How to Analyze Open-Ended Survey Responses with Claude

SurveyMonkey launched its Claude connector in May 2026, which means you can now create surveys, pull responses, and analyze open-ended verbatims without leaving a chat window. For small surveys, that workflow is excellent. For anything past a hundred open-ended responses — and certainly for NPS or CSAT programs running across quarters — the workflow that actually works looks different. This is the four-pass methodology that survey researchers have been using for years, adapted for Claude as the analyst.

The structure: four analytical passes on every batch (sentiment, themes, causation, segmentation), one framework prompt that runs all four at once, and an honest map of where Claude stops scaling so you know when to graduate.

The short answer

To analyze open-ended survey responses with Claude: set up a Claude Project with your survey context loaded once, paste responses in batches of fewer than 500, and run a framework prompt that does four analytical passes in a single run — sentiment, themes, causation, and segmentation. Output should be structured (tables, not narrative). For NPS, cluster verbatims separately by promoter / passive / detractor. For CSAT, link themes to score drops. For surveys with respondent IDs, you can ask Claude to retain those IDs in its output to enable closed-loop follow-up.

This works well for one-time surveys and quarterly programs. For continuous, longitudinal analysis where themes need to mean the same thing across multiple runs — and for respondent-level identity to persist over time — you will need infrastructure underneath Claude, not Claude alone.

How to set up Claude for survey analysis

Three things load once and apply to every run after.

The Claude Project. Create a Project specifically for survey analysis. Inside it, upload: your most recent product strategy doc, your persona definitions, your customer segments document, and any prior survey reports that established the existing theme taxonomy. Claude's context window holds roughly 200,000 tokens — enough for several hundred pages of reference material.

The custom instructions. This is your context block. Tell Claude what your product does, who your ICP is, what your current strategic focus is, and what brand voice to maintain. Tell it the analytical defaults: 8-15 themes per analysis, three quotes per theme, always flag uncertainty, structured output by default.

The survey file uploads. Export your survey responses as a CSV with one row per respondent. Include columns for: respondent ID (anonymized if needed), score (for NPS or CSAT), open-ended response, and any segmentation tags you have (tier, cohort, persona). Upload directly to Claude.

If you are using the SurveyMonkey × Claude connector, much of this is handled for you on the data-pull side, though the analytical structure below still applies.

The four-pass methodology

Survey researchers have converged on a four-pass approach for open-end analysis. The passes are: sentiment, themes, causation, segmentation. Each answers a distinct question and each compounds the value of the others. Run all four in a single prompt, not as separate analyses.

Pass 1: Sentiment. Classify each response as positive, negative, neutral, or mixed. Sentiment alone is shallow but it is the scaffolding for the other three passes.

Pass 2: Themes. Cluster responses into 8-15 named themes. Themes should be specific ("export speed", not "performance"). Each theme needs a frequency count and three representative quotes.

Pass 3: Causation. For each theme, identify what specifically drove the score. "The migration tool failed at step 3" is causation. "Migration was bad" is not. Causation is what makes the output prioritization-ready.

Pass 4: Segmentation. Break themes down by segment (tier, cohort, persona). Themes that appear in one segment but not others are the high-leverage findings. Aggregate analysis flattens this — segmented analysis surfaces it.

The framework prompt

This prompt runs all four passes in a single call. Paste it after your context block is loaded.

Analyze the open-ended survey responses below using a four-pass methodology.

PASS 1 — SENTIMENT
Tag each response: Positive / Negative / Neutral / Mixed.

PASS 2 — THEMES
Cluster responses into 8-15 themes. For each theme:
- Specific name (avoid generic categories)
- Response count
- Three representative quotes
- Dominant sentiment

PASS 3 — CAUSATION
For each theme, identify the specific cause. Be concrete: "the export feature times out on files over 100MB" not "performance issues."

PASS 4 — SEGMENTATION
Break themes down by segment (using the segmentation columns in the CSV). Flag any themes that appear in one segment but not others — those are the segment-specific findings worth attention.

OUTPUT FORMAT
1. A table: Theme | Count | Sentiment | Causation | Top Quote
2. A segmentation matrix: Theme × Segment, showing where each theme is concentrated
3. The top 5 themes worth immediate action with a one-sentence "why" each

If you are uncertain about how to group a response or classify causation, flag it explicitly rather than forcing a confident answer.

The "flag uncertainty" instruction is the single most important line in this prompt. Without it, Claude grades on confidence and produces categories that look clean but mask the edge cases.

NPS-specific: clustering by promoter, passive, detractor

NPS verbatims hide their value in the segment structure. A single theme list flattens it. Use this prompt variant for NPS.

Below are NPS responses with score (0-10) and open-ended follow-up.

Run the four-pass methodology separately for three segments:
- Promoters (9-10)
- Passives (7-8)
- Detractors (0-6)

For detractors, order themes by severity (deal-breaker / strong friction / mild friction) before frequency. For promoters, order by frequency. For passives, identify the gap themes — what would move them to a 9 or 10.

Then surface: which themes appear in detractors but not promoters (these are the highest-leverage churn drivers).

Detractor themes ordered by severity is the move that traditional NPS reports miss. The aggregate score change does not tell you whether your detractors are mildly annoyed or actively churning. Severity ordering does. The deeper walkthrough is in analyzing NPS verbatims at scale.

CSAT-specific: linking themes to score drops

CSAT analysis is most useful when themes are connected to score movements over time. Use this variant.

Below are CSAT responses with score (1-5 or 1-7) and open-ended follow-up. The data spans [time period].

Run the four-pass methodology, then add:

THEME → SCORE LINKAGE
For each theme, compute the average score of responses that mention it. Themes with averages well below the dataset's overall average are the score-suppressing themes.

TIME COMPARISON (if dates are present)
Compare theme frequency in the most recent month vs the prior month. Surface themes that are trending up sharply — those are emerging issues. Surface themes that have dropped — those are resolved or muted.

The score-linkage move is what makes CSAT data actionable. A theme that appears in 30% of responses but has an average score of 4.8 is not the problem. A theme that appears in 12% but has an average of 2.1 is the priority. The how to quantify qualitative feedback post goes deeper on this principle.

Where this workflow breaks

Three limits that Claude alone cannot overcome.

Theme stability across runs. Run the same analysis on Monday and Wednesday and Claude will return slightly different themes. For one-off surveys, this is fine. For quarterly NPS programs where you need to compare Q1 to Q2 to Q3, taxonomy drift makes longitudinal analysis unreliable. The fix requires an external, persistent taxonomy fed back into Claude on every run — which most teams do not have the infrastructure to maintain manually.

Respondent ID persistence. Claude can hold respondent IDs within a single conversation, but it does not retain them across conversations or quarters. For closed-loop follow-up — contacting the detractor who left a specific comment, tracking whether their score recovered after intervention — you need a system that links every response to a persistent customer identity. Sopact's NPS analysis methodology calls this "verbatim decay": the value of an open-ended response declines rapidly if you cannot tie it back to the customer and act on it within days.

Cross-source unification. When NPS, CSAT, support tickets, and app reviews all describe the same underlying issue in different language, you need one theme taxonomy that means the same thing across all sources. Claude can analyze each source in isolation. Unifying them requires a shared, persistent taxonomy and identity layer — infrastructure, not prompting.

When to graduate to feedback intelligence infrastructure

You are ready to move beyond Claude alone when:

• You run the same survey program quarter after quarter and need themes to mean the same thing over time
• Your survey volume exceeds a few hundred responses per batch
• You need to link survey verbatims to broader customer signals (support, sales calls, app reviews) under one taxonomy
• You need to trigger downstream actions from survey responses (CSM alerts, roadmap updates, closed-loop responses)

This is what Enterpret was built for: continuous, identity-linked, multi-source feedback intelligence with a persistent adaptive taxonomy that learns from your product's actual language, a customer context graph that holds respondent identity across time, and a Wisdom MCP Server that brings the whole layer directly into Claude. You keep working in Claude — the infrastructure persists state underneath. The deeper context is in Customer Context Graph inside Claude.

FAQ

How do you analyze open-ended survey responses with Claude?

Set up a Claude Project with your product context loaded once, then run a four-pass framework prompt on batches of fewer than 500 responses. The four passes — sentiment, themes, causation, segmentation — produce a structured output table that is prioritization-ready. Always specify the output format and ask Claude to flag uncertainty rather than force confident grouping.

How many open-ended responses can Claude analyze at once?

Claude's context window technically supports several thousand short responses, but the practical quality ceiling is around 500 per batch. Past that, themes start collapsing into vague categories and consistency between runs drops sharply. For larger surveys, batch into runs of 200-400 and reconcile the theme lists.

Can I analyze NPS open-ended responses with Claude?

Yes. The right structure is to cluster verbatims separately for promoters (9-10), passives (7-8), and detractors (0-6) — never as a unified theme list. Detractor themes should be ordered by severity, not frequency. Promoters by frequency. The themes that appear in detractors but not promoters are the highest-leverage churn drivers.

What is the SurveyMonkey × Claude integration?

Announced in May 2026, the SurveyMonkey connector for Claude lets you create surveys, pull responses, and analyze open-ended verbatims through natural-language prompts in Claude. It is most useful for small or one-off surveys. For continuous quarterly programs where themes need to remain stable over time, the connector alone does not solve the taxonomy-drift problem — that requires persistent infrastructure underneath.

When should I move beyond Claude for survey analysis?

When you need themes to be stable across runs (longitudinal analysis), when you need respondent identity to persist beyond a single conversation (closed-loop follow-up), or when you need to unify survey data with other feedback sources under one taxonomy (multi-channel intelligence). At that point you need infrastructure, not prompting.

‍