18 January 2026 · Measurement

When CSAT misleads the funnel

A kinder score after a bot launch is not automatically a healthier path. Sometimes the people who would have scored you poorly never reached the survey.

Quiet office aisle suggesting survey bias after hours

Customer satisfaction surveys sit at the far end of many support stories. They are easy to put on a slide. They are a terrible primary metric for Support Funnel App Analytics. The survey only samples people who stayed long enough to be asked. If a bot or a dead-end article pushes the impatient away before ticket birth, CSAT among those who remain can rise while the Atlas is leaking.

We saw this pattern after an in-app help redesign at a consumer app whose night coverage ran from Bangkok. Article views climbed. Bot “resolution” climbed. CSAT on closed tickets ticked up by a few points. Repeat contacts on the same intent, when finally tagged, also climbed. The survey was interviewing a calmer remainder. The funnel had not improved; the sampling frame had shrunk.

Identity Nodecore does not tell desks to abandon CSAT. We tell them to place it after named edges, never instead of them. In a leadership brief, CSAT may appear as a secondary observation, with the response rate and the share of intents that never created a ticket sitting beside it. If you cannot estimate that share, you write “unknown” rather than implying the score represents the whole path.

Agent-level CSAT is a separate political object. Funnel work that starts by ranking people usually dies in the room. The Method keeps the unit of analysis on edges and intents. When a score moves, we ask which stage changed: search, article, bot, first reply, promise, close. If nobody can point to a stage, the score is gossip with a decimal.

Thai-language and English surveys on the same desk add another distortion. Mixing them without a language split makes a blended number that no Atlas stage owns. Split first. Then argue.

← Back to the journal