Turn Churn-Survey Verbatims into Retention Plays

Workflow area: Marketing · Tier: Advanced · Hands-on time: ~45 minutes

The failure mode with verbatim analysis is cherry-picking: someone reads thirty responses, remembers the three angriest, and a quarter of roadmap gets bent around an outlier. Systematic clustering with counts keeps the loudest quote from beating the most common one.

When to use — and when not to

• Use it when you have 40+ free-text responseschurn surveys, cancellation reasons, detractor comments. Enough volume that themes are real and counting matters.
• Use it when you can act on the outputthe play format below assumes someone can own an intervention. Analysis without an owner is a slide, not a play.
• Skip it below ~20 responsesjust read them all yourself; AI clustering on tiny samples manufactures pattern where there is only anecdote.
• Skip the AI clustering for legally sensitive churncustomers alleging discrimination, safety issues or contract breaches go to a human immediately, not into a theme bucket.

The recipe

1. 1.

   Prepare the verbatims. Export survey responses to a sheet with: response ID, verbatim text, plan/segment, tenure band, and cancellation date. Delete names, emails and account numbers — the analysis needs words and segments, not identities. Number every row; the IDs are how you will audit the clustering later.

2. 2.

   Cluster with counts and evidence. Paste the verbatims (or upload the CSV) into Claude or ChatGPT:

   Copy

   These are [N] churn-survey verbatims, one per line with an ID.
   
   Cluster them into root-cause themes:
   - 5 to 8 themes maximum; name each with a specific phrase
     ("Missing integrations with Tally/Zoho", not "Product gaps").
   - For each theme: count of responses, % of total, the response IDs
     assigned, and 2 representative verbatim quotes (exact text).
   - A response may sit in up to 2 themes; say when it does.
   - Put genuinely unclassifiable responses in "Other" with IDs — do
     not force-fit them.
   - Separate SYMPTOM from CAUSE where the text allows: "too
     expensive" from a customer describing unused features is a
     value-realization cause, not a pricing cause — but only make
     that call when the verbatim supports it, and quote the evidence.
   
   Do not editorialize or propose fixes yet.
   

3. 3.

   Audit the clustering before trusting a single count. Pick two themes and read every response ID assigned to them. Check three things: the counts add up against your row total, the representative quotes exist verbatim in your data (models sometimes "improve" quotes), and the assignments are defensible. If more than one or two responses per theme are misfiled, tighten the theme names and rerun. This ten-minute audit is what makes the numbers reportable.

4. 4.

   Convert the top three themes into plays.

   Copy

   Take the top 3 themes by count. For each, draft a retention play:
   
   - Play name and the theme it addresses (with its count).
   - Trigger: which customers get this, and when (be specific —
     "day-45 accounts with zero API calls", not "at-risk users").
   - Intervention: what actually happens — email sequence, in-app
     prompt, success-team call, pricing offer, product fix.
   - Owner role: who runs this (CS lead, product, lifecycle marketing).
   - Measure: the number that says it worked, and the review date.
   - Cost/effort: S / M / L.
   
   Ground every play in the verbatim evidence. If a theme needs a
   product change we can't ship soon, say so and propose the best
   available mitigation instead.
   

   A good play is uncomfortably specific:

   Copy

   Play: "Unused-seat rescue" — addresses "Paying for seats we don't
   use" (14/86 responses, 16%).
   Trigger: accounts at renewal-minus-60 days with <40% seat activation.
   Intervention: CS call offering seat-count rightsizing + training
   session; follow-up email with usage report attached.
   Owner: CS team lead. Measure: renewal rate of triggered cohort vs.
   matched control, reviewed quarterly. Effort: S.
   

5. 5.

   Package the report. One page: theme table with counts and quotes, the three plays, and a short "what we are NOT acting on and why" section — the themes below the line, with counts, so nobody claims they were hidden. Every count in the report must reconcile to your audited step-3 numbers.

Make it yours

• Segment cut. Rerun the clustering split by plan tier or tenure band — enterprise churn and starter churn usually have different top causes, and one blended list hides both.
• Rolling quarterly pipeline. Keep a master themes sheet; each quarter's verbatims get classified against the existing theme list first ("assign to these themes; propose a new theme only if 5+ responses don't fit"), so trend lines stay comparable across quarters.
• Win-back variant. Point the same pipeline at lost-deal reasons from the CRM and the plays become competitive win-back sequences — pb-25's outreach patterns handle the drafting.

Quality check

• Counts reconciletheme counts sum correctly and you audited two clusters response-by-response.
• Quotes are verbatimevery representative quote exists character-for-character in the source.
• Themes are specific enough to act ona theme name should imply its own fix.
• Every play has a trigger, owner and measureno play ships as "we should improve X."
• The below-the-line themes are visibledeprioritized causes are listed with counts, not deleted.