CG-CAHPSthe standard, gone dark
four composites + a 0–10 rating
Built and maintained by AHRQ, it measures the reported experience of care — access, communication, coordination, office staff — not satisfaction. But AHRQ suspended data collection in 2021, so the public comparison data stops at 2018–2019.
NPSborrowed from retail
promoters (9–10) − detractors (0–6)
One question, designed for and validated against commercial industries. Against patients’ own reported experience it correlates only 0.44 — weaker than the plain 0–10 rating it replaces — and the benchmarks in circulation have no method behind them.
The whole article in one line: CG-CAHPS asks what actually happened in the visit, but its national comparison database was switched off in 2021; NPS asks one commercial question that was never validated for clinical settings, and the "healthcare averages" quoted for it contradict each other. Either way you are left comparing yourself to yourself.
This guide covers what each instrument actually measures, why the benchmarks are fiction, what the satisfaction-versus-quality evidence really says (including the part that undercuts it), and how a small practice can run a survey worth the effort.
The two instruments — and what they actually measure
CG-CAHPS
The Clinician & Group survey in the CAHPS family, developed and maintained by AHRQ. Version 3.0 measures four composites — access to care, provider communication, care coordination, and courteous and helpful office staff — plus an overall provider rating (0–10). The willingness-to-recommend item was in version 2.0 and is not in 3.0.
Note the wording: it measures the reported experience of care, not "satisfaction". That distinction is the point. "Did the provider explain things in a way that was easy to understand?" produces something you can act on. "How satisfied are you?" produces a mood. CG-CAHPS is also the de facto research standard — 52 peer-reviewed studies used its data between 2008 and 2023. Its questions are public; for an independent practice the useful move is to borrow them, not to enrol in anything.
NPS
The Net Promoter Score comes from Fred Reichheld's 2003 Harvard Business Review article, "The One Number You Need to Grow." One question, 0–10: promoters (9–10) minus detractors (0–6), giving a score from −100 to +100. It was designed for, and validated against, commercial industries. It was never built for clinical settings — a fact that turns out to matter.
There is no benchmark. Not for either one.
CG-CAHPS: the national database was switched off
In AHRQ's own words: "AHRQ suspended data collection for the Clinician & Group Survey in 2021." That database ran until 2020; the publicly available comparative data covers 2018–2019. Submissions are closed. (The CAHPS Database still operates for health plan surveys — it's the clinician-and-group side that went dark.)
So when a vendor offers to show you how your practice compares "to the national benchmark," ask which one. At best they're comparing you to data from 2019. More likely they're comparing you to their own book of business — the practices that happen to pay them.
NPS: the benchmarks contradict each other
Go looking for the average healthcare NPS and you'll find a number. Go looking twice and you'll find a different one. Here's what's actually behind the figures in circulation:
Look at that last row. Patient sentiment did not collapse by 16 points in twelve months — the sample changed. And the published range across sources, 37 to 65, is a 28-point spread on a scale where 21 points is enormous.
To be fair to the field: one organisation does hold a genuinely large patient-experience dataset. Press Ganey reports across 6.5 million patient encounters, and found medical practices reached a five-year-high likelihood-to-recommend score of 84.1 out of 100 (2023 data, published in its 2024 report). But note two things: that's likelihood-to-recommend top-box, not NPS — a different metric — and the sample is Press Ganey's own client base, which over-represents large systems. It is the best number available. It is still not a random sample of US practices.
NPS has a peer-reviewed problem in healthcare
Beyond the fake benchmarks, there's a deeper issue: NPS may simply be a poor summary of patient experience. A study of about 16,900 patients (6,018 inpatient, 10,902 outpatient) tested it against the established summary measures and found:
- NPS correlated only 0.44 with an overall score built from patients' actual reported experiences — weaker than the plain 0–10 global rating, which managed 0.52–0.54.
- 17% of patients classed as "passives" said they would definitely recommend. The bucketing throws away real signal, and treating a 6 as a "detractor" underestimates willingness to recommend.
- Drawing on earlier NHS research, the authors note patients find it strange to "recommend" a healthcare provider at all — you don't wish illness on your friends. The core question is semantically awkward in medicine.
Their conclusion is blunt: "it is still unclear what the NPS specifically adds to patient experience surveys."
The uncomfortable part: satisfaction is not quality
Here's the research that survey vendors have a structural incentive never to mention — and then the caveat that almost everyone else leaves out too.
In "The Cost of Satisfaction" (Archives of Internal Medicine, 2012), researchers followed 51,946 adults in a national survey panel, with mortality tracked in a subsample of 36,428. Compared with the least satisfied quartile, the most satisfied patients had 8.8% higher total healthcare expenditures, 9.1% higher prescription drug spending, higher odds of inpatient admission (1.12) — and an adjusted mortality hazard ratio of 1.26.
Two details the popular retellings drop. First, it wasn't "more of everything": the most satisfied patients had lower odds of any emergency-department visit (0.92). Second, the obvious objection — that sick people are grateful people — was tested. When the authors re-ran the models excluding everyone with poor self-rated health and three or more chronic conditions (about 16% of the cohort), the mortality association got stronger, not weaker: it rose to 1.44.
Then the same group revisited it — at two and a half times the scale
The 2012 paper drew heavy criticism: too few deaths, not enough adjustment for how sick people really were. So in 2019 the same research group went back to it with a much longer run of the same national panel — 92,952 people, 2000–2015. (Worth being precise: this is not an independent replication. The later sample contains the earlier one, and the authors describe it as having "revisited" the question.) With extensive, sequential adjustment for illness burden, the association held, as a clean dose-response gradient:
Adjusted 2-year mortality by patient-satisfaction quartile
And here's the part that should stop you cold
That gradient does not exist before case-mix adjustment. Adjusting only for sociodemographics, the most satisfied quartile came out at 0.91 — numerically protective, and not statistically significant (p = 0.42). The mortality association appears, in the authors' own words, only after the CAHPS-recommended adjustment for how sick patients are.
That is a real and serious caveat, and you will not find it in any article that cites this research to make a point — including, until I checked, a draft of this one. It doesn't make the finding false. Case-mix adjustment is standard practice, and there are good reasons to adjust. But it does mean the effect is a creature of a particular statistical model, not a raw fact you can see in the data. Anyone who tells you "the science shows satisfied patients die more" is overselling it. So, for that matter, is anyone who tells you the opposite.
The other side — and how to hold both
Because the opposite literature also exists, and it's substantial.
A systematic review of 55 studies (BMJ Open, 2013) found positive associations between patient experience and clinical safety and effectiveness outnumbering null findings 429 to 127. The strongest domain was adherence: an included meta-analysis found positive associations between clinician–patient communication and treatment adherence in 125 of 127 studies. And again, read the authors' own caveats, because they're the honest part: "associations do not entail causality", and "As always, there may be a publication bias in favour of studies showing positive associations."
At the hospital level, a study of 171 hospitals (Annals of Surgery, 2014) found satisfaction tracked mortality — but not safety. Compliance with process measures and patient-safety indicators didn't correlate with satisfaction; complication rates came out at p = 0.491. What satisfaction did track was large hospitals, high surgical volume and low mortality. The authors put it carefully: "The present tool is not a surrogate marker for safety but is an important measure of patient centered care."
So which is it?
Possibly both — because they measure different things, at different levels. This is the most useful idea in this article:
- At the individual patient level, satisfaction with your own clinician may partly reflect getting what you asked for. Fenton's team raised exactly this hypothesis — that a clinician who never says no earns a higher score and delivers more care. But note: no study here actually measured request-granting. It's a plausible mechanism, not a demonstrated one.
- At the organisation level, a practice that scores well on patient experience tends to be a well-run practice that also does other things well. Good clinics are good at lots of things at once.
- All of it is observational. Nobody has randomised patients to satisfaction. Nobody — including me — can claim causation in either direction.
Why it still matters for the business
None of that makes the score unimportant — it makes it a business metric rather than a clinical one. And as a business metric it's a strong one, because patients leave over the experience far more readily than over the medicine.
Accenture has reported that patients switch more readily over how easy a practice is to deal with than over the care itself. I am no longer linking a figure to it: the page I cited was replaced in July 2026 by a different report, and the older numbers are not retrievable. Its current read is that only 65% of patients say they will definitely stay. The direction I first quoted was that patients were twice as likely to switch providers after a negative front-desk or online experience than after a poor clinical one. (Fair warning, and it's the same standard I've applied to everyone else here: Accenture publishes no methodology for that survey — no sampling frame, no weighting. Treat it as a large, suggestive industry finding, not as evidence of the same grade as the peer-reviewed work above.) Directionally, it matches what practices see: the unglamorous stuff — the phones, the booking, the waiting room — is the retention stuff. It's the same mechanism behind patient retention, and it feeds straight into patient lifetime value.
One useful corrective, though, on the thing owners worry about most. The best nationally representative peer-reviewed data on how patients actually choose a physician (JAMA, n = 2,137) ranked online ratings dead last of seven factors:
An honest caveat: that survey was fielded in 2012. Review-seeking has almost certainly grown since. That it remains the best nationally representative evidence we have is itself telling about how thin this evidence base is.
First, the psychometric floor nobody mentions: a solo site clears 0.70 reliability on access, but not on communication or staff courtesy
So if you're a solo practice, be sceptical of any precise composite score you're handed — the instrument wasn't reliable at that scale in its own validation. (Reliability does improve with more responses per clinician, so it's a floor to respect rather than a wall.) Borrow the questions; don't over-read the number.
Second, the liberating finding about response rates
Everyone worries their response rate is too low to mean anything. A randomised comparison of CAHPS survey modes found:
- Web-only (emailed link): 20% response
- Mail-only: 43% response
- Sequential web-then-mail: 41%
Discouraging so far. But here's what matters: there were "no statistically significant differences among the three protocols … in any of the substantive results", and "little evidence of nonresponse bias". The people who answered the free email survey said substantially the same things as the people who answered the expensive mail one.
Then, the practical build
How a small practice actually measures it
CG-CAHPS composites are not statistically reliable at single-clinician sites. In the survey's validation study (21,318 responses across 450 practice sites), solo sites cleared the 0.70 reliability threshold on access only (0.77). Provider communication managed 0.62; office staff, 0.40. Reliability was acceptable across all composites only at sites with four or more clinicians.
Ask five questions, not forty
Map them to what CG-CAHPS actually measures: could you get an appointment when you needed one (access); did the provider explain things clearly and listen (communication); were the front-desk staff courteous and helpful (staff); an overall rating 0–10; and one free-text "what would you change?"
Send it after every visit, not in quarterly waves
Continuous, small and consistent beats a big annual push — you get a trend line instead of a snapshot, and you catch a problem in the month it appears.
Know roughly what a sample can carry
The rule you'll see everywhere — about 200 responses overall, at least 50 per physician before judging an individual clinician — traces back to a 1999 practice-management article quoting the head of a patient-survey company. So treat it as a rough floor, not a law. The peer-reviewed constraint above (reliability at solo sites) is the one to actually respect.
Keep it anonymous and keep the scale consistent
Use the same scale every time or your trend is meaningless. Avoid double-barrelled questions ("were the doctor and nurse courteous?") — you won't know which one they're answering about.
Read the comments before you look at the score
AHRQ's own synthesis is explicit that verbatim comments "add specific information that may not be captured in CAHPS survey scores" and give more actionable feedback than the closed questions alone. The number tells you whether. The comments tell you what.
What to do with the number
Segment it — by clinician, by site, by visit type. A single practice-wide score hides the one provider or the one location dragging it down, and that's the only actionable thing in the data. Feed it back regularly: at UCLA Health, sharing CAHPS scores and verbatim comments in a monthly report is credited with strengthening physician buy-in.
And two honest warnings from that same AHRQ synthesis. Coaching works, but the gains fade — programmes pairing mid-performing providers with high-performing coaches improved communication scores, and then those improvements decayed over time. And the time lag between collecting data and getting it to the person who can act on it is a documented barrier — as is the belief among some clinicians and staff that the results do not represent their whole patient population.
Patient satisfaction is one of the 12 KPIs every practice should track — and of all twelve, it's the one most often measured badly and understood worst. Track your own trend, read the comments, split it by provider, and ignore anyone selling you a benchmark.
Frequently asked questions
There is no credible benchmark for either. AHRQ suspended data collection for its national CG-CAHPS database in 2021, so the newest national comparison data is from 2019. Published "average healthcare NPS" figures have ranged from about 37 to 65, on samples that are usually self-selected. The largest patient-experience dataset that does exist — Press Ganey's, covering 6.5 million encounters — reports a different metric (likelihood-to-recommend top-box, 84.1 for medical practices on 2023 data), and reflects their own client base rather than a random sample of US practices. Measure your own score and track your own trend.
The Clinician & Group survey in the CAHPS family, developed and maintained by AHRQ. Version 3.0 measures four composites — access to care, provider communication, care coordination, and courteous and helpful office staff — plus an overall provider rating (0–10). The willingness-to-recommend item was in version 2.0 and is not in 3.0. It measures the reported experience of care rather than "satisfaction", and it's the de facto research standard for ambulatory patient experience.
As an internal trend line, it's fine. As a benchmark, no. A peer-reviewed study of about 16,900 patients found NPS correlated only 0.44 with an overall score built from patients' actual reported experiences — weaker than the plain 0–10 global rating (0.52–0.54) — and that 17% of patients classed as "passives" said they would definitely recommend, signal the buckets throw away. The authors concluded it's "still unclear what the NPS specifically adds to patient experience surveys."
No — but the evidence is more tangled than either side admits. Two US studies of national survey data found patients in the highest satisfaction quartile had higher spending and higher mortality (1.26 in 2012; a gradient up to 1.57 in a 2019 extension with 92,952 people). Crucially, in the 2019 study that association did not exist before case-mix adjustment — adjusting only for sociodemographics, the top quartile was numerically protective (0.91, not significant) — and it was statistically significant only in women. Meanwhile other research finds patient experience is positively associated with adherence and safety at the organisation level. All of it is observational. Satisfaction is a valid measure of the experience of care; it is not a proxy for clinical quality.
The widely repeated rule — about 200 responses overall, at least 50 per physician — traces to a 1999 practice-management article quoting the head of a patient-survey company, so treat it as a rough floor rather than a law. The solid, peer-reviewed constraint is different: CG-CAHPS composites aren't statistically reliable at single-clinician sites, where only the access composite reached the 0.70 threshold; reliability was acceptable across all composites only at sites with four or more clinicians.
WRITTEN BY
Olha · clinic data analyst
I build the reporting our managers open every morning at a multi-branch medical clinic — and package it so other practices don't have to start from scratch.
Every figure here was checked against its primary source, and where a source is weak I've said so in the text rather than quietly leaning on it. The satisfaction–outcome studies are all observational: associations are not causation, and I've reported the authors' own limitations — including the one that most undercuts the argument — rather than paraphrasing around them. Lucid Vitals is not affiliated with AHRQ or Microsoft.