Most satisfaction programmes fail at the questionnaire rather than at the fieldwork. A survey asking fourteen agree/disagree statements produces fourteen numbers that all move together, and a report that states one headline satisfaction figure and shows everything correlating with everything else.
Why the fourteen statements all agree
Because respondents answer the overall impression, not the individual item. Someone who liked the visit agrees with all fourteen; someone who did not disagrees with all fourteen. The instrument records one judgement fourteen times and presents it as fourteen findings.
If every item correlates above 0.8 with every other item, the questionnaire has one question in it.The correlation matrix test
The tell is in the correlation matrix, and it is worth looking at before publishing anything. If every item correlates above 0.8 with every other item, the questionnaire has one question in it.
Shorter instruments answer more
Ask fewer questions and make them do different jobs. An overall satisfaction measure, one behavioural intent measure, a small number of attribute questions covering things you can actually change, and at least one open response. The open response is usually where the finding is, and it is the first thing cut when a questionnaire is being shortened by someone counting questions rather than reading them.
Score against what the sample can support
A national number and a branch league table need very different sample sizes. Publishing the second from a sample built for the first is the most common failure in this work, and it is invisible in the output — a branch score of 68 based on nine responses looks identical to one based on four hundred.
Report the sample size and the response rate next to every number. Where a cut is too small to publish, say so rather than publishing it with a footnote nobody reads.
Who invites decides what you learn
Where staff hand out the survey, scores rise and the data stops being usable. It is not dishonesty — a customer asked face to face by the person who just served them answers a different question from one asked by an automated message the next day.
Invitations should come from your own system after the transaction, or from independent intercept. If neither is possible, the limitation belongs in the report rather than in a footnote.
Arabic is not a translation
A question that reads neutrally in English can read as leading, or as an accusation, in Arabic. Instruments are written to carry the same meaning in both languages and piloted separately in each, because a translated questionnaire that has only been piloted in English has been piloted in half the languages it will run in.
Do not reset a trend you already paid for
Where an NPS or CSAT programme already exists, replacing it costs you the comparison you have been building. Benchmark against it and add the diagnostic questions it is missing. A new metric that starts the trend line again is a cost dressed as an improvement — which is why our satisfaction survey work begins by reading your existing instrument rather than replacing it.