Labvanced Eye Tracking 2.0: accurate webcam eye tracking for all online research

Caspar Goeke, Head of Labvanced Research · Published October 2026

Labvanced Eye Tracking 2.0 tracks gaze with a standard webcam at a median accuracy of 1.8 degrees of visual angle (5.2 % of the screen diagonal) in unsupervised online participants. The result comes from 494 Prolific participants on their own computers, and the anonymized data is published with this note.

After more than a year of development, Eye Tracking 2.0 brings the accuracy that webcam eye tracking reached in the laboratory to real online studies.

Key findings

• Median accuracy is 1.8° (5.2 % of the screen diagonal) with a two-minute calibration and balanced screening.
• With the long calibration and very strict screening, median accuracy reaches 1.4° (4 %).
• Gaze is steadier than in the first generation: sample-to-sample noise is about 40 % lower than that of Eye Tracking 1.x recorded in the laboratory.
• 99.3 % of paid participants delivered a complete dataset, and balanced screening raises the recruitment cost by about 10 %.

Outline

• 1Abstract and introduction
• 2How we measured: the task, the metric and the participants
• 3An improved gaze engine with dual-eye fusion that works on laptops, tablets and phones
• 4A shorter and better calibration used for individual neural network fine-tuning
• 5A better screening method, technical pre-checks, and a simplified setup
• 6Live corrections for head movements and lighting changes
• 71.4 to 1.8 visual degrees of accuracy, verified on about 500 real participants on Prolific
• 8A passive tier that works out of the box without any calibration, great for infant research
• 9Discussion: which setup for which study?
• 10Try it yourself
• 11Data and reproducibility

Abstract

Webcam eye tracking has already become a powerful tool for online behavioral research, and the first generation of Labvanced's system was validated against an EyeLink 1000 in the laboratory at an accuracy of about 1.4° of visual angle (Kaduk et al., 2023). Several questions remained open. How much of that accuracy survives the uncontrolled conditions of a real online study? Why do some participants deliver far worse data than others, and can they be identified before a researcher pays for them? Can the calibration be shortened to make it more practical in ordinary online studies? And what should happen to participants who do not pass it, or otherwise do not qualify? Here we show how we improved eye-tracking accuracy, while simultaneously shortening the calibration, simplifying the setup, and optimizing the whole procedure for real online data collection with our partner Prolific. We then validated the system in seven recording batches with about 500 completed participants of a 45-target accuracy task, and publish their anonymized data so that every result can be verified. With our new standard calibration (about 2 minutes) and balanced screening, the median participant's error was 5.2 % of the screen diagonal, 1.8° of visual angle on a laptop at 60 cm. And with the 5 minute calibration and very strict screening, the median accuracy of online participants reached 4 % of the screen diagonal, which is equal to 1.4° of visual angle at a standard distance. In summary, with this release of Eye Tracking 2.0, we show that the accuracy reported in our peer-reviewed laboratory comparison can be replicated online.

1. Introduction

Eye movements reveal attention continuously and without asking, and for decades measuring them required laboratory equipment. With our existing first-generation webcam system (referred to here as Eye Tracking 1.x) we showed that a webcam can come very close in accurately predicting eye movements: Recorded simultaneously with an EyeLink 1000 in five standardized tasks, our system reached an accuracy of 1.4° and a precision of 1.1°, about half a degree behind the lab "gold standard" (Kaduk, Goeke, Finger & König, 2023). Since then, independent groups have confirmed the picture and put the system to work across a remarkable range of fields:

• Reading. Recorded alongside an EyeLink, Labvanced webcam gaze correlated with the laboratory tracker at a median r of 0.81 and reproduced the word-frequency effect (Serrano-Carot, Angele, Xu & Vasilev, 2025).
• Infant development. Infants' webcam gaze recorded at home was compared with laboratory eye tracking in an audio-visual synchrony task (Bánki, de Eccher, Falschlehner, Hoehl & Markova, 2022).
• Decision making and human-computer interaction. Remote gaze revealed early decision processing (Bertrand, Ouellette Zuk & Chapman, 2023), the coordination of eye and cursor in a digital object task (Bertrand & Chapman, 2023) and users' frustration (Stone & Chapman, 2023), brought together in a doctoral thesis on the utility of webcam eye tracking (Bertrand, 2023).
• Language, attention and perception. Studies tracked how learners attend to text, audio and pictures when acquiring second-language vocabulary (Zhang & Zhang, 2025), how words bias visual attention (Calignano, Lorenzoni, Semeraro & Navarrete, 2024), how partial information shapes dynamic face perception (Alp, Lale, Saglam & Sayim, 2024) and how video-game players detect change (Haverly, 2022).
• Education and science communication. Gaze synchrony between students watching lecture videos predicted their test performance (Sauter, Hirzle, Wagner, Hummel, Rukzio & Huckauf, 2022), and fixations showed how online advertisements draw attention away from scientific explanations (Tsutsuse, 2025).
• Marketing, social and environmental psychology. Instagram users' gaze showed that they seek out an influencer's follower count (Wies, Bleier & Edeling, 2023); attention to résumés explained how 1,159 participants judged discriminatory hiring (Zhang & Powdthavee, 2026); and gaze on urban scenes was related to emotional experience (Sander, Mazumder, Fingerhut, Parada, Koselevs & Gramann, 2024).
• Health and human-AI interaction. A randomized controlled trial measured where pharmacists look when an AI assists them in verifying medication (Tsai et al., 2025; Kim et al., 2026), and a clinical study recorded webcam gaze with EEG in children with ADHD (Zhao et al., 2025).

Methodological reviews and primers now list Labvanced among the established webcam eye-tracking systems (Niehorster et al., 2025; Jenke & Sullivan, 2025; Patterson, Nicklin & Vitta, 2025; Tsuji, Amso, Cusack, Kirkham & Oakes, 2022). Those results established webcam eye tracking, and Labvanced in particular, as a serious instrument for academic research.

Importantly, to compare our webcam-based data with the EyeLink, our previous comparison project, which resulted in our peer-reviewed publication, had to be conducted inside a lab. Hence we could control the entire environment including the computer, the screen, the webcam, the room, and the lighting. We could also instruct participants in person, and being in such a laboratory setting, most participants arguably gave their best or at least a good effort when doing the experiment. Almost none of this is ensured in an unsupervised environment with real online participants. An online participant sits where they sit, on the device they own, in the light they have, and some just don't pay attention whatsoever. Given these known shortcomings, we are both happy and proud that many projects with our existing eye-tracking technology were remarkably successful and were published in leading journals. However, some researchers did not get the accuracy they expected or faced other difficulties, such as participants being frustrated by a long calibration, or a high number of participants who failed for technical and various other reasons. For example, even though we had already built infant-friendly calibrations, two infant studies found our first-generation eye tracking not yet reliable enough for their youngest participants (Grosse Wiesmann et al., 2025; Golan, Abo Shkara & Havron, 2023).

In 2025 we started to evaluate these shortcomings systematically and to address them one by one. At the end of this journey, though surely not at the end of webcam eye-tracking development, stands Eye Tracking 2.0. This lab note describes what changed and why, how we validated it, and what accuracy you can now expect with real online participants. The data behind all of this is open access and available for everyone to independently replicate and verify (Section 11). We are also planning a new peer-reviewed study, but as that takes time, we wanted to share these results now.

2. How we measured

All participant data in this note were collected over a period of about one month, within more or less the same study paradigm. However, the eye-tracking version changed between almost every one of these recording batches, to test or verify our latest developments (e.g. testing pose correction). Some of the batches are also smaller and some larger, for various reasons. This note is not a peer-reviewed publication, but we believe that all analyses and results reported here are scientifically rigorous, genuinely insightful, and everyone is welcome to download the data and replicate or verify them. A peer-reviewed article on Eye Tracking 2.0, and potentially a few other topics, will follow.

Task. Prolific participants ran through the technical feasibility check (see Section 5) and completed the study's calibration; the admitted participants then did a short training block and finally a 45-target accuracy task: a 9 by 5 grid of positions covering the whole screen, presented one at a time in random order. The participants' task was to fixate the target (a plain circle) and click it as soon as its color changed. If a participant clicked too early or too late, the trial was repeated. You can run this task yourself (Section 10).

Metric. The main metric we are interested in is the target error or accuracy. This is defined as the distance between the target location and the median gaze position over the last 1000 ms before the target changed. We also implemented a new fixation detection algorithm and used the centroid of the fixation at the change; both measures (the gaze median and the fixation) yielded almost identical accuracy results. More information about this new fixation detection mechanism will be available in the next lab note. A participant's accuracy is then defined as the median of their 45 errors, and a group's accuracy as the median over participants. We report this median throughout and filter neither trials nor participants. On a 1920 by 1080 screen (full HD), 1 % of the diagonal is 22 px and 5 % is 110 px. Because of the remote recording setup, we could not reliably record either the true physical screen size or the participant's exact viewing distance. The error in visual angle is therefore computed for an assumed reference screen: a 14.2-inch laptop (e.g. a 14-inch MacBook Pro), the middle of the 13 to 16 inch range typical online participants use, viewed at a typical distance of 60 cm.

Sample. Ten recording batches between 25 August and 28 September 2026 produced about 700 completed sessions. The results in this note are based on the 494 completed sessions of the seven most recent batches: 48 of a 2.0.9 batch, 296 on the High tier without screening, 50 on the High tier with balanced screening and the live corrections running, and 100 on the Highest tier. The roughly 200 sessions of the earliest batches shaped the system during earlier development and hence are not part of the reported results.

Data. The anonymized data of these 494 sessions, together with the scripts that compute every number and figure in this note, can be downloaded (see Section 11).

3. A dual-eye gaze engine

3.1 How the dual-eye engine works

How does Labvanced Eye Tracking 2.0 predict gaze and fixations?

Figure 1. The processing chain. Everything runs on the participant's device (edge inference); no image or video leaves the device.

What does the dual-eye fusion model refer to? The 2.0 gaze engine treats the two eyes as two separate measurements and combines them, so an eye that is half closed, in shadow or caught in a reflection counts less. Information from neighboring frames is used as well, without smoothing away real eye movements. The result is a more accurate and more robust raw gaze estimate than before, as well as a confidence value that is more indicative of actual precision and accuracy.

Speed and devices. The gaze engine was rewritten for speed and makes use of the most recent web technologies such as WebGPU and WebAssembly. Taken together this considerably speeds up both the live prediction and the fine-tuning step (calibration). And as a result, Eye Tracking 2.0 runs smoothly on most laptops (apart from older Chromebooks and similar devices), on desktop computers and on modern tablets such as iPads. Phones use a shortened calibration of about 15 seconds, and in our own phone measurements the gaze was quite accurate. A field validation on tablets and phones will follow separately.

3.2 Result: 40 % less sample noise than 1.x

Precision. With dual-eye fusion, gaze is markedly more stable and precise. Online, 2.0 gaze samples scatter by a median of 2.2 % of the screen diagonal (about 0.75°), and their sample-to-sample noise is about 40 % lower than that of Eye Tracking 1.x recorded in our 2023 laboratory study (Section 7.2).

4. A shorter and better calibration

4.1 How the calibration works

Calibration in 2.0 is significantly shorter and works in a fundamentally new way. The quality of each calibration point is checked while it is recorded, so points taken while the participant blinked or looked away are repeated on the spot. The participant's network is then fine-tuned on their own device, in most cases considerably faster than before, and the fine-tuning is limited in time, so no participant can ever get stuck.

Gaze Control. Gaze Control is a short docking task that runs before the first eye-tracking task, after any repositioning of the head, and at regular intervals set by the accuracy tier: every three minutes on High, every minute on Highest. It never interrupts a trial. Participants see their head position as a circle and move their head until it overlaps a target circle, which corrects head position and distance to the camera in a single intuitive step. Real online participants take a median of 6.7 seconds from start to end. Each run also serves as a check of the current gaze quality, which the live corrections use (Section 6).

Probe. The first Gaze Control run of a session is slightly longer, about 17 seconds in total, and serves as a probe: it predicts how accurate a participant's gaze will be over the rest of the session. This is the basis of the screening levels (Section 5): at higher levels, only participants with better predicted accuracy are admitted to the study.

4.2 Result: 4 minutes on High, 7 minutes on Highest

Instructions + calibration takes 4 minutes on High and 7 on Highest. From seeing the first instruction text to the end of the probe, real online participants needed a median of about 4 minutes on the High tier (about 2 minutes of calibration points) and about 7 minutes on the Highest tier (about 5 minutes of calibration points). The fine-tuning itself took a median of about 37 seconds after the two-minute calibration and about 67 seconds after the Highest one. Note, subjects can mostly relax during the fine-tuning step.

Two minutes of calibration are enough for most studies, but five will improve it. With balanced screening (level 3), the two-minute calibration (High) delivers a median of 5.2 % of the diagonal (about 1.8 visual degrees). For the majority of online paradigms this will be accurate enough. The Highest calibration with very strict screening (level 5) delivered about 4 % (Figure 2), which is about 1.4 visual degrees.

5. Better screening, technical pre-checks and a simpler setup

5.1 How pre-checks and screening work

Screening happens in two stages, supported by new functionality from our partner and participant recruitment provider, Prolific. First, a technical check rejects unsuitable devices without payment; then come the calibration, the probe and the screen-out, with a reduced participant fee of the researcher's choosing.

• Technical rejection. The device's capabilities and its camera are checked, and the gaze sampling rate must reach 10 Hz. A failure here results in a technical rejection: it takes the participant only a few seconds (maybe up to a minute on really slow devices), and they are told why they are rejected. The best way to communicate this to participants coming from Prolific is to create a rejection link, select "incompatible device" as the reason, and tick the checkbox "ask participant to return the study". Most participants will happily return it, and for those who still submit for the full amount, you will see their status as "incompatible device", so the rejection reason is straightforward. As a result, only participants with technically capable devices enter the actual study and are allowed into the calibration.

• Screen-out after calibration. After the calibration and the probe, the system decides whether the participant is allowed into the eye-tracking task, based on the study's screening level. Level 0 does not screen anyone out and allows everyone to start the task. Level 1 screens out only failed calibrations, while levels 2 to 5 use the accuracy predicted by the probe and admit about 65, 50, 30 or 15 of 100 participants. A participant who does not pass is screened out on the spot. If you use Prolific, we recommend using a dedicated screen-out link and paying participants 0.50 GBP for the two-minute and 0.80 GBP for the five-minute calibration (screen-out links are partially paid on Prolific). These amounts apply when the eye-tracking calibration is more or less the first thing the participant does. If the eye-tracking part starts much later in your study, consider enabling a calibration retry (off by default) and, for participants who fail the retry as well, paying a higher screen-out fee.

Taken together, the technical rejection and the screen-out make eye tracking with our 2.0 system a smooth experience on crowdsourcing platforms like Prolific, with both high accuracy and high completion rates for the participants who are admitted.

5.2 Result: 99.3 % complete datasets, about 10 % extra cost

The pre-check catches what would have failed later. In one of our online data collections, 37 of 251 entrants (about 15 %) were turned away before the start, most for a camera below the required resolution or a computer too slow to track a face (Figure 3). Each of them would otherwise have produced a failed or unusable session. We now also enforce a measured gaze rate of at least 10 Hz, which our new fixation algorithm needs (more on that in the next Lab Note).

Each screening level buys a predictable gain. Applied to the 295 participants of the High tier without screening, each level trades admission for accuracy in steps of 0.2 to 0.7 points (Figure 5). At level 3, balanced, about half of all entrants are admitted, the median of admitted participants falls by about a point, and the share of admitted participants worse than 10 % falls from 17 % to 3 %. A later live batch confirmed this exactly as designed: at balanced screening, half of the participants were admitted, and they delivered 5.24 % (Section 7.1).

Screening costs little because screen-outs are only partially paid. A screened-out participant receives a small part of the full fee, about a tenth in our case, so level 3 raises the estimated recruitment cost by about 10 % and level 5 by about 60 % (Figure 6).

Almost every paid participant delivers a complete dataset. Of the 450 participants approved and paid on Prolific in the last five batches (24 to 28 September), 447 (99.3 %) delivered the full dataset (Figure 7). Just as important, participants stay once they are admitted: 96 to 97 % of those who passed the probe went on to complete the study. Participants turned away at the technical pre-check are not paid and screened-out participants receive only a small part of the fee, so the full fee goes almost exclusively to usable recordings.

6. Live corrections for head movement and lighting changes

6.1 What the live corrections do

Screening decides who enters the study. For the admitted participants, two live corrections then reduce the error that head movement and changing light add over the course of a session. Both are on by default on every calibrating tier, but can be switched off in the study settings.

• Head-pose correction. When the head moves away from its position during calibration, the gaze estimate shifts with it. This correction compensates for that shift continuously.
• Gaze-check correction. The regular Gaze Control checks (every three minutes on High, every minute on Highest) are used to remove the drift that builds up after calibration, including the drift a change of light produces. A lighting-change detector triggers an extra check when the light on the face changes noticeably.

Raw gaze stays reconstructable: the corrections are applied by default but can be turned off, and the correction applied to each gaze sample can also be recorded in Labvanced (as parameters of the gaze trigger event), so researchers who prefer the uncorrected signal can recover it exactly.

6.2 Result: a sixth less error, most of the tail removed

The corrections remove a sixth of the error and most of the tail. On the High tier with balanced screening (level 3, 143 participants) the median moves from 6.37 % raw to 5.72 % with the head-pose correction and 5.28 % with both, and the share of participants worse than 10 % falls from 15 % to 3 % (Figure 8). Without any screening the gain is the same, a sixth: 7.42 % to 6.21 % on 295 participants. About three participants in four improve (Figure 9).

They hold on participants they were never developed on. The corrections were set once and then applied unchanged to five later batches, three offline and two with the corrections running live in the study (Figure 10).

Stable light keeps accuracy high, and Gaze Control checks catch the changes. On the Highest tier, 70 of 99 participants kept the brightness of their face within 5 % of the calibration, and their median raw error was 5.49 %. For the 13 participants whose light changed by more than 10 %, it was 10.94 % when the face became darker (5 participants) and 7.78 % when it became brighter (8 participants) (Figure 11). Regular Gaze Control checks catch such a change and correct for it.

7. Validated accuracy of 1.4 to 1.8° with Prolific participants

7.1 Accuracy: 1.8° on High, 1.4° on Highest

With the two-minute calibration and balanced screening, the median participant's error is 5.2 % of the diagonal, about 1.8°, and fifteen of sixteen admitted participants stay below 10 %. The Highest tier with very strict screening reaches 4 %, about 1.4°.

Error is nearly uniform across the screen, and the outer targets are not worse than the center (Figure 13). The calculator below converts the expected accuracy to any screen and viewing distance.

7.2 Precision: 0.75° scatter, 0.6° sample-to-sample

Precision improved markedly over the first generation. To compare like with like, we recomputed the precision of Eye Tracking 1.x on the recordings of our 2023 laboratory study, where every participant was recorded with the webcam and an EyeLink 1000 at the same time. For each grid trial we took the steadiest one-second window according to the EyeLink, so that both trackers were measured during a genuine fixation, and applied the same two standard measures as for 2.0: the scatter of samples around their mean (SD) and the root mean square of sample-to-sample distances (RMS-S2S).

Values in percent of the screen diagonal, with visual angle on each study's own screen (15-inch monitor at 60 cm in the laboratory, 14.2-inch laptop at 60 cm online). Both webcam systems sampled at about 30 Hz.

The online 2.0 signal scatters about 10 % less than the laboratory 1.x signal, and its sample-to-sample noise is about 40 % lower, from about 1.0° to 0.6°. Notably, the comparison favors 1.x on two counts: it was recorded under laboratory control, and its windows were chosen as the steadiest second by the EyeLink, which is not possible online. The improvement is therefore, if anything, underestimated. The EyeLink of course remains the most precise system, but 2.0 closes a good part of the gap.

8. A passive tier without calibration

8.1 How the passive tier works

With Eye Tracking 2.0 it is possible to run a passive model without calibration, which can be very useful in certain experimental scenarios. All it requires is that participants grant camera access; the eye-tracking prediction then starts immediately. It is made for participants who cannot or should not calibrate, infants first among them: calibration has always been the hard part of infant eye tracking, and many infant studies use so-called preferential-looking paradigms, i.e. they ask whether a child looks left or right. Being able to classify this reliably, in real time and entirely on the participant's own machine, can be a game changer for infant attention research, but also for other domains where eye tracking is best done silently in the background. You can run an infant study yourself in Section 10.

8.2 Expected accuracy: about 4° (estimate)

The passive tier was not part of the Prolific batches, and its field validation is still to come. Infant labs will rightly want that validation with actual infants, which goes beyond what Prolific offers. If you work in or represent an infant lab and would like to help validate the passive tier, please reach out: we are very interested in collaborations.

That said, from our extensive internal tests, our working estimate is an error of about 11-12 % of the screen diagonal, about 4° on a laptop at 60 cm. Figure 14 shows what that means for a typical experimental design: at the passive tier's median error, a look at a target in the middle of the left half of the screen stays well clear of the right half. There is even a good chance that 4 to 6 screen regions can be distinguished with enough confidence. More data on this will be provided over time.

Figure 14. The median error of each tier drawn to scale as a circle around a target in the left half of a 16:9 screen. The dashed line splits the screen into left and right. The passive figure is an estimate; the other two are measured.

9. Discussion

Our peer-reviewed 2023 comparison showed that a webcam in a laboratory can track gaze to about 1.4°. This note shows that with our Webcam Eye Tracking 2.0 engine, a webcam in a participant's home, on their own laptop and in their own light, tracks gaze to a median of about 1.8° with a two-minute calibration and balanced screening, and to about 1.4° with the long calibration and very strict screening. In effect, the accuracy the previous generation reached under laboratory control is now reached online. Three mechanisms account for most of the change: an improved dual-eye gaze engine, a calibration that better captures individual differences and drift, and live corrections that remove the systematic errors of head pose and lighting.

For researchers the consequences are very practical. Most studies in domains such as psychology, linguistics, sociology, economics, marketing and others, can run the High tier and expect about 1.8°, which will most likely be good enough to capture the expected effect. Studies that investigate smaller regions, such as reading tasks with many words and small gaps between them (try the reading study), can use the Highest tier and get about 1.5° with strict screening, or 1.4° with very strict screening, from a webcam. At the other end, researchers who study attention in infants and toddlers now have the option to track it without any calibration, and so without the difficulties calibration brings to data collection (Section 8). And of course there are several steps in between these extremes. The setup in the Labvanced interface is much simplified, and the expected accuracy it shows for each choice is based on real measurements. In short, your next eye-tracking project can start today.

Which setup for which study?

The table below is a rough guideline, not a strict rule set. For each task and study, look at the number of areas of interest, their size and, above all, the distance between them, and choose the settings accordingly: two large regions far apart tolerate far more error than many small ones side by side. If you are not sure which settings fit your design, reach out to us and we will look at it with you.

Expected accuracies are the median of admitted participants with the corrections on, measured on the High tier and the Highest row and estimated for the others; the extra recruitment cost counts a screened-out participant at a tenth of a completed one. Screening does not apply to the Passive and Medium tiers: everyone is admitted.

The setup wizard in the Labvanced study settings asks the same two questions, the accuracy tier and the screening level, and shows the expected accuracy of your choice before the study starts.

Frequently asked questions

How accurate is webcam eye tracking?

With Labvanced Eye Tracking 2.0, the median online participant is accurate to 1.8° of visual angle after a two-minute calibration with balanced screening, and to about 1.4° after the long calibration with very strict screening. In the laboratory, the first generation reached 1.4° in a direct comparison with an EyeLink 1000 (Kaduk et al., 2023).

Is webcam eye tracking accurate enough for reading studies?

For reading at word level we recommend the Highest tier with strict screening, which delivers about 1.5°. Larger areas of interest, such as faces, scenes or regions of a page, work on the High tier at 1.8 to 1.9°. The table above lists which setup fits which study design.

How long does the calibration take?

Instructions and calibration take about 4 minutes on the High tier and about 7 minutes on the Highest tier. The shorter tiers take between 30 seconds and 3 minutes, and the passive tier needs no calibration at all.

Does webcam eye tracking work with infants?

The passive tier starts tracking as soon as the camera is on, without any calibration. Its estimated error is about 4°, which is enough to tell a look to the left from a look to the right in preferential-looking studies. A field validation with infants is still to come.

How does webcam eye tracking compare with a laboratory eye tracker?

A laboratory tracker such as the EyeLink 1000 remains more precise, with a scatter of about 0.1° against 0.75° for the webcam. Webcam eye tracking reaches an accuracy of about 1.4 to 1.8° on participants' own computers, which is sufficient for most designs based on areas of interest.

Can I switch my existing study to 2.0?

Yes, at any time: choose Eye Tracking 2.0 in the study settings. Existing studies stay on 1.x until you switch them, and new studies start on 2.0.

Is it included in my license?

Eye Tracking 2.0 is part of all Premium and Ultimate licenses, just as Eye Tracking 1.x was before.

What does a 2.0 study record?

The same gaze output as before, x and y positions, now more accurate and more precise, with a confidence value per sample. Fixations are detected with the new merged detector by default. If you want the uncorrected signal, the correction applied to each sample can be recorded as parameters of the gaze trigger. Participants who are screened out or technically rejected are listed with their status in the study's Group Statistics.

Does anything change in my data?

The structure stays the same. The confidence value is now on a different scale: it is computed from completely different quantities than in 1.x, so its numbers are generally lower. This is a change of scale only, not of data quality, but it means that confidence thresholds from 1.x studies do not carry over to 2.0. Confidence is also participant-specific: use it to compare samples or trials within one participant, not across participants. A confidence of 0.3 for one participant does not mean the same as 0.3 for another. Our data shows this clearly: participants' typical confidence ranged from 0.09 to 0.73, yet it said nothing about how accurate they were. The participants whose typical confidence was around 0.3 delivered anything from 3.7 % to 20.8 % error. Within a participant, by contrast, the trials with lower confidence tended to be less accurate.

What comes next?

Labvanced Webcam Eye Tracking 2.0 was a project of several months that touched almost every part of our eye-tracking code. It is certainly not the last improvement we will make. Here are five topics we are working on now.

1. Blink detection

Blinks briefly disturb the gaze estimate. We will detect them reliably and hold the gaze steady through them, so that a blink no longer looks like an eye movement. A clean blink signal is also interesting in its own right: blink rate and timing are established measures of fatigue, arousal and cognitive load, and could become a new output of every eye-tracking study.

2. Even accuracy across the screen

Accuracy is already nearly uniform across the screen (Figure 13). We are working on removing the small systematic differences that remain between screen regions, without adding a single calibration point.

3. Adaptive calibration

Today the researcher chooses the length of the calibration in advance, and every participant gets the same. We are working on a calibration that adapts to the participant: short for the easy cases and longer only where it pays off.

4. Fixation detection

Our next Lab Note, Fixation Detection Analysis, compares fixation detection algorithms on webcam data, in recordings made simultaneously with an EyeLink 1000 and in a fast-paced fixation task that our online participants completed. It shows how the new default detector of Eye Tracking 2.0 doubles the share of steady looks reported as exactly one fixation.

5. Session quality in your data

Every session already measures a lot about its own quality while it runs. We will add a compact quality output per session, with an overall quality score, so that researchers can rank, weight or exclude sessions or participants after data collection.

10. Try it yourself

The best way to judge webcam eye tracking is to experience it. The studies below run Eye Tracking 2.0 in your browser, with your own webcam: allow camera access, follow the calibration, and see for yourself. Like everything in Eye Tracking 2.0, the processing runs on your device and no image or video leaves it. The accuracy task is the study behind the results of this note; the other studies are demonstrations of typical research designs and were not part of the validation.

11. Data and reproducibility

Every result in this note can be checked independently. We publish the anonymized data of the 494 completed sessions the results are based on, together with the scripts that compute every number and every figure from it.

Get the data bundle on OSF (ZIP, 9 MB)

What it contains. One row per participant (tier, screening, calibration results and timing), one row per trial of the 45-target accuracy task with its error, raw and with the corrections, and for the last second before each target change, the window the accuracy metric uses: gaze samples with the model's confidence, fixations and, on the Highest tier, face brightness. It also contains the timing of every Gaze Control run and aggregate tables for the values that are not shared per participant (technical rejections before the start, retention after the probe, and the 2023 laboratory precision values). A README documents every column.

How to reproduce. The scripts need only Python 3 and its standard library. Run scripts/accuracy.py and then scripts/article_numbers.py: together they apply the metric exactly as described in this note and write the data of every figure and every number quoted in the text. The only values not computed from the data are the aggregate tables, the expected accuracies of the Highest tier in Figure 12, which are set in the product, and the estimated accuracy of the passive tier.

Privacy. The data contains no identifiers of any kind: no Prolific or Labvanced IDs, no names, no demographics, no device or browser information and no absolute dates or times. Each participant has a random ID that is not linked to any other record, and no key exists. No camera images or video were ever recorded. All participants consented to the anonymized use of their data.

License. Copyright © 2026 Scicovery GmbH (Labvanced). The data and scripts are licensed under CC BY-NC 4.0: you may use, reanalyze and share them and publish your results for non-commercial purposes, with credit to Labvanced and a citation of this note. Commercial use requires our written permission.

How to cite. APA style, ready to copy:

Describing it in your methods section. You are welcome to adapt this paragraph (replace the values in brackets with your setup): Gaze was recorded online with the participants' webcams using Labvanced Eye Tracking 2.0 (Labvanced, 2026), with the [High] calibration tier and [balanced] participant screening. Live head-pose and Gaze Control corrections were applied. In the validation of this configuration, the median accuracy was [5.2 %] of the screen diagonal, about [1.8°] of visual angle at 60 cm.

References

• Alp, N., Lale, G., Saglam, C., & Sayim, B. (2024). The effect of processing partial information in dynamic face perception. Scientific Reports, 14, 9794. https://doi.org/10.1038/s41598-024-58605-7
• Bertrand, J. K. (2023). The utility of webcam eye-tracking: Lessons learned and practical applications in decision-making and digital human-computer interactions [Doctoral dissertation, University of Alberta]. https://doi.org/10.7939/r3-83ge-t169
• Bertrand, J. K., & Chapman, C. S. (2023). Dynamics of eye-hand coordination are flexibly preserved in eye-cursor coordination during an online, digital, object interaction task. Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. ACM. https://doi.org/10.1145/3544548.3580866
• Bertrand, J. K., Ouellette Zuk, A. A., & Chapman, C. S. (2023). Continuous measures of decision-difficulty captured remotely: II. Webcam eye-tracking reveals early decision processing. bioRxiv. https://doi.org/10.1101/2023.06.06.543799
• Bánki, A., de Eccher, M., Falschlehner, L., Hoehl, S., & Markova, G. (2022). Comparing online webcam- and laboratory-based eye-tracking for the assessment of infants' audio-visual synchrony perception. Frontiers in Psychology, 12, 733933. https://doi.org/10.3389/fpsyg.2021.733933
• Calignano, G., Lorenzoni, A., Semeraro, G., & Navarrete, E. (2024). Words before pictures: The role of language in biasing visual attention. Frontiers in Psychology, 15, 1439397. https://doi.org/10.3389/fpsyg.2024.1439397
• Golan, M., Abo Shkara, M., & Havron, N. (2023). Children infer the meaning of novel verbs through morphological bootstrapping. PsyArXiv. https://doi.org/10.31234/osf.io/xk8eu
• Grosse Wiesmann, C., Rothmaler, K., Hasan, E., Habdank, K., Yang, C., Yeung, E., & Southgate, V. (2025). The self-reference memory bias is preceded by an other-reference bias in infancy. Nature Communications, 16, 6311. https://doi.org/10.1038/s41467-025-61642-z
• Haverly, C. (2022). Influence of video game playing on change detection: An eye tracking study. Horizons, 7, 68–73. University of Hawaiʻi at Mānoa. ScholarSpace
• Jenke, L., & Sullivan, N. (2025). Attention and political choice: A foundation for eye tracking in political science. Political Analysis, 33(4), 298–314. https://doi.org/10.1017/pan.2025.8
• Kaduk, T., Goeke, C., Finger, H., & König, P. (2023). Webcam eye tracking close to laboratory standards: Comparing a new webcam-based system and the EyeLink 1000. Behavior Research Methods, 56(5), 5002–5022.
• Kim, J. Y., Rowell, B., Whitaker, M., Chen, Q., Al Kontar, R., Lester, C., & Yang, X. J. (2026). Effects of AI assistance timing on pharmacists' trust in automated pill recognition technology. JMIR Human Factors, 13, e92822. https://doi.org/10.2196/92822
• Niehorster, D. C., Nyström, M., Hessels, R. S., Andersson, R., Benjamins, J. S., Hansen, D. W., & Hooge, I. T. C. (2025). The fundamentals of eye tracking part 4: Tools for conducting an eye tracking study. Behavior Research Methods, 57, 46. https://doi.org/10.3758/s13428-024-02529-7
• Patterson, A. S., Nicklin, C., & Vitta, J. P. (2025). Methodological recommendations for webcam-based eye tracking: A scoping review. Research Methods in Applied Linguistics, 4(3), 100244. https://doi.org/10.1016/j.rmal.2025.100244
• Sander, I., Mazumder, R., Fingerhut, J., Parada, F. J., Koselevs, A., & Gramann, K. (2024). Beyond built density: From coarse to fine-grained analyses of emotional experiences in urban environments. Journal of Environmental Psychology, 96, 102337. https://doi.org/10.1016/j.jenvp.2024.102337
• Sauter, M., Hirzle, T., Wagner, T., Hummel, S., Rukzio, E., & Huckauf, A. (2022). Can eye movement synchronicity predict test performance with unreliably-sampled data in an online learning context? ETRA '22: Symposium on Eye Tracking Research and Applications. ACM. https://doi.org/10.1145/3517031.3529239
• Saxena, S., Fink, L. K., & Lange, E. B. (2023). Deep learning models for webcam eye tracking in online experiments. Behavior Research Methods. https://doi.org/10.3758/s13428-023-02190-6
• Serrano-Carot, M., Angele, B., Xu, H., & Vasilev, M. R. (2025). Webcams can be used to study eye movements during reading. PsyArXiv. https://doi.org/10.31234/osf.io/bzt2h_v1
• Stone, S. A., & Chapman, C. S. (2023). Unconscious frustration: Dynamically assessing user experience using eye and mouse tracking. Proceedings of the ACM on Human-Computer Interaction, 7(ETRA), 168. https://doi.org/10.1145/3591137
• Tsai, C. C., Kim, J. Y., Chen, Q., Rowell, B., Yang, X. J., Kontar, R., Whitaker, M., & Lester, C. (2025). Effect of artificial intelligence helpfulness and uncertainty on cognitive interactions with pharmacists: Randomized controlled trial. Journal of Medical Internet Research, 27, e59946. https://doi.org/10.2196/59946
• Tsuji, S., Amso, D., Cusack, R., Kirkham, N., & Oakes, L. M. (2022). Editorial: Empirical research at a distance: New methods for developmental science. Frontiers in Psychology, 13, 938995. https://doi.org/10.3389/fpsyg.2022.938995
• Tsutsuse, K. S. (2025). Effects of online advertisements on processing and evaluating scientific explanations [Master's thesis, University of Hawaiʻi at Mānoa]. ScholarSpace
• Wies, S., Bleier, A., & Edeling, A. (2023). Finding Goldilocks influencers: How follower count drives social media engagement. Journal of Marketing, 87(3), 383–405. https://doi.org/10.1177/00222429221125131
• Zhang, P., & Zhang, S. (2025). Attention and learning in L2 multimodality: A webcam-based eye-tracking study. Language Learning & Technology, 29(1), 1–27. https://doi.org/10.64152/10125/73626
• Zhang, X., & Powdthavee, N. (2026). Attention as a bottleneck in moral judgment: How comparative context distorts discrimination recognition. PsyArXiv. https://doi.org/10.31234/osf.io/pnurb_v1
• Zhao, Y., Li, Y., Zhang, K., Li, Z., Hu, Y., Tan, L., Jia, H., Wang, S., Gao, Z., Song, Y., Li, X., Zhao, C., & Cao, A. (2025). Transcranial light stimulation improves selective attention in children with ADHD. medRxiv. https://doi.org/10.1101/2025.09.18.25335086
• Labvanced Eye Tracking documentation: Study settings, Data output.