A continuous glucose monitor does not measure blood glucose. It measures glucose in the fluid between cells a few millimetres under the skin, infers a value from an electrochemical signal, and displays it as though it were a blood reading. Nearly every accuracy complaint traces back to that sentence.
The gap is not a defect. It is a design consequence, and once you know its shape you can read a sensor correctly — which mostly means knowing which readings to trust and which to discard.
The Verdict
What MARD actually promises
MARD — mean absolute relative difference — is the average percentage gap between sensor readings and paired laboratory values across a study. A sensor advertising 9% MARD was, on average, 9% away from the reference.
The word doing the work is "average". A 9% MARD is fully compatible with individual readings 25–30% off, because the metric conceals its own tail. Two further caveats matter for consumer use. Accuracy is not uniform across the glucose range — most sensors perform worse at low values than at normal ones, which is precisely where a reading feels most urgent. And manufacturer figures come from controlled trials with careful insertion and frequent reference draws, not from a sensor that has been slept on for six consecutive nights.
The practical translation: use the sensor to answer "did this meal spike me, and how long did it take to come down" — questions the error band does not touch. Do not use it to answer "is my fasting glucose 97 or 103", because it cannot reliably distinguish those.
Which sensor is the most accurate?
Published MARD figures across current consumer sensors sit in a narrow band, and the differences between products are smaller than the differences technique introduces. Listed alphabetically by manufacturer, then product.
| Sensor | Access | Manufacturer-published accuracy | Sampling | Display constraints |
|---|---|---|---|---|
| Abbott FreeStyle Libre 3 / 3 Plus | Prescription | Abbott publishes 7.9% overall MARD in adults for Libre 3 | Every minute | Full range, with alarms |
| Abbott Libre Rio | Over the counter | Libre-class sensing platform; manufacturer figure published on the product page | Every minute | Reports a 70–200 mg/dL band; outside it shows out-of-range, not a number |
| Abbott Lingo | Over the counter | Libre-class sensing platform; manufacturer figure published on the product page | Every minute | Raw values plus a proprietary spike score; no alarms |
| Dexcom G7 | Prescription | Dexcom publishes 8.2% overall MARD in adults | Every 5 minutes | Full range, with alarms |
| Dexcom Stelo | Over the counter | G7-class sensing platform; manufacturer figure published on the product page | Every 15 minutes | More smoothing than G7; display capped at 250 mg/dL; no alarms |
Two things follow from this table, and the second matters more than the first. The first is that published MARD figures cluster tightly — the gap between the two prescription flagships is a fraction of a percentage point, well inside the error band of any single reading. Verify the current figure for the specific product you are buying on the manufacturer's own page (Abbott, Dexcom), since these are revised with each hardware generation and the over-the-counter lines are newer than the prescription ones.
The second is that the over-the-counter sensors run the same sensing platforms as their prescription siblings — Stelo on G7-class hardware, Lingo and Libre Rio on Libre-class hardware — so the raw measurement is not where they differ. They differ in the display layer, and that difference is larger than any MARD gap. A 250 mg/dL cap or a 70–200 mg/dL reporting band truncates exactly the reading you were trying to characterise, which is a bigger practical accuracy limit than a half-point of MARD. Over-the-counter CGMs covers those constraints in detail.
So the honest answer to "which is most accurate" is that the question is mis-framed. Choose on sampling interval, display range, and whether you need alarms. Then get the technique right, because insertion site, first-day warm-up, and compression move a reading further than the choice of brand does.
The five sources of error, and how to recognise each
Listed alphabetically by error source. Each has a distinct signature, which is what makes them diagnosable rather than mysterious.
| Error source | Cause | What it looks like | What to do |
|---|---|---|---|
| Compression low | Sleeping on the sensor restricts local blood flow | A sharp drop to 50–70 mg/dL that recovers within 30–60 minutes, usually overnight | Ignore any nocturnal low with a vertical drop and vertical recovery. Move the sensor to the other arm. |
| First-day warm-up error | The filament is still equilibrating with the surrounding tissue | Readings in the first 12–24 hours run high or low by 10–20 mg/dL more often than later readings | Discard day one when comparing meals. Start a sensor the evening before you plan to test anything. |
| Interstitial lag | The sensor reads fluid between cells, not blood | A 5–15 minute delay, largest when glucose is changing fastest | Never compare a CGM to a fingerstick during a rise or fall. Compare when the line is flat. |
| Sensor-to-sensor variance | Manufacturing tolerance and insertion depth differ between units | Two sensors worn simultaneously can differ by 10–20 mg/dL | Do not compare absolute values across sensors. Compare each sensor's own averages and patterns. |
| Site and technique | Scar tissue, poor perfusion, or an unapproved site changes the signal | Persistently low or erratic readings for the whole sensor life | Use the manufacturer-approved site — the back of the upper arm for most consumer sensors. |
Why the sensor disagrees with a finger prick
Interstitial lag explains most CGM-versus-fingerstick disagreements, and in nearly all of them neither device is wrong. Glucose has to diffuse from the bloodstream into the interstitial space before the sensor can read it, which takes roughly 5–15 minutes.
That delay is invisible when glucose is stable and maximal when it is changing fastest. Forty-five minutes after a meal, when blood glucose is climbing steeply, a fingerstick can read 150 while the sensor reads 125 — a 17% gap that is entirely explained by timing. An hour later, when the curve flattens, the two converge.
The rule that follows: only compare the two when the trend arrow is flat. Before a meal, or three hours after one, the compartments have equilibrated and a genuine disagreement means something. During a rise or a fall, a disagreement means nothing at all.
CGM vs finger prick vs HbA1c
A CGM, a glucometer and an HbA1c answer three different questions, and each has its own accuracy standard. Note the glucometer row in particular: the ISO 15197:2013 criterion switches from an absolute tolerance to a percentage one at 100 mg/dL, which means fingerstick accuracy is specified most loosely in the low range where a reading feels most urgent — the same weakness CGMs have.
| Criterion | Continuous glucose monitor | Finger-prick glucometer | Laboratory HbA1c |
|---|---|---|---|
| What it measures | Interstitial fluid glucose, inferred electrochemically | Capillary blood glucose, measured directly | Average glycation over ~3 months |
| Typical accuracy | MARD 8–10% against lab reference | 95% of readings within ±15 mg/dL below 100 mg/dL, or ±15% at 100 mg/dL and above (ISO 15197:2013) | Standardised assay, tight precision |
| Best used for | Shape of the curve, timing, and trends | Confirming a single suspicious reading | Confirming the CGM average is roughly right |
| Where it misleads | Lag during rapid change; compression lows | A single point tells you nothing about pattern | Anemia, transfusion, or hemoglobin variants shift it |
| Verdict | Use for patterns, never for a single decisive number | Use to sanity-check, not to track | Use as the annual cross-check on your sensor average |
The cross-check worth running is the first and third columns against each other. A 14-day sensor average of roughly 95 mg/dL corresponds to an HbA1c near 5.0%, and roughly 115 mg/dL to near 5.6%. If your sensor average and your lab HbA1c disagree substantially — and you do not have anemia, a recent transfusion, or a hemoglobin variant, all of which independently shift HbA1c — the more likely explanation is sensor drift than a physiological discovery. That is worth a second sensor before drawing any conclusion from either number.
How to get more accurate data without a better sensor
Most of the accuracy available to a consumer comes from technique rather than from product choice, as the sensor table above shows.
Four habits do the work. Start a sensor the evening before anything you intend to measure, so the warm-up period is spent asleep rather than during your test meal. Use the manufacturer-approved site — the back of the upper arm for most consumer sensors — and avoid scar tissue and areas you lie on. Discard day one from any meal-to-meal comparison. And compare within a sensor, not across sensors: because two units can differ by 10–20 mg/dL, a change you observe when a new sensor starts may belong to the hardware rather than to you.
The one thing worth adding on top is context, not precision. A glucose number describes the outcome of your metabolism; it says nothing about how much insulin was required to produce it. Someone with developing insulin resistance can hold a normal-looking curve for years while insulin output climbs to keep it there. Fasting insulin and HOMA-IR read that effort directly, which is why a blood panel belongs alongside sensor data rather than after it.
Frequently Asked Questions
How accurate are continuous glucose monitors?
Modern consumer sensors run a mean absolute relative difference of roughly 8–10% against laboratory reference values, which is good for tracking patterns and imprecise for any single decision. In practical terms, a displayed 100 mg/dL means the true value is most likely between 90 and 110. That is accurate enough to see that a meal spiked you and not accurate enough to distinguish a 104 from a 110. Treat the shape of the curve as the signal and any individual number as approximate.
What is MARD and why is it a weak guarantee?
MARD is the mean absolute relative difference between sensor readings and paired laboratory values, expressed as a percentage. It is an average, which is exactly its weakness: a sensor with 9% MARD can still produce individual readings 25–30% off, because the average conceals the tail. MARD also varies with glucose level — most sensors are less accurate at low values than at normal ones — and manufacturer figures come from controlled studies in optimal conditions, not from a sensor that has been slept on for six nights.
Why does my CGM disagree with my finger prick?
Usually because of interstitial lag, and usually neither reading is wrong. A CGM measures glucose in the fluid between cells, which trails blood glucose by roughly 5–15 minutes. During a sharp rise after a meal, the fingerstick is already at 150 while the sensor still reads 125 — both are correct measurements of different compartments at different points on the same curve. The only fair comparison is when glucose is stable: before a meal, or several hours after one, with the trend arrow flat.
Can a continuous glucose monitor give false readings?
Yes, and the most common false reading has a specific cause: compression. Sleeping on the sensor restricts blood flow to the tissue around the filament, and glucose in that pocket falls locally while blood glucose stays normal. The result is a dramatic overnight low — often 50–70 mg/dL — that recovers as soon as you roll over. The signature is a near-vertical drop and near-vertical recovery, usually between 1am and 5am. It is a local artefact, not hypoglycemia, and it is the single most common reason people panic about a sensor.
Which continuous glucose monitor is the most accurate?
Published MARD figures cluster tightly across current sensors — Abbott publishes 7.9% for FreeStyle Libre 3 and Dexcom publishes 8.2% for G7, a gap well inside the error band of any single reading. Chasing the lowest advertised figure is therefore not a useful buying criterion, because the differences between models are smaller than the differences introduced by insertion site, first-day warm-up, and compression. What varies meaningfully is the display layer: Stelo caps its display at 250 mg/dL and smooths more than G7, and Libre Rio reports only within a 70–200 mg/dL band. Those constraints limit a reading more than a half-point of MARD does.
Why is my CGM reading low when I feel fine?
Three explanations cover almost all cases. Compression from lying on the sensor is the most common, and shows the vertical-drop-and-recovery signature. The second is a first-day sensor still equilibrating, which can read 10–20 mg/dL off in either direction for the first 12–24 hours. The third is that consumer sensors are less accurate at low glucose values than at normal ones, so the reading has a wider error band exactly where it looks most alarming. If you feel well and the drop was abrupt, confirm with a fingerstick before acting on it.
How do I check whether my sensor is roughly right?
Compare your 14-day sensor average against a laboratory HbA1c. The two answer overlapping questions — average glucose over a period — so a large disagreement suggests sensor drift rather than a physiological finding. A 14-day average of 95 mg/dL corresponds roughly to an HbA1c near 5.0%, and 115 mg/dL to roughly 5.6%. Wide divergence between the two, in someone without anemia or a hemoglobin variant, is worth a second sensor before drawing conclusions from either.
Does sensor accuracy change over the wear period?
Yes, at both ends. The first 12–24 hours are the least reliable while the filament equilibrates with the surrounding tissue, and many users see a systematic offset that resolves on day two. Accuracy is generally most stable in the middle of the wear period. Toward the end of a 14–15 day sensor, some units drift as the local tissue response to the filament develops. The practical rule is to discard day one from any comparison, and to be sceptical of a dramatic change that appears only in the last day of a sensor's life.
Related
- CGM platforms compared — Levels, Nutrisense and Signos on one rubric
- Over-the-counter CGMs — Libre Rio, Lingo and Stelo
- Levels review
- Nutrisense review
- HOMA-IR — what a glucose curve cannot show you