Eye tracking is the measurement of where a person is looking, for how long and in what order. A camera, usually helped by infrared light, records the eyes; software converts those images into a gaze point on a screen, a page or a scene. The technique exists because looking is the one step between delivery and memory that ordinary ad logs cannot see. An impression records that an advertisement reached a device. Eye tracking records whether a human eye actually landed on it.

In advertising the term now carries a specific job. It supplies the ground truth behind most commercial attention metrics: vendors collect gaze data from consenting panellists, then train models that predict attention for billions of impressions where no camera is present.

How the measurement works

The eye does not sweep smoothly. It stops on a point, a fixation, then jumps to the next, a saccade. Fixation counts, durations and sequences are the raw material of nearly every eye-tracking report.

Laboratory hardware relies on a method called pupil centre corneal reflection (PCCR). Near-infrared light, invisible to the participant and less affected by room lighting, shines on the eye and produces a reflection on the cornea known as the glint. Sensors locate the centre of the pupil and the glint; the vector between them shifts as the eye rotates, and algorithms turn that vector into a gaze point. Tobii's Pro Spectrum captures an image every 0.833 milliseconds, a rate of 1,200 hertz, fast enough to record microsaccades smaller than 0.1 degree of visual angle, according to Tobii.

Webcam eye tracking trades precision for scale. Instead of infrared sensors, software analyses ordinary video from a laptop or phone camera. WebGazer.js, released by Brown University researchers Alexandra Papoutsaki and Jeff Huang in June 2016, finds the eyes with a face-detection library, separates iris from sclera and calibrates itself from clicks, on the assumption that people look where they click. It placed gaze within 100 to 200 screen pixels, less accurate than dedicated hardware, according to Brown University. Lumen Research, the UK vendor most associated with advertising eye tracking, says its software calibrates a phone or computer camera to each user's eyes and estimates attention within a circle about 2.5 centimetres across. Lumen says all processing happens on the participant's device and that no video or personally identifiable information is uploaded.

A third route sits in the living room. TVision places computer vision sensors in about 5,000 US households to record who is in the room and whether their eyes are on the television, second by second, alongside content recognition of what is playing, as reported by PPC Land on August 30, 2026.

From panel to bid request

Raw gaze data is collected from a sample. The programmatic market needs a number for every impression. The bridge is a predictive model. Vendors record panellists viewing real feeds, pages and programmes, log the placement signals around each exposure (ad size, time in view, screen share, clutter, scroll speed, sound) and learn which combinations predict eyes on the ad. The model then scores impressions with only those signals.

Lumen's partnership with Snapchat, which began in 2021, followed this pattern: laboratory eye tracking was combined with an AI attention model so advertisers could measure attention on Snap Ads, Commercials and AR Lenses "without going to the lab", in the words of Lumen chief executive Mike Follett. DoubleVerify later paired its exposure data with Lumen's eyes-on-screen signals for impression-level reporting on Snapchat. Integral Ad Science (IAS) built Quality Attention on viewability, ad environment, user interaction and Lumen's eye-tracking data.

Training sets can be small. The model behind Teads' connected television (CTV) home screen product rested on more than 15 hours of live eye tracking with 100 smart TV viewers, completed in June 2025, while Lumen's wider dataset spans more than 50 countries, according to the May 6, 2026 announcement.

Origin and evolution

Louis Emile Javal observed in 1879 that reading consists of fixations and saccades rather than a continuous sweep. Edmund Huey built an early tracker using a contact lens fitted with a pointer and published his findings in 1908. Guy Thomas Buswell in Chicago reflected light off the eye onto film, the first non-intrusive method, and his 1935 book studied how people look at pictures. Alfred Yarbus showed in 1967 that the task given to a viewer changes where the eyes go. In 1980 Marcel Just and Patricia Carpenter proposed the "eye-mind" hypothesis, that there is no appreciable lag between what is fixated and what is processed, the assumption on which commercial use still largely rests.

Marketing adopted the method in print and packaging research, then on the web. In March 2005 Enquiro, Did-it and Eyetools presented the "golden triangle" study of Google results pages, finding that listings below fifth position were viewed by only half of participants and ads below the top three had less than a 20% chance of being noticed. Jakob Nielsen published the F-shaped reading pattern on April 17, 2006. Tobii, founded in Sweden in 2001 by John Elvesjo, Marten Skogo and Henrik Eskilsson, made infrared trackers a standard research instrument.

The shift to media trading came with webcams and machine learning. Follett founded Lumen in 2013. Karen Nelson-Field's Amplified Intelligence used device cameras to separate active attention (eyes on the ad) from passive attention (eyes on the screen but not the ad).

Why it matters

Viewability, the dominant quality standard since the Media Rating Council (MRC) guidance of 2014, counts an ad as viewable when half its pixels are on screen for one second, or two seconds for video. It says nothing about whether anyone looked. Eye tracking is the evidence attention vendors cite to argue that two viewable impressions can carry very different value. TVision's August 2026 figures illustrate the gap: of more than 200 daily minutes with video, 24 are advertising opportunities and nine are attentive ad time.

That evidence now feeds buying systems. Google made Adelaide's attention score a Custom Bidding signal in Display and Video 360 (DV360) from mid-July 2026, and Adelaide segments reached Amazon DSP in June 2026.

Limitations and disputes

The method measures looking, not thinking. Research summarised since the 1980s on covert attention shows people can attend to things they are not fixating, and attention tends to lead the eye by roughly 100 to 250 milliseconds. A gaze point is therefore evidence of exposure, not of processing.

Scale is the second problem. Laboratory precision does not survive the move to webcams, and webcam accuracy varies with lighting, glare, device angle and screen size. Panels of thousands then stand in for billions of impressions, so every predictive score inherits its training sample's blind spots. The IAB and MRC attention guidelines finalised on November 12, 2025 respond directly: eye-tracking tools "must be calibrated for accuracy across measured devices", providers must disclose error margins, and deterministic collection should reflect real-world scenarios, with biases from "simulated or forced collection scenarios (such as simulated feeds or auto-scrolling)" studied and accounted for.

Outcomes are contested. A Kantar analysis of 873 campaigns and $3.2 billion in spend found that channel-level attention did not directly track brand contribution or cost-effectiveness. Because vendors weight inputs differently, scores do not compare across providers; Anya Libova of DoubleVerify told an IAB Europe panel on July 3, 2025 that the definition of attention "really depends on who is talking".

Privacy is the sharpest dispute. The IAB/MRC guidelines require explicit consent and compliance with the General Data Protection Regulation (GDPR) and the California Consumer Privacy Act "due to sensitive nature of gaze and facial data". Under GDPR, eye images become biometric special-category data when processed to identify a person, and gaze patterns can reveal health conditions, fatigue or interests. Illinois' Biometric Information Privacy Act (BIPA) covers "a retina or iris scan", and courts are still weighing whether gaze data falls within it, mainly in exam-proctoring cases, according to a July 16, 2024 analysis by law firm Perkins Coie. The Federal Trade Commission's 2023 biometric policy statement treats recordings of characteristic movements as biometric information. Consumer devices have drawn a firm line: Apple's privacy overview for Vision Pro states that where a user looks is processed at system level and not shared with apps, websites or Apple before the user selects something.

Not the same as

  • Attention metrics are the products; eye tracking is one input. Many scores, including Adelaide's AU, are predicted from placement signals such as ad size, time in view, clutter and position rather than measured gaze in each case.
  • Viewability measures whether an ad could be seen, based on pixels and time on screen. Eye tracking measures whether it was seen.
  • Facial coding reads expressions to infer emotion or presence. Kantar's attention framework pairs it with gaze data, but it does not locate where the eyes point.
  • Gaze interaction, as in Vision Pro, uses the eyes as an input device. The goal is control, not measurement, and the data stays on the device.

Aggregates built from gaze data, such as cumulative attention and attention per mille, are covered in separate explainers.

Recent developments

Eye tracking has moved into premium video. Lumen extended measurement to Netflix ad inventory in five European markets on March 5, 2026, and Teads made Lumen's CTV home screen model available globally on May 6, 2026. TVision launched Ad Scoreboard 2.0 on August 25, 2026. With MRC accreditation audits open since the guidelines were finalised, the next test for eye-tracking vendors is whether panels, calibration and model validation survive independent audit.

Timeline

  • 1879 - Louis Emile Javal describes fixations and saccades in reading
  • 1908 - Edmund Huey publishes work based on an early contact-lens tracker
  • 1935 - Guy Thomas Buswell publishes his study of how people look at pictures, using non-intrusive photographic tracking
  • 1967 - Alfred Yarbus shows that viewing task shapes eye movements
  • 1980 - Just and Carpenter propose the eye-mind hypothesis
  • 2001 - Tobii founded in Sweden
  • March 2005 - Enquiro, Did-it and Eyetools present the Google "golden triangle" study
  • April 17, 2006 - Jakob Nielsen publishes the F-shaped reading pattern
  • 2013 - Mike Follett founds Lumen Research
  • April 22, 2015 - Tobii lists on Nasdaq Stockholm
  • June 1, 2016 - Brown University announces WebGazer.js webcam eye tracking, presented at IJCAI in July 2016
  • 2021 - Lumen and Snapchat begin eye-tracking attention research
  • 2023 - FTC biometric policy statement covers recordings of characteristic movements
  • January 2024 - IAS launches Quality Attention with Lumen eye-tracking data
  • February 2, 2024 - Apple Vision Pro goes on sale in the US with on-device gaze processing
  • March 26, 2024 - Snapchat and Lumen launch rich media attention measurement
  • May 12, 2025 - IAB and MRC release draft attention guidelines covering visual tracking
  • June 2025 - DoubleVerify pairs exposure data with Lumen eye tracking on Snapchat; MediaMento CTV eye-tracking study completed
  • July 31, 2025 - Nielsen launches Outcomes Marketplace with Realeyes
  • November 12, 2025 - IAB and MRC finalise Attention Measurement Guidelines
  • March 5, 2026 - Lumen measurement extends to Netflix in five European markets
  • May 6, 2026 - Teads launches Lumen CTV home screen attention model globally
  • July 2026 - Adelaide attention scores enter DV360 Custom Bidding
  • August 25, 2026 - TVision launches Ad Scoreboard 2.0

Summary

Who: Research vendors such as Tobii, Lumen Research, TVision and Amplified Intelligence collect gaze data; verification companies including IAS and DoubleVerify, platforms such as Snapchat, Netflix and Teads, and standards bodies IAB and MRC use or govern it.

What: Eye tracking measures where, how long and in what order a person looks, using infrared trackers, webcams or in-home sensors, and in advertising supplies the ground truth for predictive attention models.

When: Scientific study dates to 1879; marketing use grew from the 2000s; the IAB and MRC finalised attention guidelines covering visual tracking on November 12, 2025.

Where: Laboratories, consenting online panels on laptops and phones, and panel households in front of televisions, with the resulting models applied across web, social, CTV and programmatic bidding.

Why: Impressions and viewability show that an ad could be seen, not that it was; eye tracking provides direct evidence of looking, though its link to outcomes, its comparability across vendors and its privacy treatment remain disputed.