A 95% CSAT score looks definitive. It fits neatly on a dashboard, reassures leadership, and gives the support team a number to rally around.
But before celebrating, ask a second question:
Your surveys only hear from a fraction of customers. Supportman predicts CSAT for every unrated Intercom conversation, so you see satisfaction across 100% of your support volume.
What percentage of customers actually rated the conversation?
If 20 customers respond to 100 survey requests and 19 choose a positive rating, the reported CSAT is 95%. That is an accurate summary of the 20 responses. It says nothing directly about the other 80 conversations.
This is the blind spot inside survey-only CSAT.
CSAT score and response rate measure different things
CSAT is positive ratings divided by all submitted ratings. Response rate is submitted ratings divided by eligible survey requests. The first tells you how respondents felt; the second tells you how much of your customer base those respondents represent.
Support teams need both numbers, side by side. A 95% CSAT on a 20% response rate is a real result about a small, self-selected group—useful feedback, but never a census of every customer experience. For the full three-metric breakdown, including how coverage differs from response rate, see What Is 100% CSAT Coverage?
The silent majority may be different
The deeper problem is that survey response is voluntary—and volunteering is not random. Customers who rate may differ systematically from customers who do not.
Every support team runs versions of this selection process without noticing it. An agent who just rescued a frustrated account closes with "mind leaving a rating?"—so happy customers get invited into the sample at a higher rate than everyone else. In-chat rating prompts catch customers in the moment, while follow-up emails land in an inbox hours later, so a quiet shift in channel mix changes who gets counted. The customer angry enough to churn often skips the survey entirely; they are not looking to give feedback, they are looking for the cancel button. And enterprise customers who route everything through a CSM or a shared Slack channel may never see a survey at all. None of these customers appear in your CSAT—but all of them appear in your renewal numbers.
A study examining more than 170,000 Samsung customer-service chat sessions found that only 16.2% contained a submitted rating. Rated sessions were overwhelmingly positive, while the researchers’ models estimated that unrated sessions would have received lower scores on average. The authors described this as a positivity bias in customer-satisfaction ratings. Read the 170,000-chat Samsung study.
That result does not prove that every company’s unrated conversations are less positive. It does show why leaders should not assume that nonrespondents would have answered like respondents.
What the bias costs: a worked example
Selection bias is invisible on a dashboard, so put numbers on it.
Imagine 1,000 eligible conversations in a month. 200 customers rate, and 190 of those ratings are positive—a 95% CSAT. Now suppose the 800 silent conversations, had they been rated, would have come in at 80% positive: 640 positive experiences. Across all 1,000 customers, that is 830 positive out of 1,000—an all-customer satisfaction of roughly 83%, not 95%. A twelve-point gap, and nothing on the dashboard hints at it. The reported score never moved, no alert fired, and the team celebrated a number that describes one-fifth of the queue.
The 80% figure is an assumption—and that is exactly the point. At a 20% response rate, the difference between a true 95% and a true 83% rests entirely on an assumption you cannot check from survey data alone. The Samsung study suggests the assumption usually errs in the flattering direction.
Why improving response rate does not solve everything
Make surveys easy to complete: ask in the channel where the conversation happened, immediately after it closes, with a one-click first question. Those changes earn more first-party feedback and more labelled examples for checking any predicted score against reality.
But even a strong response rate leaves most conversations unrated, and the customers who do respond are still volunteers—a bigger sample of a biased sample. Customers do not owe companies a survey response, and chasing them with reminders creates more friction than insight. The realistic goal is to preserve direct feedback while building visibility into the conversations that stay silent.
What survey-only CSAT can tell you
Surveyed CSAT is still valuable. It is the customer’s direct expression of satisfaction, and nothing inferred by a model should be passed off as equivalent.
Use surveyed CSAT to:
- Identify strong praise and urgent dissatisfaction, in the customer’s own words
- Track changes among respondents over time
- Calibrate automated quality and satisfaction models
Just be precise about the population measured.
Instead of reporting:
Our customers are 95% satisfied.
Report:
Among 200 customer ratings from 1,000 eligible conversations, 95% were positive. Survey response rate was 20%.
That one sentence is more honest and more useful.
Add a signal for the unrated conversations
Predicted CSAT uses the content and context of a support conversation to estimate the satisfaction rating a customer may have given. It creates a separate, clearly labelled operational signal for the conversations the survey did not measure—it never turns an inference into a survey response.
A trustworthy model follows three rules:
- A real customer rating always takes precedence.
- Predicted ratings are clearly labelled as AI-inferred.
- Surveyed and predicted results remain separable in reporting.
This produces a two-layer view:
| Signal | Purpose |
|---|---|
| Surveyed CSAT | Direct customer feedback |
| Predicted CSAT | Coverage for unrated conversations |
Applied to the worked example above, the 200 real ratings stay untouched and the 800 silent conversations each get a labelled estimate—which is precisely where a twelve-point blind spot stops being invisible. The full model, including eligibility rules and how to report the results, is covered in What Is 100% CSAT Coverage?
Measure the customers who answer—and learn from those who do not
Surveyed CSAT deserves a permanent place in customer-support reporting. It captures direct customer judgment in a way no inferred metric can replace—and it measures only the customers who respond.
Here is the Monday-morning version. Put your survey response rate on the dashboard next to your CSAT score this week. If fewer than one in five eligible conversations gets rated, read the score as a poll of your happiest and angriest customers, not a measure of all of them. And if you add a predicted signal, watch the conversations where the prediction disagrees with a real customer rating—that disagreement is where you learn what your surveys are missing.
Supportman’s predicted CSAT adds a clearly labelled satisfaction signal to every unrated eligible Intercom conversation, so the silent majority stops being a blank space in your reporting.
See how Supportman scores the conversations your surveys miss →
Frequently asked questions
Does a low response rate make CSAT useless?
No. The responses are still real customer feedback. A low response rate limits how confidently the result can represent all customers and increases the importance of checking for selection bias.
Is predicted CSAT the same as surveyed CSAT?
No. Surveyed CSAT comes directly from the customer. Predicted CSAT is an AI inference and should always be labelled accordingly.
Should the two scores be combined?
Keep them separate in primary reporting. A combined operational measure may be useful if its calculation and source composition are clearly disclosed.