Customer Experience and Journey Analysis

Why do our satisfaction scores look fine while complaints are rising?

Several structural reasons, usually together. The programme may measure only the moments the organisation controls and none of the waiting in between, which is where frustration accumulates. The survey may fire when the case is closed in the system rather than when the customer got what they were waiting for, so it measures the interaction rather than the resolution. And a low, self-selected response rate can improve mechanically as dissatisfied customers stop responding, so a rising score on a falling base is not evidence of improvement. Check the response rate trend alongside the score trend before reading either.

What is the difference between touchpoint and relationship measurement?

They differ in base, in timing and in what they draw on. A touchpoint measure asks about a specific event soon after it, of the people who had it. A relationship measure asks about the overall connection at an arbitrary moment, of a defined customer population. They should never be averaged into a composite, and a difference between them is not a change over time. When they disagree, that is a finding: uniformly good touchpoint scores with a weak relationship score usually means the individual interactions are fine and the accumulation is not, pointing at effort, repetition or the gaps between touchpoints.

Can I say a touchpoint drives overall satisfaction?

Not from a single self-report survey. Three specific problems inflate that association independent of any real influence: common-method bias, because the touchpoint rating and the overall rating come from the same person in the same moment and share mood and halo; reverse causation, because already-satisfied customers rate everything more generously; and recency and peak weighting, because overall evaluations disproportionately reflect the most recent and most intense moments. The defensible statement is that these are the strongest correlates in this data, that direction is not established, and that a change would need testing.

How should I prioritise pain points?

On three dimensions, not on mention volume, which measures who complains and is dominated by whichever channel is easiest to complain in. Frequency: how many customers encounter it, from operational data where possible. Impact: how much it degrades the experience when it happens, measured as the score difference between those who encountered it and those who did not, with bases. Recoverability: whether the organisation can put it right afterwards. That third dimension is what makes the list a prioritisation, because a frequent, severe, unrecoverable failure needs prevention while a recoverable one needs detection and a recovery capability.

Is a dissatisfying moment the same as a churn-driving one?

Often not. People tolerate considerable friction where switching is hard, and leave over things they rated adequately, sometimes months later. Establishing the difference needs outcome data joined to experience data: subsequent retention of those who encountered a pain point against those who did not, with bases and a stated time window. Two cautions: the comparison is observational, so those who encounter a failure may differ in ways that also affect retention, and a window that is too short will show nothing. Where outcome data does not exist, say the study measures dissatisfaction and cannot identify churn drivers.

What is the survivorship problem in satisfaction data?

Satisfaction is measured among people who are still customers, so those most damaged by the experience have already left and are not in the sample. The measured distribution is truncated at the bad end, and the effect worsens the more churn there is. The real fix is surveying the lapsed population, which is rarely done. A cheaper partial fix, available in any programme with individual-level linkage, is analysing the last measured score of customers who subsequently left, which frequently shows they scored adequately shortly before leaving. At minimum, state the limitation next to every level and trend.

Does a well-handled failure really leave customers better off than no failure?

Sometimes, and it should be tested rather than assumed. It depends on the severity of the failure, the speed and completeness of the recovery, and whether the customer had to fight for it. Compare four groups on subsequent satisfaction and retention: no failure, failure with no recovery attempt, failure with an attempted recovery, and failure with a resolved recovery. Also estimate what proportion of failures the organisation ever detects, which is usually far below the failure rate, because most dissatisfied customers do not complain. That detection gap is frequently the biggest single opportunity in a programme.

What breaks when I join operational data to survey responses?

Five things, all of which need documenting. The match rate, because not every response links to a record and the unmatched are not random. Identity across channels, because the same person may be different identifiers in different systems. Timing alignment, because joining a response to the wrong event is easy and silently wrong. Definitional mismatch, because the operational definition of a resolved case and the customer's sense of resolution routinely differ. And consent, because joining survey responses to service records may exceed what the privacy statement covered, which is an ethics decision and not a data engineering one.

How do I analyse a journey that crosses several channels?

Group interactions into job-level episodes using a defined rule (same identifier, same subject, within a time window) and measure at the episode level: attempts per job, channels per job, elapsed time, first-attempt resolution rate. Channel-level measurement produces fragments that individually look fine and collectively describe a bad experience. Report the distribution rather than the average, because the story is in the tail: the small share of jobs taking four or more contacts generates most of the cost and most of the damage. State the episode rule, since it is a modelling choice that changes the numbers.

Can AI do customer experience analysis?

It is well suited to the inventory work, the linkage bookkeeping and the classification of large volumes of contact text against journey stages, all of which are tedious and error-prone by hand. The characteristic risks are filling unmeasured journey stages with plausible narrative, composing a verbatim that captures a stage perfectly with no participant behind it, describing correlates in causal language, and collapsing a severity assessment into one index with invisible weighting. Require every stage claim to carry an evidence reference or be marked unmeasured, every quote to carry an identifier, and the severity dimensions to stay separate.