NPS gets a bad reputation because most teams run it badly. The number is treated as a quarterly trophy, the follow-up is silent, and nobody can tell you what changed when the score moves. Done right, it is still one of the cheapest signals you can put in front of a product team.

Here is how to run NPS as an actual program in 2026.

What does NPS measure that other metrics don't?

NPS asks one question: "How likely are you to recommend us, 0–10?" The score is the percentage of promoters (9–10) minus the percentage of detractors (0–6). The methodology was introduced by Bain & Company and Fred Reichheld in 2003 and the canonical definition has not changed since.

It is good at: trending sentiment over time, catching big shifts, segmenting customers by attitude.

It is bad at: telling you why anything changed, comparing across industries, predicting individual churn, or being the single KPI of anything.

The trap is treating the number as the answer. The number is the signal. The follow-up is the answer.

How should you word the NPS question in 2026?

The classic NPS phrasing is "How likely are you to recommend us to a friend or colleague?" In 2026, that wording is showing its age. Two updates we use:

  • Replace "a friend or colleague" with someone the respondent might actually recommend you to. For B2B SaaS, "another team at a company like yours" beats "a friend". For consumer products, "a friend" still wins.
  • Anchor the scale at both ends. "0 = definitely not, 10 = definitely yes" noticeably reduces middle-clicking in our experience (the effect is well documented in survey-methodology literature on labelled endpoints).

The 0–10 scale stays. Five-point variants ("CSAT-style NPS") lose the bimodal distribution that makes NPS readable.

Who should you sample, and how often?

The most common NPS mistake is asking everyone, every quarter. Two problems with that:

  • You ask new customers who have not formed an opinion yet, which adds noise.
  • You ask the same loyal customers every quarter, which both annoys them and biases the panel toward survivors.

A better cadence:

  • Transactional NPS: triggered N days after a meaningful event (onboarding completed, first successful job ran, first invoice paid). The cleanest signal because you know what they are scoring.
  • Relationship NPS: a rolling sample, 1/12th of your customer base per month, no customer asked more often than every 90 days. Gives you a continuous read without survey fatigue.

Skip "everyone, every quarter". It produces a score with no actionable subgroup data.

What follow-up question turns the score into action?

The follow-up is the whole game. The single most important screen in an NPS survey is the second one.

For promoters (9–10): "Great — what's the one thing you'd tell that other team about us?" The answer is your next testimonial.

For passives (7–8): "What's the one thing that would have made this a 9 or 10?" The answer is your next feature priority.

For detractors (0–6): "What's the one thing you'd fix first?" The answer is your next sprint.

Three different follow-up screens, branched by score. This is where conditional logic in your survey tool earns its keep. If your survey can't branch by score, you are running it on the wrong tool.

How do you close the loop on a response?

Half the value of NPS is in the response, not the score. The rule we use:

  • Detractor responses get a personal email from a human within 48 hours. Not a templated apology — a question.
  • Passive responses get a templated thank-you and a link to roadmap items that match their feedback.
  • Promoter responses get an ask: testimonial, referral, case study, review.

In our experience, teams that do this consistently see their detractor rate fall meaningfully — often by a quarter or more — over two cycles. Not because the product changed, but because detractors became people who felt heard.

What to do with the actual number

Trend it, do not target it.

Targeting an NPS number turns the survey into a gameable metric. Reps start asking customers for 10s, support agents start asking favors. The number goes up, the underlying truth does not.

Trending it works:

  • Plot it monthly, not quarterly. Quarterly hides shifts.
  • Segment by plan tier, by tenure, by acquisition channel. The aggregate number is almost never the interesting one.
  • Alert on a 10-point drop in any segment, week over week. That is the signal worth interrupting your week for.

When should you stop using NPS?

NPS stops being useful when your respondent base is too small. In our experience, below roughly fifty responses per period, the score is noise — Bain's own NPS guidance is upfront that the math only stabilises with a meaningful sample. In that case, switch to CSAT ("How would you rate your experience, 1–5?") which is more sensitive to small samples, or a simple thumbs-up/thumbs-down with a mandatory comment.

NPS also stops being useful when your product has multiple very different audiences. A two-sided marketplace, for example, should run NPS separately per side. A blended score across audiences with opposite incentives means nothing.

The shape of a good NPS program

  • Triggered by an event, not a calendar.
  • Branched follow-up by score.
  • Closed-loop response for every detractor within 48 hours.
  • Trended monthly, segmented by tier and tenure.
  • Never tied to a comp plan.

That is the program. The number on the dashboard is the cheapest part.

Related from this desk