CSAT Trends Report

Chart survey/sentiment scores across a period and slice them by client, tech, or
category — with RESPONSE-RATE HONESTY as the organizing principle. CSAT's dirty secret
is that most tickets never get scored; this report always shows the N next to the score
and says plainly when the N is too small to mean anything.

1. Confirm the period (default: last 6 months, monthly buckets) and the slice requested
   (overall, by client, by tech, by category). Establish what satisfaction signal
   exists for this desk — explicit survey scores, thumbs ratings, or sentiment analysis
   on threads — and label the output with which one; they are not interchangeable and
   must never be blended into one trend line (if both exist, show two labeled lines).

2. Run split searches per bucket: total resolved tickets, tickets with a satisfaction
   response, and the response values. Record result caps; never present capped searches
   as exact totals.

3. Compute per bucket: score, response count (N), and response rate (N / resolved).
   EVERY score is displayed with its N. Confidence labeling: under ~10 responses =
   "anecdote — do not act on this number alone"; 10-30 = "directional"; 30+ = "usable".
   Low N is said plainly, every time, next to the number — never in a footnote. Do not
   merge months to hide a thin N without saying so.

4. TREND READ — is the score moving, or is the response RATE moving? A falling response
   rate with a stable score often means only angry people still answer (or only happy
   ones); check whether the responding population changed before declaring a
   satisfaction change.

5. SLICE — per client / tech / category tables, sorted by score, but only for slices
   clearing the directional threshold; lump the rest into "insufficient responses"
   rather than printing misleading rows. Per-tech CSAT is role- and mix-sensitive:
   techs handling escalations and outages meet angrier people — benchmark role-aware,
   never rank on CSAT alone, and never volume alone either.

6. For any genuinely declining slice with adequate N, pull 3-5 negative responses and
   summarize the recurring theme (AGGREGATE the complaint, don't quote clients verbatim
   into a report others will circulate; name a specific ticket only when a single
   incident explains a score cliff).

7. Output: trend table (bucket / score / N / response rate / confidence), slice tables,
   one paragraph on the biggest real movement, and a methodology note (signal type,
   period, thresholds, searches, caps).