One recommend question cannot carry an improvement decision
Aliases: single-item NPS · recommend intent is not a fix plan · NPS underdetermines action
What it is
Net Promoter Score (NPS) is usually one item: how likely you are to recommend the product to a friend or colleague, answered 0–10. The item measures recommend intent, not defect location, task failure, or which people were harmed. A rising or falling score can create the feeling that action is required without giving action a direction. Using it alone to justify a redesign, a roadmap, or a budget treats a coarse thermometer as a construction drawing. Open comments can supply reasons, but that is no longer the single-item score, and comments are incomplete and not one-to-one with the number.
Why it happens
Recommend intent compresses many inputs: the last outcome, price, brand, whether there is anyone to recommend to, social risk. Opposite stories can sit behind the same digit—a low score from a failed critical task, or from having no use worth speaking of. Decisions need an intervenable mechanism (which step, for whom, what failure); a single item carries none of those coordinates. Once an organization treats the item as the whole “voice of the customer,” discussion stops at the tick of the score and the motive to collect other evidence drops. The more it is the only indicator, the less it can answer “what to change.”
Studying it
Align NPS with concurrently collected task outcomes, error types, and support themes, and test whether reasons inside the same score band are homogeneous. Heterogeneous reasons under one score mean the item has low information for improvement. A follow-up “single main reason” coded against the score can show the same: a scattered cross-tab means the score does not map onto actions. Comparing decisions made from NPS alone with decisions made from coded reasons, then checking whether they pointed at the same change, estimates the mis-fix rate when the item is used as a drawing.
Where it stops holding
As a long-running alarm, the item can say “direction changed, go look.” An alarm is not a diagnosis. For a tiny product with one obvious touchpoint, the score sits closer to that touchpoint’s experience and still needs behavioral evidence to change design. Forcing everyone to write a reason lowers completion; the reasons are from people willing to write, and cannot be filled back onto every rater. Treating recommend intent as loyalty or as actual later recommending is a further jump: stated likelihood is not whether someone later speaks.
Applying it
- Let NPS enter only the threshold “do we open an investigation,” not the list “which screen to change.”
- When the score moves past a threshold, run a pre-specified reason collection (session replay, tickets, targeted interviews). No redesign is funded from the score without reason material.
- Split the report: score in one column, reason categories in another. A colored arrow is not a conclusions page.
- Audit the last change made because NPS moved. If it cannot be traced to a concrete failure location, strip the metric of decision rights.