Why this is hard to get right
Picture this: Your Q2 NPS results just landed in your inbox. Overall score: 34. That's down from 42 last quarter. Leadership wants to know why — and they want it in Thursday's business review, two days away.
You open the raw export. You've got 412 rows: numeric scores, customer plan tier, account tenure, and 280-odd open-ended comments ranging from three words to three paragraphs. You know there's a story in there. You just don't have 12 hours to find it manually.
So you try the obvious move — you paste a chunk of the data into ChatGPT and ask it to "analyze my NPS results." The output comes back looking polished. It summarizes that customers value ease of use but have concerns about support response times. It recommends "improving the onboarding experience" and "investing in customer support."
The problem: that output is basically true of every SaaS company that has ever run an NPS survey. It doesn't tell you whether it's your Enterprise accounts or your SMB accounts driving the drop. It doesn't tell you whether the support complaints cluster in the 0-12 month tenure band — which would implicate onboarding — or the 3-year band — which would implicate a product regression. It doesn't give your VP of Customer Success anything actionable to bring to the team.
This is the gap that vague prompting creates. The AI isn't wrong. It's just answering a different, much simpler question than the one you actually need answered.
The professionals who get real value from AI analysis don't dump their data and hope for the best. They architect the prompt: they define the segments, describe the data structure, specify the comparison they need, and constrain the output format so the result is usable — not just readable.
That's exactly the kind of prompt this page helps you build.
Common mistakes to avoid
Skipping Data Structure Description
If you don't tell the AI what columns or fields your dataset contains, it will invent a structure that fits a generic NPS dataset — not yours. Always describe your actual fields, segment labels, and response volume before asking for analysis.
Asking for 'Insights' Without Defining Segments
Prompting for general insights produces general answers. NPS analysis only becomes useful when you compare specific groups — plan tiers, tenure bands, geographies. Name the comparisons you care about explicitly or the AI won't make them.
Requesting Actions Without Specifying Ownership
AI-generated NPS recommendations default to vague advice like 'improve onboarding.' If you specify which team owns the output — product, support, CS — the AI ties recommendations to functions, making them immediately assignable.
Ignoring Verbatim Comments in the Prompt
Many users only mention the numeric scores and forget to instruct the AI to analyze open-ended comments. Verbatims are where drivers live. Explicitly ask the AI to group themes from comments by promoter and detractor status.
Not Constraining Output Length or Format
Without format guidance, NPS analysis outputs often run 1,500+ words with no clear hierarchy. Specify an executive summary length, section labels, and whether you want tables — otherwise you'll spend as much time editing as you saved on analysis.
The transformation
Analyze my NPS data and tell me what's affecting the score and what I should do about it.
**Act as a customer experience analyst** with expertise in NPS driver analysis. I will provide you with NPS survey results from our B2B SaaS platform (Q2 2025, n=412 responses). The dataset includes: numeric scores (0-10), customer segment (SMB vs. Enterprise), tenure band (0-12 months, 1-3 years, 3+ years), and open-ended verbatim comments. **Your task:** 1. Identify the top 3-5 drivers of promoter scores (9-10) vs. detractor scores (0-6) 2. Surface recurring themes in verbatim comments, grouped by sentiment 3. Highlight any segment-specific patterns (SMB vs. Enterprise, or by tenure) 4. Flag statistically notable gaps between segments 5. Recommend 3 prioritized actions, each tied directly to a specific driver **Format:** Executive summary (3-4 sentences), then a structured breakdown with labeled sections. Use a table for driver comparison where helpful. Keep total output under 600 words.
Why this works
Specificity
Naming the exact dataset (n=412, Q2 2025, B2B SaaS) anchors the AI in your actual context. It stops the model from applying generic NPS benchmarks that don't reflect your industry, customer profile, or business model.
Segmentation
Explicitly naming SMB vs. Enterprise and tenure bands forces the AI to run the comparisons that matter for real decisions. Without named segments, the AI averages across your customer base and obscures the patterns you most need to see.
Structure
The numbered task list breaks a complex, multi-part analysis into sequential steps. This dramatically reduces the chance of the AI collapsing multiple distinct analytical tasks — theme extraction, driver ranking, recommendations — into a single undifferentiated response.
Constraints
The 600-word cap and format requirements (executive summary, labeled sections, driver table) ensure the output is presentation-ready. Constraining format isn't limiting — it's the difference between a deliverable and a draft.
Role Priming
Opening with 'act as a customer experience analyst' activates domain-specific reasoning patterns. The AI applies CX-specific logic — like weighting detractor risk differently than passive scores — rather than defaulting to general data analysis behavior.
The framework behind the prompt
NPS driver analysis draws on two well-established research traditions: key driver analysis (KDA) and text analytics.
Key Driver Analysis, rooted in regression-based market research, identifies which variables have the strongest statistical relationship with an outcome — in this case, the likelihood-to-recommend score. In traditional research, this requires statistical modeling. With AI prompting, you replicate the core logic by asking the model to surface patterns that correlate with score extremes rather than the average.
The qualitative side of NPS analysis connects to thematic coding frameworks used in qualitative research. Systematic thematic analysis — as formalized by Braun and Clarke — involves identifying, grouping, and naming recurring patterns across a text corpus. When you instruct an AI to group verbatim comments by sentiment and label recurring themes, you're applying this framework at speed.
The SERVQUAL model (reliability, assurance, tangibles, empathy, responsiveness) also informs best-practice NPS analysis. Structuring verbatim themes around service quality dimensions helps teams connect customer language to operational categories they already manage.
Effective NPS driver analysis combines all three: statistical thinking (what correlates with score?), thematic coding (what do customers actually say?), and service quality framing (what does this mean for how we operate?). A well-constructed AI prompt sequences these three moves deliberately.
Prompt variations
Act as a qualitative research analyst specializing in customer feedback.
I'm sharing 150 open-ended NPS comments from our e-commerce platform (collected post-purchase, Q1 2025). Scores are not included — focus entirely on the verbatims.
Your task:
- Identify and label the top 5 recurring themes across all comments
- Classify each theme as primarily positive, negative, or mixed
- Pull 2-3 representative verbatim quotes per theme
- Rank themes by frequency of mention
- Suggest one concrete response for the top negative theme
Format: Theme table with columns for theme name, sentiment, frequency, and sample quote. Follow with a brief recommendation section.
Act as a product analytics specialist with experience connecting NPS data to product decisions.
I have NPS results from our project management SaaS (n=280). Each response is tagged with the primary feature area the customer most recently used (reporting, task management, integrations, or mobile app).
Your task:
- Compare average NPS scores across the 4 feature areas
- Identify which feature area generates the most detractors (scores 0-6)
- Surface the top 2 complaint themes per feature area from verbatim comments
- Recommend 2 product changes per high-detractor feature area, ranked by estimated impact
Output: Feature comparison table, then a prioritized recommendation list. Keep under 500 words. Assume the audience is a product leadership team reviewing roadmap priorities.
Act as a customer insights analyst preparing a quarterly trend report for senior leadership.
I will provide NPS data for three consecutive quarters (Q4 2024, Q1 2025, Q2 2025) for our HR software platform. Each quarter includes overall score, segment scores (Enterprise vs. SMB), and top verbatim themes already summarized.
Your task:
- Identify score trends over the three periods overall and by segment
- Flag any quarter-over-quarter shifts greater than 5 points and explain likely causes
- Identify themes that are newly emerging vs. persistently recurring
- Provide a 3-sentence executive narrative summarizing the overall trend
- Recommend one priority focus area for Q3 based on trend direction
Format: Trend summary table, narrative paragraph, then prioritized recommendation. Under 500 words total.
When to use this prompt
Customer Success Leaders
Identify which customer segments are at churn risk based on detractor patterns and tenure-band signals, then build targeted intervention playbooks for at-risk accounts.
Product Managers
Extract feature-specific complaints and praise from NPS verbatims to inform roadmap prioritization, making the case for investment with direct voice-of-customer evidence.
CX and Insights Teams
Synthesize quarterly NPS cycles into a structured driver report that executives can review in under 5 minutes, replacing ad hoc slide decks built from gut instinct.
Marketing Teams
Identify promoter language and recurring praise themes to fuel testimonial programs, case study development, and messaging that reflects what customers actually value.
Operations and Strategy Teams
Track NPS driver shifts across multiple quarters to detect emerging satisfaction trends before they become retention problems, supporting proactive resource allocation.
Pro tips
- 1
Specify your exact sample size and collection method (in-app survey, email, post-support ticket) so the AI can flag potential bias in your data before drawing conclusions.
- 2
Name the segments you care about — don't just say 'different customer types.' Include plan tier, industry vertical, or tenure band so the AI surfaces comparisons that map to real business decisions.
- 3
Define what 'action' means in your context. If you need recommendations tied to a specific team (product, support, onboarding), say so — otherwise the AI defaults to broad, org-agnostic suggestions.
- 4
Include your current NPS score and trend direction (e.g., dropped from 42 to 34 quarter-over-quarter) so the AI frames its analysis around change, not just the static snapshot.
The quality of your prompt output depends heavily on how well you've structured your data before you start. Here's a practical preparation checklist:
1. Standardize your score field. Make sure numeric scores are clean integers (0-10), not ranges or letter grades. Flag and exclude incomplete responses.
2. Label your segments clearly. If your data has a customer tier field, make sure the values are consistent — 'SMB', 'smb', and 'Small Business' are three different values to an AI parsing text.
3. Summarize before you paste. For datasets over 100 rows, don't paste everything. Instead, calculate: total n, score distribution (% promoters, passives, detractors), average score per segment, and the top 15-20 verbatim comments by representativeness.
4. Flag known context. If you changed your survey delivery method, ran a promotion, or experienced a major product outage during the survey period, mention it in the prompt. The AI can't account for confounding factors it doesn't know about.
5. Decide your output destination. Is this going into a slide deck, a Slack summary, a long-form report? Tell the AI — format requirements should match how the output will actually be used.
Basic NPS driver analysis identifies what's driving scores. Layered driver analysis goes further — it asks why those drivers matter differently to different customers.
Layer 1 — Frequency: Which themes appear most often? This is what most NPS analysis stops at.
Layer 2 — Correlation: Which themes correlate most strongly with score extremes (0-2 vs. 9-10)? A theme mentioned frequently by passives (7-8) may be less actionable than a theme mentioned rarely but exclusively by detractors.
Layer 3 — Segment interaction: Does the 'support response time' theme cluster in a specific tenure band or plan tier? Segment-level correlation reveals whether a problem is systemic or targeted.
Layer 4 — Trend over time: Is this driver new this quarter, or has it appeared in previous cycles? Recurring drivers that haven't been addressed become urgency signals.
To prompt for layered analysis, add a step explicitly asking the AI to distinguish between high-frequency/low-correlation themes and low-frequency/high-correlation themes. This single addition often produces the most actionable output in the entire analysis.
The analysis is only useful if it drives action. Here's how to structure your prompt's recommendation section to get output that's immediately usable:
Tie each recommendation to a named driver. Avoid generic recommendations. Ask the AI to format each action as: Driver identified > Recommended action > Team owner > Success metric.
Prioritize by impact vs. effort. Instruct the AI to rate each recommendation on estimated customer impact (High/Medium/Low) and implementation effort (High/Medium/Low). This gives you a natural 2x2 for prioritization without building it manually.
Request timeline framing. Ask the AI to classify recommendations as quick wins (0-30 days), medium-term (30-90 days), or strategic (90+ days). This makes the output directly usable in a planning conversation.
Sample prompt addition:
For each recommendation, format as: [Driver] → [Action] → [Owner: product/support/CS/marketing] → [Metric to track] → [Timeline: quick win / medium-term / strategic]
This structure turns a 600-word analysis into a ready-made action tracker your team can copy into a project management tool immediately.
When not to use this prompt
This prompt pattern isn't the right fit when your NPS sample is fewer than 30 responses — the AI may identify "patterns" that are statistical noise, and presenting them as findings creates false confidence in leadership reviews.
It's also not appropriate as a substitute for quantitative significance testing. If your executive team needs p-values or confidence intervals to make budget decisions, you need a statistician or analytics tool, not an AI-generated driver analysis.
Finally, avoid using this prompt when your data contains PII (personally identifiable information). Scrub customer names and account identifiers before pasting any data into an AI tool.
Troubleshooting
The AI produces generic NPS advice that doesn't reflect my actual data
This almost always means the data description in your prompt is too thin. Add: your exact sample size, score distribution (e.g., '38% promoters, 41% passives, 21% detractors'), the specific segment labels in your dataset, and 3-5 representative verbatim quotes. The more concrete the data description, the more specific the output.
Recommendations don't map to any specific team or action
Add a constraint to the prompt: 'Each recommendation must name a specific team owner (product, customer success, support, or marketing) and include one measurable success metric.' Without this instruction, AI recommendations default to organizational-level generalities that no single team can act on.
The output mixes promoter and detractor themes without clear separation
Restructure your task list to explicitly separate the two groups. Add: 'Analyze promoter comments (scores 9-10) separately from detractor comments (scores 0-6). Do not mix themes from both groups in a single section.' Numbered steps work better than paragraph instructions for maintaining this separation.
How to measure success
A strong AI output for this prompt will include at minimum: a clear promoter/detractor distinction (not a blended summary), at least 3 named themes per group with supporting verbatim evidence, segment-level differences called out explicitly (not buried in a footnote), and recommendations that name a specific team or function as owner.
Check that the executive summary could stand alone — a reader who only reads those 3-4 sentences should understand the score direction, the primary driver, and the recommended priority. If the summary is generic or score-agnostic, the prompt needs more data context before you re-run it.
Now try it on something of your own
Reading about the framework is one thing. Watching it sharpen your own prompt is another — takes 90 seconds, no signup.
an NPS driver analysis report for leadership
Try one of these
Frequently asked questions
Yes — just remove the verbatim analysis steps from the task list and focus the prompt on score distribution, segment comparison, and trend analysis. Numeric-only NPS analysis is still valuable for identifying which segments are struggling, even without qualitative context.
As a general rule, 50+ responses give you enough signal for basic pattern recognition. Below 30, the AI should caveat findings heavily. Always include your sample size in the prompt so the AI can flag confidence limitations on its own.
For smaller datasets (under 100 rows), pasting directly works well. For larger datasets, summarize the data structure and key statistics in the prompt, then paste a representative sample. Most AI tools have context limits that make full large-dataset pasting impractical.
Add one sentence naming your industry and its typical NPS benchmarks (e.g., 'B2B SaaS industry average NPS is approximately 30-40'). This lets the AI frame your scores in competitive context rather than in a vacuum, making its recommendations more relevant.
Yes, with minor adjustments. Replace NPS-specific score ranges (0-10 promoter/detractor logic) with the relevant scale for CSAT (1-5) or CES (1-7), and reframe the driver analysis around satisfaction or effort themes rather than likelihood-to-recommend signals.