Key points

  • Accuracy: the difference between measured gaze and actual gaze. It sets the size of the targets you can tell apart.
  • Precision: the stability of the measurement when the eye is not moving. It determines how well fixations are detected.
  • Sampling rate: the number of samples per second. It matters most for saccades.
  • Latency and data loss: two figures often missing from data sheets, and decisive for some protocols.
  • Figures from two manufacturers are not measured under the same conditions: they give an order of magnitude, not a ranking.

Accuracy: is gaze in the right place?

Accuracy is the mean difference, in degrees of visual angle, between the point the participant is really looking at and the point the device reports. It is measured by having participants fixate targets whose position is known.

A degree means nothing until it is related to a distance. Here is what it represents.

AccuracyOn a screen at 60 cmIn a scene at 1 m
0.3°0.3 cm0.5 cm
0.5°0.5 cm0.9 cm
0.6°0.6 cm1.0 cm
1.3°1.4 cm2.3 cm
1.8°1.9 cm3.1 cm
2.5°2.6 cm4.4 cm

Calculation: offset = distance × tangent of the angle.

The practical rule: your areas of interest should be clearly larger than this offset, and separated by at least as much. To tell two neighbouring words apart, you need a screen-based device. To know whether a driver is looking at the road or at their phone, 2° is enough.

Precision: is the measurement stable?

When the eye fixates a point, successive samples do not fall in exactly the same place. Precision describes this dispersion. It determines how well the software can recognise a fixation and measure its duration.

It is expressed in three ways, which do not give the same figure for the same signal:

  • the root mean square of sample-to-sample differences (RMS);
  • the standard deviation of positions;
  • the area of an ellipse that contains the samples.

In an example cited by the 2023 international guidelines, the same recording gave 0.30° as root mean square and 0.42° as standard deviation. A data sheet that states a precision without saying which one cannot be compared with another.

Filtering also changes the figure: Tobii states 0.2° for the Pro Fusion under optimal conditions, and 0.04° after filtering.

Sampling rate: how many samples per second?

At 60 Hz, the device gives one sample every 16.7 ms; at 250 Hz, every 4 ms. Sampling rate describes the time step, not spatial fineness.

What you measureSensitivity to sampling rate
Areas looked at, scan pathsLow
Fixation durationModerate
Saccade onset and offset, reaction timeHigh
Saccade peak velocityVery high
Gaze-contingent displayHigh, together with latency

A 2017 modelling study put numbers on it: measured peak velocity decreases as the sampling rate drops, and the effect is stronger the smaller the saccade.

Latency: with what delay?

Latency is the time between the moment light reaches the camera and the moment the device delivers the gaze position. It does not matter for analysis done afterwards. It becomes decisive in two cases:

  • gaze-contingent display, where the image changes according to where the participant looks: the screen's latency adds to the eye tracker's;
  • combination with other measurements, such as EEG, where each device has its own delay.

Some data sheets give it in milliseconds, others in number of frames. Three frames at 250 Hz is 12 ms.

Data loss: how many samples are missing?

A 250 Hz eye tracker should deliver 250 positions per second. If it delivers only 225 usable ones, the loss is 10%. Blinks explain part of it; the rest comes from tracking itself: glasses, make-up, drooping eyelids, head movements.

This figure almost never appears on a data sheet, because it depends on the participants. That is one more reason to test the device on your population. Two useful values to ask for: the time the device takes to recover gaze after a blink, and after a loss.

Why the figures compare poorly

What variesExample
Measurement conditionsEyeLogic states a “typical value”, Tobii a value “under optimal conditions”
CalibrationNeon states 1.8° uncalibrated; Tobii Pro Glasses 3, 0.6° after a one-point calibration
Filtering0.2° or 0.04° for the same device, without or with a filter
The precision measureRoot mean square, standard deviation or ellipse
The participantPublished values are obtained on participants who are easy to track
How the device is wornGlasses that slip can add an error

On this last point, a 2020 study tested four pairs of glasses while participants talked, pulled faces or moved the frame. Two systems stayed stable; the others showed additional errors of 0.8° to 3.1°. The authors recommend that each team run this test with its own equipment.

Six questions to ask the supplier

  1. Is the accuracy a typical value or a value under optimal conditions?
  2. Is the precision a root mean square or a standard deviation, and on filtered data?
  3. On how many participants, and which ones, were these values measured?
  4. What is the latency, in milliseconds, at the sampling rate I will use?
  5. How long does it take to recover gaze after a blink or a loss?
  6. Can I test the device on my participants, with my stimuli?

What to report in a paper

The 2023 international guidelines list the minimum to report in any eye tracking study. Among the twelve items common to all studies:

  • the manufacturer and model, and software versions;
  • the sampling rate;
  • constraints on head movement;
  • the calibration procedure and the validation error;
  • measurement uncertainty, including precision;
  • the data processing steps;
  • data loss.

Latency is added for studies of gaze dynamics, and screen distance for screen-based studies.

Frequently asked questions

What is the difference between the accuracy and the precision of an eye tracker?

Accuracy is the mean difference between the point actually looked at and the point measured. Precision is the dispersion of samples when the eye fixates a single point. A device can be precise without being accurate, and vice versa.

What does 0.5° of accuracy represent on a screen?

At 60 cm from the screen, 0.5° corresponds to about 0.5 cm. At the same distance, 1° corresponds to about 1 cm.

Is a high sampling rate always necessary?

No. Fixation duration is measured correctly at modest sampling rates. Saccade peak velocity, however, is underestimated at low rates, especially for small saccades.

Can the accuracies stated by two manufacturers be compared?

With caution. One states a “typical value”, another a value “under optimal conditions”, and some precision figures are given after filtering. Ask for the measurement protocol and test the device on your participants.

How is precision expressed?

Most often as the root mean square of sample-to-sample differences (RMS), as the standard deviation of positions, or as the area of a dispersion ellipse. These three measures do not give the same figure for the same signal.

Sources

The scientific publications come from PubMed.

  • Dunn MJ, Alexander RG, Amiebenomo OM, et al. Minimal reporting guideline for research involving eye tracking (2023 edition). Behav Res Methods. 2024;56(5):4351-4357. doi:10.3758/s13428-023-02187-1
  • Mack DJ, Belfanti S, Schwarz U. The effect of sampling rate and lowpass filters on saccades – A modeling approach. Behav Res Methods. 2017;49(6):2146-2162. doi:10.3758/s13428-016-0848-4
  • Niehorster DC, Santini T, Hessels RS, Hooge ITC, Kasneci E, Nyström M. The impact of slippage on the data quality of head-worn eye trackers. Behav Res Methods. 2020;52(3):1140-1160. doi:10.3758/s13428-019-01307-0
  • EyeLogic, LogicOne and Lite product pages, consulted on 19 November 2026: eyelogicsolutions.com
  • Pupil Labs, Neon specifications, page consulted on 19 November 2026: pupil-labs.com
  • Tobii, comparisons of screen-based eye trackers and glasses, pages consulted on 19 November 2026: tobii.com

This article is general information for research teams. It is based on the manufacturers' pages and the publications cited. Gaze Intelligence distributes EyeLogic and Pupil Labs eye trackers and supports SMI systems: this article is written by a distributor. The trademarks mentioned belong to their owners.