NPS Text Analysis: Turn Open-Ended Comments into Decisions

NPS text analysis is the process of coding the written comments that follow the 0–10 recommendation question, grouping them into themes and sentiment, and linking those themes back to Promoter, Passive and Detractor scores. The score shows how many customers would recommend you. The comments show why. This guide covers the workflow, the choice between manual and automated coding, the checks that keep findings accurate, and how to turn themes into actions.

TL;DR

• Keep every comment attached to its score and respondent ID, then code each one into themes using a code frame refined as you read.
• Compare themes separately for Promoters, Passives and Detractors. The same topic can describe very different experiences.
• Use manual coding for small samples and codebook design. Use text analytics when volume makes reading every response impractical. Most teams need both.
• State the denominator for every percentage, and rank themes by both volume and relationship to NPS, not by mention count alone.
• Redact identifying details in wide reports and share individual records only with authorized people.

Table of Contents

• What NPS text analysis adds to the score
• A step-by-step workflow for NPS verbatim analysis
• Reading themes and sentiment by score group
• Manual coding, text analytics or both
• Checking that findings are accurate and representative
• From themes to improvements
• Privacy and data handling
• Where Lexic fits
• Perspective
• FAQ
• Sources

What NPS text analysis adds to the score

The Net Promoter method asks one question on an 11-point scale. According to Bain’s description of the method, respondents scoring 9–10 are Promoters, 7–8 are Passives and 0–6 are Detractors. NPS is the percentage of Promoters minus the percentage of Detractors. The same source describes a follow-up that asks respondents to explain their score in their own words.

That follow-up produces verbatims: unstructured text, one response per customer. A score of +32 says nothing about what to fix. A verbatim such as “setup took three calls and nobody followed up” does. Text analysis converts thousands of such sentences into a ranked, measurable picture.

The analysis answers four questions:

1. Which topics do customers raise?
2. Is the tone about each topic positive, negative or neutral?
3. Which topics separate Promoters from Detractors?
4. Which topics are large enough, and tied closely enough to the score, to justify action?

A step-by-step workflow for NPS verbatim analysis

The workflow below follows the practical sequence set out by MeasuringU’s guide to coding verbatim comments, adapted for NPS data.

1. Export comments with identifiers and scores. Each row needs a unique respondent ID, the 0–10 score, the comment and any permitted metadata, such as segment, region, product or tenure. Without the ID you cannot audit a code later.
2. Read an orientation sample. Review a random set of comments before building anything. Reading 50 to 100 is a common starting point, and it shows how long, noisy or off-topic the responses are.
3. Clean the data. Remove blanks, single characters, test entries and obviously unusable responses. Keep a count of what you removed, because it affects your denominator.
4. Split multi-idea comments. “Great support but the invoice portal is confusing” holds two ideas. Break each response into atomic units so each unit carries one theme.
5. Build a code frame. A code frame is the set of codes for one question. Draft themes from the sample, name them in plain language and write a one-line definition for each.
6. Let themes emerge, then refine. Avoid starting with a fixed conclusion. Make sure the frame includes contrasting categories, merge redundant codes, and recode earlier responses when a new category appears.
7. Assign codes. Give each unit one or more themes. Decide in advance whether themes are single-label or multi-label, and record the choice.
8. Quantify. Count how many respondents mention each theme and express it as a share of an explicit base. Add uncertainty estimates such as confidence intervals where samples are small.
9. Link themes to NPS. Calculate theme incidence by score group and, where the data allows, the NPS of respondents who mention each theme versus those who do not.
10. Keep representative quotes. A few verbatims per theme, chosen to represent the typical case rather than the most dramatic, make findings credible to stakeholders.

Reading themes and sentiment by score group

A theme list is only the first layer. The more useful question is how each theme behaves across Promoters, Passives and Detractors.

The approach described in Quirk’s article on structured verbatim analysis is to split each verbatim into text units, code each unit positive, negative or neutral, compare tone across the score groups, and then build themes iteratively. Its example found that Detractors still made positive and neutral comments. An entire response, or an entire score band, should not be treated as uniformly negative.

In practice this gives three useful views:

• Detractor themes: the pain points most associated with low scores. Look for themes where the negative share is high and the volume is meaningful.
• Promoter themes: what customers value enough to recommend. These themes protect revenue and belong in messaging and onboarding.
• Passive themes: often the most overlooked group. Their comments are frequently lukewarm or conditional, and they show what stops a satisfied customer from becoming an advocate.

The same topic can mean different things in different groups. “Pricing” among Promoters may be praise for value. Among Detractors it may be a complaint about a renewal increase. Counting the word “pricing” without sentiment and score context hides that difference. Ipsos recommends beginning with the relationships among topics, often shown on visual maps, to support broader theme categories and to give context for what customers said.

Word clouds and raw topic frequency are a weak substitute. A frequently used word does not show whether the underlying experience was positive or negative, or what the customer meant by it.

Manual coding, text analytics or both

The right method depends on volume, cadence and how much nuance the comments carry. Ipsos defines text analytics as a set of computer-processing techniques that identify and quantify themes and sentiment in unstructured customer text, using semantic processing, language parsing and statistical analysis.

Automated tools also differ in depth. The simplest count single words. The next level counts multi-word concepts, then concepts qualified by sentiment. The most capable group differently worded expressions with similar meaning into a taxonomy, so “waited forever” and “slow response” land in the same theme.

A sensible sequence for most teams:

1. Code a first sample by hand to learn the language of your customers and draft the codebook.
2. Apply automated classification to the full dataset using that codebook.
3. Audit a random sample of machine-coded comments each cycle and correct the codebook where errors cluster.
4. Keep humans responsible for deciding which themes matter.

Checking that findings are accurate and representative

A clean chart can still mislead. Build these checks into every report.

Keep traceability. Retain participant IDs so any theme count can be traced to its source comments. Review a random sample and example verbatims for each theme, as MeasuringU advises.

Merge redundant themes. Two codes for the same idea split its true volume and make both look minor.

Check coding reliability. Have a second coder, or a second pass, code a sample and compare agreement. Low agreement means the definitions are unclear.

State the denominator. NPS usually covers all valid ratings. A theme percentage may cover only respondents who left a usable comment. These bases differ, and a report should say which one each figure uses. If themes are multi-label, shares can add up to more than 100%, and the report should say so.

Respect small bases. Do not rank subgroups with few comments as though small differences were stable. Show the count next to the percentage, and add confidence intervals where possible.

Watch for comment bias. Customers who write comments are not always a mirror of those who do not. Compare the score distribution of commenters with that of all respondents before generalizing.

From themes to improvements

The goal is a decision, not a dashboard. Two ideas help.

Prioritize by impact and reach. Ipsos describes text-based impact analysis as relating themes to a KPI such as NPS and combining that relationship with theme volume. The result flags issues that both move the KPI materially and touch many customers. A rare theme with a severe effect and a common theme with a weak effect need different responses.

Do not equate frequency with influence. The topic customers mention most is not always the one most associated with the score. Pair comment analysis with statistical driver analysis, and ask which themes explain the gap between Promoters and Detractors.

Two operating loops keep the work practical:

• Individual follow-up. Review single responses, especially from Detractors, and resolve the customer-level problem quickly.
• Systemic improvement. Track recurring themes over time and assign owners in product, support, operations or billing to fix root causes.

Comments also gain meaning when set beside operational data. If support tickets show a spike in a category, NPS comments often explain why. Build a simple review table for each priority theme:

Privacy and data handling

NPS comments often contain names, account details or sensitive remarks that customers did not expect to circulate. Good practice is to:

• Keep response IDs, scores, original comments and relevant permitted context together in one controlled dataset.
• Report the comment denominator separately from valid rating responses.
• Redact identifying details in wider reports and presentations.
• Share individual records only with people authorized to act on them.
• Confirm with your legal or privacy team which rules apply to your customers’ data and any outside tools that process it.

Where Lexic fits

Lexic works with the same raw material, customer conversations, at a broader scale. Lexic Pulse analyzes 100% of customer interactions across calls, chats, emails and tickets, and generates sentiment, churn and opportunity insights, alongside mood, NPS and urgency scoring, for teams across CX, Product, Sales, QA and Operations. That places survey verbatims in context with what customers say in everyday channels.

Organizations that deploy conversational AI agents face a related question: how those agents behave in real conversations. Lexic Compass independently audits agents from any vendor across customer experience, security, technical reliability and regulatory compliance. Lexic’s audits found that 85% of the agents assessed did not escalate issues to a human representative appropriately. Audit options, including Flash Preview, Audit Sprint and Enterprise Continuous Trust, are described on the audit engagement page.

Perspective

Survey comments are one slice of customer language. They capture what a customer chooses to write after a rating, often days after the interaction. The conversations themselves, the calls, chats and tickets, hold the unprompted version of the same story.

We believe NPS text analysis works best when it is connected to that wider record. A theme found in survey comments can be checked against what customers said in the moment, and a pattern spotted in conversations can be tested against the score. This is the logic behind our approach to conversational intelligence.

The same discipline applies to AI agents that now handle a growing share of customer conversations. Their behavior needs independent evidence, gathered by interacting with them as a real user would, rather than vendor-reported metrics alone. Our work on Lexic Compass is built on that principle, and it is why we pair customer insight with AI oversight.

Sergio Llorens

FAQ

How many NPS comments do I need before analysis is worthwhile?

There is no fixed minimum. Reading 50 to 100 comments is a common way to identify recurring themes and draft a code frame. Quantified shares become more stable as the base grows, so show counts and confidence intervals alongside percentages, and avoid ranking small subgroups.

What is the difference between NPS verbatim analysis and sentiment analysis?

Sentiment analysis labels the tone of text as positive, negative or neutral. Verbatim analysis is broader: it also identifies what topic each comment concerns and links topics to score groups. Used together, they show which themes are praised, which are criticized and how that differs among Promoters, Passives and Detractors.

Should I analyze comments manually or with software?

Start manually for small datasets and to build your codebook, because it surfaces nuance. Move to automated text analytics when volume or survey frequency makes reading every response impractical. A combined approach is common: automation screens, clusters and quantifies, and people validate themes and set priorities.

Why do theme percentages sometimes add up to more than 100%?

When one comment can carry several themes, each theme is counted separately, so totals exceed 100%. This is normal with multi-label coding. State in the report that themes are multi-label and name the base, for example all respondents who left a usable comment.

How do I decide which theme to fix first?

Combine two measures: how strongly a theme relates to NPS and how many customers mention it. A theme with both material impact and broad reach is a strong candidate. Check it against operational data, such as support tickets, and assign an owner and review date.

Sources

• Bain & Company, The Ultimate Question 2.0 (excerpt)
• MeasuringU, How to Code & Analyze Verbatim Comments
• Quirk’s, A structured approach for qualitative verbatim analysis
• Ipsos, A Guide to Text Analytics