Symptom: Tags multiply and become impossible to maintain.
Cause: No governance: anyone creates new tags without checking for existing ones.
Fix: Define a small taxonomy first; use bulk-tag and merge instead of creating one-off tags.
Technical SEO
Title, meta description, canonical URL, OG tags, Twitter cards, BreadcrumbList, FAQPage schema, and WebApplication schema are configured for this insight clustering helper route.
Affinity Mapping Tool for User Research (Free)
Cluster user interview notes into evidence-backed candidate patterns, with every claim traceable to its source. Each pattern shows supporting evidence units beside unique sources, preserved contradictions, segment differences, and explicit caveats, so a stakeholder can challenge the synthesis instead of trusting a mention count.
Practical product-research affinity synthesis, not academic thematic analysis and not an AI research repository: local TF-IDF clustering drafts the groups, you keep full control, and frequency is never presented as priority. No login required; everything runs in your browser.
- Import notes via CSV (paste or upload), tag and cluster them with auto-assist (TF-IDF and k-means), and build themes with evidence and segment breakdowns.
- Export CSV, Markdown, JSON, or XLSX and share a URL snapshot so your team can review or continue synthesis later.
- Use the stakeholder summary and co-occurrence view to spot patterns and align on top themes before roadmap planning.
No login. Runs in your browser. We do not store your data on our servers.
No login. Runs in your browser. We do not store your data on our servers.
0 evidence unit(s) across 0 unique source(s) · Frequency is coverage, not priority.
Paste CSV below, then click Import.
Notes (0)
| Select | ID | Text | Source | Segment | Tags | State |
|---|
Paste CSV or load an example to add notes.
How it works
Paste or upload qualitative notes as atomic observations (one insight or quote per row). Organize them with tags and run local clustering (TF-IDF + k-means) to discover groups. Turn high-cohesion clusters into named themes with clear insight statements, evidence quotes, and segment breakdowns. Export or share a URL so your team can review or continue synthesis later.
Clustering
Auto-cluster uses TF-IDF and k-means (cosine similarity) to group similar notes into 5–12 clusters. You can set k manually or use auto. Clusters get suggested titles and keywords; rename and merge them to match how you think about the data. Treat clustering as a draft, not final truth.
Candidate patterns, not validated findings
Patterns are insight statements backed by inspectable evidence. Select notes (e.g. from a cluster or by tag), then write the pattern with the observation kept separate from your interpretation and the product implication. Mark contradicting evidence instead of hiding it, split by segment when disagreement clusters by context, and record caveats. Copy the synthesis record to hand problem framing an artifact it can trust, and challenge.
Provenance: no synthesis without inspectable source evidence
Every evidence unit carries source, participant, segment, and evidence type (quote versus observed behavior), and the ledger reports evidence units beside unique sources, never calling notes “users.” Duplicate text is flagged so five copies of one note cannot pose as five voices, rows can be excluded without deletion, and unclustered outliers are allowed to stay unclustered.
Pattern stability without fake confidence
There is no 87% confidence score and no saturation badge. Each pattern gets a descriptive diagnostic: supporting sources, contradicting sources, source concentration, duplicate risk, and segment spread. A pattern resting on one source is flagged fragile; a contested pattern points at the segment split that would resolve it. Frequency measures coverage, a rare but severe issue can still matter, and nothing here auto-prioritizes the backlog.
Exports and handoff
Export CSV (one evidence unit per row with provenance, exclusions, and cluster memberships), Markdown research synthesis records with ledger, contradictions, and caveats, JSON (lossless versioned workspace), or XLSX (Notes, Clusters, Themes, Tags, Segments). Patterns hand off to problem statements, opportunity trees, or further research, clustering never writes a validated problem, a persona, or a priority by itself. Share URLs encode the workspace for review; redact customer details first, because anyone with the link can reconstruct the content.
Pro tips
- Keep notes atomic: one observation or quote per row so tags and clusters map cleanly.
- Define a small tag taxonomy before bulk-tagging so you avoid tag sprawl and near-duplicates.
- Use segments (e.g. participant_id, question, source) so you can break down themes by cohort or question.
- Run auto-cluster with k=5–8 first, then rename clusters and merge or split based on your judgment.
- Turn high-cohesion clusters into candidate patterns with a context-specific statement, 2–3 evidence quotes, and explicit caveats.
- Preserve contradicting evidence per pattern and split by segment instead of averaging disagreement away.
- Report evidence units beside unique sources; five notes from one participant are one source, not five users.
- Export the stakeholder summary early and share it with one partner before socializing broadly.
- Keep theme statements testable: “Users struggle with X when Y” is better than “Users want better X.”
- Write observations separately from interpretations so a quote never silently becomes a need or a solution.
Common mistakes
Symptom: Clusters feel random or don’t match how you think about the data.
Cause: Relying only on auto-cluster without renaming or merging clusters.
Fix: Treat auto-cluster as a draft: rename clusters, merge similar ones, and assign notes manually when needed.
Symptom: Themes are vague and not defensible with evidence.
Cause: Writing theme statements without tying them to specific notes or quotes.
Fix: Attach 2–3 evidence note IDs to each theme and add a short quote in the stakeholder summary.
Symptom: One theme dominates because the same note was pasted several times.
Cause: Duplicate evidence inflates mention counts while unique sources stay flat.
Fix: Check the stability ledger for duplicate-risk units; collapse duplicates and report units beside unique sources.
Symptom: Contradictory interviews get averaged into one vague theme.
Cause: Dissent is treated as noise instead of segment or context signal.
Fix: Mark contradicting evidence on the pattern and split by segment, role, or lifecycle before deciding.
Symptom: Stakeholders disagree on what the top themes mean.
Cause: Theme titles and statements are ambiguous or not tied to a clear insight.
Fix: Write theme statements in one sentence with a clear “so what” and keep evidence visible.
Symptom: Export doesn’t match what you see in the UI.
Cause: Export runs on a stale or partial state (e.g. before last clustering).
Fix: Re-export after any major change; use JSON export as the source of truth for roundtrips.
Symptom: Auto-tag suggestions are noisy or irrelevant.
Cause: Suggestions are purely frequency-based and ignore context or segment.
Fix: Use suggestions as a starting set; then refine by merging and deleting low-value tags.
FAQ
Is this insight clustering tool free?
Yes. The tool is free and runs in your browser with no login. You can import notes, tag, cluster, build themes, and export CSV, Markdown, JSON, or XLSX without limits. There are no paywalls, trial caps, or account requirements; all processing happens locally so your qualitative data never leaves your device.
Can I use this for both interviews and support tickets?
Yes. The tool works with any atomic notes: interview highlights, survey open-ends, or support ticket summaries. Use the segment column to separate sources (e.g. interview vs support) and break down themes by segment. The same tagging and clustering workflow applies so you can compare patterns across data types.
How does auto-clustering work?
Notes are turned into text vectors using TF-IDF, then grouped with k-means (cosine similarity) into 5–12 clusters. You can set k or use auto. Clusters get suggested titles and keywords, but every cluster stays editable: rename, merge, move evidence, or leave notes unclustered. All processing runs locally in your browser, so no note text is sent anywhere.
What is the difference between tags and clusters?
Tags are labels you assign (or suggest) to notes; they can be hierarchical and you control the taxonomy. Clusters are groups from an algorithm; they help you discover structure. Often you turn high-cohesion clusters into themes with insight statements and evidence so stakeholders can act on them.
How do I create a theme?
Select notes from a cluster or tag, then create a candidate pattern with a statement, an observation, and your interpretation kept separate. The tool records evidence IDs, unique-source counts, segment breakdown, contradicting evidence, and caveats. Copy the synthesis record to hand a defensible, inspectable pattern to problem framing instead of a bare claim.
Can I share my workspace with someone else?
Yes. Share generates a URL that encodes the full workspace so a teammate can open and continue it. Research notes may contain customer names or confidential details, so redact before sharing: anyone with the link can reconstruct the content. The link is compressed, not encrypted, and nothing is stored on CraftUp servers.
What CSV format do I need for import?
Use the template: note_id, text, source, participant_id, segment, question, timestamp, evidence_type, excluded. Only text is required; participant or source IDs preserve provenance, segment enables divergence views, and evidence_type labels quotes versus observations. You can exclude rows without deleting them, and fix source IDs after import.
Does the tool support large datasets?
It is designed for hundreds of notes (e.g. 500) in the browser. For very large datasets, consider splitting by segment or importing in batches and merging tag/themes after. Performance stays reasonable because clustering and tagging run locally without server round-trips.
Is theme frequency the same as importance?
No. Counts measure coverage, not priority or truth. Five notes from one participant are one source, not five users, and duplicates can inflate a theme. The ledger shows evidence units beside unique sources, flags concentration, and never auto-prioritizes. A rare but severe issue can still deserve action through problem framing and risk review.
How does this fit with other CraftUp discovery tools?
Cluster after interviews from the Interview Script Generator, then hand candidate patterns to problem statements, JTBD, personas, or opportunity trees. Clustering never writes validated problems, personas, or priorities by itself. Survey measurement, triangulation across methods, and prioritization stay separate decisions owned by their own workflows.
Learn more with CraftUp
Courses, blog guides, and glossary entries to deepen your qualitative synthesis and discovery skills.
Turn qualitative notes into clear themes
Use CraftUp’s discovery tools and courses to connect research synthesis with JTBD, personas, and prioritization so insights lead to action.
Freshness
Last updated: 2026-03-05
- Launched Insight Clustering Helper with tagging, TF-IDF/k-means clustering, and theme building.
- Added CSV/MD/JSON/XLSX exports and shareable URL snapshots for team handoffs.
- Included 3 example datasets (B2C learning, B2B workflow, support tickets) with 30+ notes each.
- Added stakeholder summary, co-occurrence view, and auto-tag suggestions (local, disclosed).