Guide
What gage R&R result means a supplier's measurements can be trusted?
Gage R&R is a designed study that measures how much of the spread in a set of readings comes from the measurement system itself: repeatability, the variation when one person measures the same part several times with the same gage, and reproducibility, the variation between different people, fixtures or sites measuring the same parts. Its headline result, %GRR, tells you whether a difference between two readings, or between a supplier's inspection report and your receiving inspection, is real or just gage noise.
By Saola Manufacturing engineering team · Published · Updated
This guide is for the quality engineer or buyer who has a supplier's gage study in hand, or a measurement dispute on the table, and needs to decide what the numbers are worth. It explains what the familiar 10 and 30 percent limits actually measure, why the same study can pass on one basis and fail on another, which mistakes make a report look better than the gage is, and how a correlation study settles whose readings govern. How measurement error then distorts process capability is covered in the Cpk vs Ppk guide; where a gage study fits in an approval package is covered in the PPAP guide.
On this page
What does a gage R&R study measure, and what does it leave out?
It splits the observed variation into repeatability (equipment variation), reproducibility (appraiser variation) and the real part-to-part variation. It does not measure bias, so a gage with an excellent %GRR can still read consistently wrong.
The classic study is crossed: a set of parts is measured by several appraisers, each of whom measures every part more than once, in random order and without seeing earlier readings. The usual design in industry guidance is ten parts, three appraisers and two or three trials each. The analysis then estimates three standard deviations. Equipment variation (EV) is the scatter when the same person re-measures the same part. Appraiser variation (AV) is the shift between people. Part variation (PV) is the true difference between the parts, estimated from the spread of the part averages. Combined gage variation is the root sum of squares of EV and AV, and total variation is the root sum of squares of the gage variation and PV.
There are two common ways to do the arithmetic. The average-and-range method is quick to compute by hand. The analysis of variance (ANOVA) method is preferred by most statistical software because it also separates an appraiser-by-part interaction: one inspector who reads small parts high and large parts low, for example because of how they seat the part in a fixture.
What gage R&R does not tell you matters just as much in a dispute. Bias (how far the average reading sits from the true value), linearity (whether bias changes across the range) and stability (whether the gage drifts over weeks) are separate studies; the gauge study section of the NIST/SEMATECH e-Handbook likewise treats bias as its own analysis, apart from the variability estimates. Bias is controlled by calibration against a reference, and the reference only means something if it is traceable; NIST defines traceability as a result related to a stated reference through a documented unbroken chain of calibrations. Two gages can each have a fine %GRR and still disagree with each other, because each is precise around a different center.
What do the typical gage R&R limits of 10% and 30% mean?
Per the AIAG Measurement Systems Analysis manual, a %GRR under 10% is generally acceptable, 10% to 30% may be acceptable depending on the importance of the measurement and the cost of a better gage, and over 30% is unacceptable. The limits are industry guidance, not physics, and they mean little until you know what the percentage was divided by.
%GRR compares six standard deviations of gage variation with either the tolerance width or the total observed variation. On the tolerance basis, under the typical 10% limit the gage spread uses up a tenth of the band, so only parts sitting right at a limit can be misjudged. At the typical 30% limit nearly a third of the band is measurement noise, and a reading anywhere near either limit could belong to a good part or a bad one. Above it, the gage cannot reliably tell a conforming part from a nonconforming one close to the edge, and no amount of inspection effort fixes that.
The ratios are on a standard-deviation basis, so they do not add. The squares add, as the AIAG manual itself cautions: on the total-variation basis, a %GRR of 30% means the gage accounts for roughly 9% of the total variance (0.30 squared), and %EV squared plus %AV squared equals %GRR squared. That is why, in a typical report, a study with %EV of 20% and %AV of 20% shows a %GRR of about 28%, not 40%.
The middle band is a decision, not a pass. Ask what the characteristic does: a sealing bore or a bearing fit deserves a better gage or a tighter method; a clearance hole for a bolt can typically live with 25%. The same AIAG manual adds a second test, the number of distinct categories (ndc), which counts how many groups of parts the gage can separate across the process spread. It should be at least 5. A gage with ndc of 1 or 2 is effectively a go/no-go gage, whatever its display shows.
| Criterion | Acceptable | Marginal (decide case by case) | Unacceptable |
|---|---|---|---|
| %GRR (to tolerance or to total variation) | under 10 % | 10 % to 30 % | over 30 % |
| Number of distinct categories (ndc) | 5 or more | — | under 5 |
| Gage resolution | 1/10 of tolerance or process spread, or finer | — | coarser than 1/10 |
| Attribute study: effectiveness | 90 % or more | 80 % or more | under 80 % |
| Attribute study: miss rate (bad part accepted) | 2 % or less | 5 % or less | over 5 % |
| Attribute study: false alarm rate (good part rejected) | 5 % or less | 10 % or less | over 10 % |
Should %GRR be judged against the tolerance or the total variation?
Against the tolerance when the gage is used to accept or reject parts, and against the total process variation when it feeds control charts or capability indices. The same study can pass on one basis and fail on the other, so a report that does not state its basis cannot be read.
The tolerance basis asks: can this gage sort good parts from bad ones? The total-variation basis asks: can this gage see the process move? A turning or milling process that is well inside its tolerance has a small part-to-part spread, so a gage that is perfectly adequate for inspection can still be too coarse to chart that process. The worked example below takes one study on a bored diameter and reports it three ways.
Two details change the answer without anyone noticing. The current AIAG edition uses a six-sigma multiplier; older editions and some older software use 5.15 sigma, which makes the same gage look about 14% better on the tolerance basis, because 5.15 divided by 6 is 0.86. And the total-variation result depends entirely on which parts went into the study. Ten parts picked from one short run of a stable process have almost no spread, so the gage looks terrible by comparison; ten parts hand-picked from scrap bins at both ends of the tolerance make the same gage look excellent. Per the AIAG manual, the parts should represent the expected range of the process, and if they cannot, the process variation can be taken from a capability study instead.
| Quantity | Value | Reading |
|---|---|---|
| Tolerance width | 0.020 mm | Input |
| Gage standard deviation (EV and AV combined) | 0.0006 mm | Input from the study |
| %GRR to tolerance, 6 sigma basis | 18.0 % | Marginal: may be accepted for inspection |
| %GRR to tolerance, 5.15 sigma basis (older templates) | 15.5 % | Same gage, looks better |
| %GRR to total variation, parts spanning the process (part sigma 0.0015 mm) | 37.1 % | Unacceptable for process control; ndc 3 |
| %GRR to total variation, ten consecutive parts (part sigma 0.0005 mm) | 76.8 % | Looks unusable; ndc 1; the parts, not the gage, caused it |
Which mistakes make a supplier's gage study look better or worse than the gage is?
Most misleading studies come from the choice of parts, people or method rather than the gage itself. Every one of them is visible in the raw worksheet, which is why the worksheet matters more than the summary page.
- Wrong appraisers. The study is run by the metrology technician instead of the production operators who will actually use the gage, so appraiser variation is understated.
- Trials that are not blind. If appraisers measure the parts in the same order every time and can see their earlier readings, they remember them, and repeatability looks better than it is.
- Resolution too coarse. When most of the ranges in the worksheet are zero, the gage is rounding every reading to the same value. Industry guidance is a resolution of one tenth of the tolerance or finer; a caliper reading in hundredths of a millimeter cannot support a bore with a total tolerance of 0.010 mm, such as a tight bearing seat.
- A different gage or fixture. The study was run on a bench instrument in the inspection room, but the parts are accepted with a hand gage at the machine, or the reverse.
- Warm parts. Parts measured straight off the machine and again later at room temperature produce a shift that looks like appraiser variation. Dimensions are defined at 20 °C per ISO 1, so the method should state how long parts soak before measuring.
- One study reused everywhere. A study on one feature does not qualify the same instrument on a different feature, size or material, because the fixturing and the part's own form error change the result.
- No basis stated. A single %GRR with no indication of tolerance or total variation, sigma multiplier, or method, which is the most common gap in reports from suppliers who rarely receive the request.
When supplier and customer measurements disagree, how do you decide whose numbers are valid?
First confirm that each side's gage passes its own gage R&R, then run a correlation study on the same serialized parts to measure the bias between the two systems. The system with acceptable R&R and traceable calibration, measuring by the agreed method, governs.
Disputes usually start with pass/fail results: the supplier's report says conforming, receiving inspection rejects a handful of parts. Pass/fail data cannot resolve anything. Exchange the actual readings on the disputed characteristic, then work through four steps.
- Qualify both gages. If either side's %GRR on that characteristic is above the typical 30% limit, its readings near the limits are not evidence of anything, and that side fixes its measurement before arguing further.
- Correlate on shared parts. Number a set of parts that span the range, typically ten or more, and have both sides measure the same parts by their normal method. The average of the paired differences is the bias between the two systems; compare it with each side's gage variation. A bias larger than the gage noise points to a systematic cause, not chance.
- Find the systematic cause. The usual ones are different datums or fixturing, different definitions of size (a two-point caliper reading versus a fitted diameter, which ISO 14405-1 treats as different size characteristics), temperature, probe force on thin walls, and burrs or coating on the measured surface.
- Write down the referee method. Agree in the quality agreement or on the drawing which method settles a borderline part, and which side's traceable calibration it relies on.
When a gage is in the marginal band but is the best available, guard banding keeps the risk on the right side. Per ISO 14253-1, conformance is proven only when the measured value lies inside the tolerance zone reduced by the expanded measurement uncertainty, which in practice means accepting parts only inside tightened internal limits; the method for stating that uncertainty is described in NIST Technical Note 1297. For thread gages and other attribute checks, where disagreement takes a different form, see how two go/no-go gages can disagree on the same hole.
What should the drawing, PO and RFQ say about gage R&R?
Name the characteristics that need a study, the gage to be studied, the basis for %GRR and the acceptance values, so the supplier can quote the work instead of guessing. Leave every other dimension to normal first-article and lot inspection.
A gage study takes time and parts, so asking for one on every dimension of a ballooned drawing mostly buys paperwork. Flag the few characteristics where the tolerance is tight relative to what the gage can resolve, or where the dimension feeds a capability index. The wording below is a starting point; adjust the acceptance values to your own quality manual.
| Item | Example wording |
|---|---|
| Scope | Variable gage R&R required for characteristics flagged on the drawing only, using the gage and fixture that will accept parts in production. |
| Study design | ANOVA method; 10 parts spanning the process range; 3 production appraisers; 3 trials; randomized and blind. |
| Report | %GRR to tolerance (6 sigma basis) and to total variation, %EV, %AV, ndc, gage resolution and calibration due date; raw readings attached. |
| Acceptance | %GRR under 10 % accepted; 10 % to 30 % subject to buyer approval; over 30 % rejected; ndc 5 or more. |
| Attribute checks | Go/no-go gages on flagged features: attribute agreement study with effectiveness, miss rate and false alarm rate reported. |
| Disputes | Disputed parts settled by correlation on serialized parts, then by the referee method named in the quality agreement. |
On Saola's side, the quality page lists what ships with a part: a first-article inspection report covering every toleranced dimension on the first piece from new tooling, and a dimensional inspection record with each production lot. Gage R&R studies are not among the documents that page lists, so if your part needs one, flag the characteristic and the gage on the RFQ and ask for the study to be quoted rather than assuming it is in the package. The same page says that when you send the control dimensions marked on your drawing, the inspection is built around them rather than asking you to adopt the plant's own. The CNC turning page describes thread fit checked with go/no-go gauges before a lot ships, which makes an attribute agreement study the relevant check for those threads, and the CNC milling page says every GD&T callout is measured and logged. ISO 9001 certification is in progress, not held. To ask about a specific part, send the drawing for a quote.
Frequently asked
Is a gage R&R study only needed for automotive parts?
No. The method comes from the automotive quality manuals, but any buyer can ask for it on a characteristic where the gage might be the weak link. Outside a formal PPAP, it is reasonable to ask only for the few characteristics that matter and to accept a simpler report, as described in the guide to PPAP for non-automotive parts.
How many parts, appraisers and trials does a study need?
The typical crossed design is ten parts, three appraisers and two or three trials, which gives sixty to ninety readings. Fewer parts reduce the ability to estimate part variation, and the ndc result becomes unreliable. If only one person or one automated instrument ever measures the part, reproducibility is close to zero and the study reduces to a repeatability study, but say so in the report rather than inventing appraisers.
Can you run gage R&R on go/no-go gages?
Not as a variable study, because there is no measured value. The equivalent is an attribute agreement analysis: several appraisers judge a set of parts, including some near each limit, several times, and the results are compared with a known reference. The AIAG manual treats a kappa above 0.75 as good agreement and grades effectiveness, miss rate and false alarm rate as in the table above.
What is a Type 1 gage study, and is it the same thing?
A Type 1 study measures one reference part many times with one gage and compares the scatter and bias with the tolerance, reported as Cg and Cgk. It checks the instrument on its own, before people and parts are involved, and is common in European supplier requirements. It complements gage R&R; it does not replace it, because it says nothing about appraiser or fixturing effects.
When should a gage study be repeated?
Whenever something in the measurement system changes: a new gage or fixture, a new measuring program, a different site, a revised drawing tolerance, or a run of disputes that suggests drift. Many control plans also repeat studies on a fixed cycle for critical characteristics. A study is a snapshot of one system, not a property of the instrument model.
Sources
- Measurement Systems Analysis (MSA), 4th edition, Automotive Industry Action Group, 2010
- NIST/SEMATECH e-Handbook of Statistical Methods, Gauge R & R studies, NIST, 2012
- Metrological traceability, NIST, accessed October 2026
- Technical Note 1297, Guidelines for Evaluating and Expressing the Uncertainty of NIST Measurement Results, NIST, 1994
- Production Part Approval Process (PPAP), 4th edition, Automotive Industry Action Group, 2006