is read-only. It reads conversations and ratings, matches them to billing customers by email (doubtful matches are suggested for you to confirm, never merged automatically), and uses them as inputs to the 0–100 churn score and to the churn drivers analysis, which ranks the attributes most correlated with churn (correlation, not causation). It does not change your CSAT setup or send surveys; that stays in Intercom. You can also push your own attributes by API to compare support signals with product usage. See NoChurn + Intercom.