NPS Intelligence beta

NPS ↔ CSAT — THE INTERACTIVE RELATIONSHIP

60
DUAL DISCLOSURES
0.57
CORRELATION (R)
22
BREAK THE MATH
2017–2026
PAIR RANGE
Analyst brief · Interactive · July 2026

CSAT and NPS measure the same customers. The numbers still don’t agree.

The industry’s two favourite metrics can be computed from the same 0–10 question — yet across the 60 companies in our corpus that disclosed both for the same period, the correlation is just r ≈ 0.57. This page maps the geometry that connects them, and plots where real disclosures actually land.

Ask customers one question — “how likely are you to recommend us, 0–10?” — and you can publish two different numbers from the same answers. Average the scores and you have a satisfaction figure (CSAT, expressed here as the mean × 10). Take % Promoters (9–10) minus % Detractors (0–6) and you have Net Promoter Score. Same data, different arithmetic: CSAT keeps the whole distribution, NPS deliberately throws away the middle. That one design choice means the metrics are only loosely coupled — a CSAT of 80% is mathematically compatible with any NPS from 0 to +78, and an NPS of +30 with any average score from 5.9 to 8.6.

To see how the relationship behaves in the wild, we scanned the full NPS Intelligence corpus — 45,827 classified disclosures, 2011 to today — for companies quoting both metrics for the same reporting period. Sixty did, from Jet2 (CSAT 92%, NPS 65) to Shell (8.5/10, NPS 60). Across those pairs the correlation is r ≈ 0.57: knowing a company’s satisfaction score explains only about a third of the variance in its NPS (r² ≈ 0.33). Related — nowhere near interchangeable.

22 of the 60 pairs are mathematically impossible if their “CSAT” were an average. Nasuni’s 98% satisfaction would force an NPS of at least +90 (they report 88); Navan’s 97% would force +85 (they report 45). Nobody is misreporting — they are publishing top-box % satisfied, a different statistic that systematically runs higher than a mean-based score. The practical lesson: a satisfaction figure without its definition is uninterpretable — and, as section 4 shows, an NPS without its response rate isn’t much better.

Why it matters: boards treat CSAT and NPS as proxies for the same thing, and they aren’t. CSAT averages everyone, including the ambivalent middle; NPS counts only enthusiasts minus critics, so it moves when the shape of opinion changes even when the average doesn’t. The tools below make the translation explicit: the exact wedge of possible combinations, what a sample of n = 100 does to both numbers, calculators in each direction, and every real dual disclosure we could find, plotted where it falls.

1 — The geometry

Where CSAT and NPS can meet

Grey dots are simulated survey score distributions. The green area is every mathematically possible combination; the dark line is the median relationship, dashed lines the 1st and 3rd quartiles. Your calculator inputs are plotted live. Scroll to zoom · drag to pan · double-click to reset.

Mathematically possible Simulated surveys Median relationship 1st / 3rd quartile Your mix (Calc 1) Response-rate adjusted Real pair — fits envelope Real pair — breaks envelope Your CSAT (Calc 2)
Hover a diamond to identify the company…

Diamonds are 60 companies that publicly quoted both NPS and a CSAT figure for the same period, 2017–2026 (full scan of the CustomerGauge NPS Intelligence Database). 37 fit the envelope; 22 sit below the mathematical floor — proof that their disclosed "CSAT" is the % of respondents satisfied (top-box), not the mean score ×10 this axis assumes (98% satisfied ≠ a 9.8 average). One (Raiffeisen Romania '23: NPS 80, CSAT 76%) breaks the ceiling the other way. Across all 60 pairs the correlation is only r ≈ 0.57 — related metrics, but far from interchangeable. Self-reported disclosures skew favourable.

2 — Sampling reality

Sample of n = 100: NPS vs raw average score

Each grey dot is one simulated survey of exactly 100 respondents: 100 individual 0–10 scores drawn from a random underlying distribution, then averaged (X) and scored as NPS (Y). With n = 100, NPS lands only on whole numbers and the average on steps of 0.01. The green wedge is the exact envelope: for a given NPS, the average must sit between 4.5 × (1 + NPS/100) and 8 + NPS/50. The dark line is the median relationship, dashed lines the 1st and 3rd quartiles (from 4,000 simulated n=100 surveys). Hover for a live readout. Scroll to zoom · drag to pan · double-click to reset.

Possible (avg, NPS) pairs Simulated n=100 samples One fixed population, resampled 60× Median 1st / 3rd quartile
Hover over the chart…
3 — Calculators

Mix → NPS & likely CSAT

Enter Promoters and Detractors; Passives fill the remainder.

Promoters %
Detractors %
Passives %

CSAT → likely NPS

Enter CSAT as a % (average 0–10 score × 10).

4 — The response-rate stress test

Response-rate reality check

Count non-responses as Detractors
Why would you count non-responses as Detractors?

Your NPS only describes the customers who bothered to answer — and responders skew engaged. The silent majority contains churn risk you never hear from. Treating every non-responder as a Detractor is a deliberate worst-case stress test: if your "real" NPS is +30 on a 30% response rate, the stress-tested score is deeply negative, because 70% of your base is unaccounted for.

The point isn't that the adjusted number is your true NPS — it's that a headline NPS on a thin response rate is fragile. Two programs with identical NPS but 15% vs 60% response rates are not remotely equivalent. Driving response rate up is often worth more than driving the score up.

Assumptions: single 0–10 scale; Detractors 0–6, Passives 7–8, Promoters 9–10. "Likely" bands use segment-typical averages (Promoters ≈ 9.5, Passives ≈ 7.6, Detractors ≈ 4.0) and the 10th–90th percentile of simulated distributions; "possible" bounds are exact mathematical limits. NPS is a score (−100…+100), not a percentage.