Quick answer
Customer sentiment is how customers feel about a company, product or interaction: positive, negative, neutral or mixed, and how strongly. It is a process signal that changes during a conversation, unlike satisfaction, which is an outcome rating given after it ends.
Formula: Low-score share = ratings of 1 or 2 ÷ all ratings (1–5 scale)
What is customer sentiment?
Customer sentiment is how customers feel about a company, product or interaction: positive, negative, neutral or mixed, and how strongly. It shows up in what people write and say (chat messages, reviews, emails, call transcripts) and in what they do (asking for a person, contacting you again about the same problem, leaving mid-conversation).
The useful distinction is between sentiment and satisfaction. Sentiment is a process signal: it changes during a conversation, sometimes from message to message, and tells you something is going wrong while there is still time to fix it. Satisfaction is an outcome signal: a rating given after the interaction is over, which tells you how it ended. A customer can start a chat angry, get a fast fix and rate it 5; another can stay polite throughout and still rate it 2 because the problem was never solved.
That difference decides how to measure it. Before reaching for a sentiment model, use the structured signals you already have: ratings, requests to speak to a human, the same question coming back, long waits. A text-sentiment model adds value only when there are more conversations than people can read.
The term "customer sentiment analysis" usually means the automated part: software that labels text as positive, negative or neutral. That is one way to measure sentiment, not the definition of it.
Sentiment vs satisfaction vs loyalty
| Customer sentiment | Customer satisfaction (CSAT) | Loyalty (e.g. NPS) | |
|---|---|---|---|
| What it captures | Emotion and its direction, during an interaction or across many | How one interaction or purchase went | Willingness to stay and recommend over time |
| Signal type | Process signal | Outcome signal | Relationship signal |
| How you get it | Text, tone and behaviour, read by people or models | A rating question after the interaction, often 1–5 | A survey question such as "how likely are you to recommend us", 0–10 |
| When it is available | While the conversation is happening | After it ends, only from people who answer | Periodically, from a sample |
| Typical use | Spot conversations going wrong, find recurring frustrations | Grade service quality per agent, team, period | Track the overall relationship, compare segments |
The three overlap but are not interchangeable. Satisfaction ratings come only from customers who choose to answer, so they miss the silent unhappy ones. Loyalty scores move slowly and do not tell you which conversation went wrong. Sentiment fills the gap between them, if you can measure it reliably. For calculating a satisfaction score, see the CSAT calculator; for how sentiment and satisfaction fit into the wider picture, see customer experience.
How to measure customer sentiment
Start with signals that are already recorded as numbers or events. They are cheap, they are not ambiguous, and they cover the customers who never fill in a rating.
- Ratings after the conversation. Low scores (1 or 2 on a 1–5 scale) are the most direct negative signal. Track the share of low scores, not just the average.
- Requests for a human. A visitor who types "agent" or "talk to a person" while a bot or knowledge base is answering is telling you the automated answer did not land.
- Repeat contact. The same customer back about the same issue within days, or a resolved conversation reopened, means the first answer did not resolve it. This is why resolution rate and sentiment move together.
- Waiting. Long first response times are one of the most common causes of frustration in live chat. Wait time is measured to the second, so it is easy to cross with ratings.
- Leaving. A visitor who closes the chat before an agent replies, or leaves an offline message because nobody was available, is a weak but frequent negative signal.
- What people write. Words like "still", "again", "third time", "ridiculous" and all-caps messages. Reading a sample of conversations by hand is the simplest form of sentiment analysis.
How sentiment analysis works
Automated sentiment analysis assigns each piece of text a label or a score. Two families of methods are common:
- Lexicon and rule based. Each word has a polarity score from a dictionary ("great" positive, "broken" negative), with rules for negation ("not great"), intensifiers ("very") and punctuation. Transparent and fast, but blind to context.
- Machine-learning classifiers. A model trained on labelled examples learns which patterns go with positive or negative text. Pang, Lee and Vaithyanathan's 2002 study on movie reviews found that standard classifiers did not perform as well on sentiment as on ordinary topic classification, because sentiment depends on context, contrast and sarcasm. Modern language models do better, but the same difficulties remain.
Outputs are usually a score plus a strength. Google Cloud's Natural Language API, for example, returns a score from −1.0 (negative) to 1.0 (positive) and a separate magnitude; a score near 0 can mean either neutral text or a mix of strong positive and negative, which only the magnitude separates.
The practical limits matter more than the method. Support conversations are short, full of product names and often multilingual; a model trained on reviews may read "this kills the bug" as negative. And a sentiment label tells you that a conversation went badly, not why. That is why models pay off mainly at volume, when no one can read the conversations, and why their output should be checked against the structured signals above.
Worked example: what low ratings have in common
Sample data: a support team handles 400 chat conversations in a month. Customers rate 260 of them (65.0%); 52 ratings are 1 or 2, so the low-score share is 52 ÷ 260 = 20.0%. The question is what the low scores share, using a signal already recorded: whether the customer waited more than 5 minutes for the first agent reply.
| Rated conversations | Waited > 5 min | Waited ≤ 5 min | Total |
|---|---|---|---|
| Low score (1–2) | 31 | 21 | 52 |
| Not low (3–5) | 25 | 183 | 208 |
| Total | 56 | 204 | 260 |
- Share of low-score conversations that involved a long wait: 31 ÷ 52 = 59.6%.
- Share of other rated conversations with a long wait: 25 ÷ 208 = 12.0%.
- Low-score rate when the wait was over 5 minutes: 31 ÷ 56 = 55.4%; when it was 5 minutes or less: 21 ÷ 204 = 10.3%.
More than half of the conversations with a long wait ended in a low score, against about one in ten of the rest. Without any sentiment model, the team has found its main source of negative sentiment and a lever to act on: staffing at peak hours, an away message that sets expectations, or a knowledge-base answer for the most common question. The remaining 21 low scores with a short wait are few enough to read one by one; that reading, not a model, is what explains them.
This is the core point in numbers. With 52 negative ratings a month, a person can review every one. A text-sentiment model starts to earn its cost when there are thousands of conversations a week, many of them unrated, and even then its labels are best checked against signals like wait time and repeat contact.
Common mistakes
- Treating the average rating as sentiment. A 4.2 average hides whether low scores are growing. Track the low-score share and read those conversations.
- Ignoring the people who do not rate. In the example, 140 of 400 conversations (35.0%) have no rating. Use behavioural signals for them.
- Buying a sentiment model before reading conversations. At small volume, reading is faster and explains causes; the model only labels.
- Trusting labels without checking. Sample model output against human judgement, especially for short, technical or translated messages.
- Measuring sentiment but not acting on it. Every negative pattern should map to an owner and a change, such as staffing, a canned reply or a product fix. Phrasing helps too; see empathy statements.
- Comparing sentiment across channels as if equal. People write differently in chat, email and reviews; compare each channel with itself over time.
Reading customer sentiment in TapCub
TapCub does not do sentiment analysis: it does not score or label the text of conversations. What it gives you are the structured signals the worked example uses, which is where measuring sentiment should start.
- Watch the outcome signal. Live chat › Reports shows satisfaction: visitors can rate a conversation from 1 to 5 within 24 hours after it is resolved, one score per conversation, averaged to one decimal. Next to it are conversations, resolution rate (resolved ÷ all conversations) and average first response, with the median shown as well.
- Cross ratings with waiting. Average first response counts seconds from the start until an agent's first reply and leaves out AI replies, so you can compare periods or agents where first response rose with periods where satisfaction fell. The 5-minute split in the example comes from reviewing the low-rated conversations themselves, not from a ready-made report.
- Read the low-scored conversations. In Live chat › Inbox, the Resolved list shows each conversation's satisfaction score, and the side panel shows the visitor's earlier conversations, so repeat contact about the same problem is visible without a separate tool.
- Count the visitors nobody answered. Offline messages are counted separately in the reports and are not part of resolved, so a rise in them shows demand that met no agent.
- See who asked for a person. With AI replies on, the AI stops when the visitor asks for a human and the conversation goes to an agent; those handovers are worth reading for automated answers that missed. Tips to raise the scores themselves are in how to improve CSAT.
If you need automated sentiment labels, that takes a separate tool; TapCub's role is the ratings, timings and conversation history around them.
See also: CSAT meaning: what it is, how to calculate it and sample size.
See also: What is user feedback? Types, how to analyze it and an example.
Sources
- Pang, Lee and Vaithyanathan, Thumbs up? Sentiment Classification using Machine Learning Techniques (EMNLP 2002).
- Google Cloud, Natural Language API basics: interpreting sentiment analysis values.
- Frederick F. Reichheld, The One Number You Need to Grow, Harvard Business Review, 2003.
Was this explanation clear?
We use your answers to fix and expand this page.
Couldn’t send. Please try again in a minute.
Thanks for the feedback. We’ll use it to improve this page.You’ve already sent feedback on this page. Thanks.

Growth Strategist
Explores acquisition channels, activation, and conversion experiments, turning data insights into actionable growth hypotheses and evaluation methods.