How many participants does a remote eye tracking study need

A remote eye tracking study needs 30 to 40 valid participants for every group you want to describe with a heatmap, and far fewer when the goal is exploratory. The right number comes from three questions: which decision the data will support, how many segments you need to compare, and how many sessions will fail quality control. This guide walks through all three.

Four factors that set the sample size

Asking «how many participants?» before asking «to decide what?» leads to the wrong answer. A study designed to find problems in a layout and a study designed to pick between two versions are different studies with different demands.

Four factors move the number:

• The goal. Exploration needs few cases and close observation. Comparison needs volume and control.
• The unit of analysis. An aggregate metric per area of interest stabilizes sooner than a full gaze sequence.
• The segments. Every group you read separately (age, country, customers versus non-customers) needs its own minimum.
• Expected data quality. Remote studies lose a share of sessions. Recruit with that loss in mind.

The most cited benchmark in usability comes from Nielsen Norman Group: roughly 39 participants for stable heatmaps and about 6 for qualitative findings. It is a sound starting point for interfaces and on-screen advertising, with one caveat: it covers a single group and stimuli seen by every participant.

Exploratory or comparative: two different rules

Exploratory: small samples, close reading

When the aim is to understand how people scan a package, a page or an ad, 8 to 15 valid participants usually reveal the recurring patterns: the element nobody looks at, the copy abandoned halfway, the reading order. More cases confirm those patterns but rarely change them. The value of this phase lies in watching the recordings one by one and crossing them with the survey answers collected right after.

Comparative: enough volume to separate versions

Choosing between version A and version B is a different problem. Two similar designs usually differ by small margins, and a short sample cannot tell those margins from noise. The practical recommendation is 30 to 40 valid participants per version in a between-subjects design, where each person sees one variant only. If every participant sees both variants in random order, the count per condition drops, because each person serves as their own control.

Two decisions keep the sample from going to waste:

1. Define the areas of interest before launch. Comparing «how much attention the price received» requires knowing where the price sits in each version.
2. Fix the primary metric. Time to first fixation, share of participants who looked at the area and total dwell time answer different questions and stabilize at different sample sizes.

Quotas by segment: multiply, don't divide

A common mistake is to recruit 40 people and then split them into four segments to read results by group. Each subgroup ends up with 10 cases, enough to explore and too few to compare. The rule runs the other way: decide first which segments you will analyze separately, then assign each one its minimum.

An example. A consumer goods brand wants to compare two pack fronts among regular buyers and non-buyers, in two countries. That is two versions by two segments by two countries: eight cells. At 30 valid participants per cell, the study needs 240 valid cases. If the budget covers 120, cut one dimension (a single country, or a within-subjects design) rather than spreading the sample thin and losing reading power in every cell.

When a segment depends on a criterion that is hard to recruit, such as a recent purchase in the category, the study's screener has to verify it with a control question instead of relying on the panel's self-report.

Webcam data quality and drop-out

Webcam eye tracking measures gaze with the camera on the participant's own device, at home, under whatever light they have. That is its strength (reach, speed, cost) and also the reason a share of sessions ends up discarded.

A well-designed remote study applies a known set of controls:

• Calibration up front, with a check that the participant passed it before any stimulus appears.
• Session conditions: face visible, stable distance, no sunglasses and no backlight that blinds the camera.
• Control instructions inside the task, to catch participants who answer without looking.
• Per-participant review of signal quality, dropping incomplete or unstable sessions.

Our article on how eye tracking accuracy is measured explains error in degrees and why it matters when you read a heatmap. For sample planning the practical consequence is simple: recruit above the valid minimum. The exact drop-out rate depends on the panel, the country and the device, so take it from earlier studies with the same participant source. A generous margin in the first study saves a second round of fieldwork.

The device matters too. A desktop study and a mobile study differ in screen size, viewing distance and camera quality; mixing them in one cell adds a variable you cannot remove later.

When the lab is the right call

Remote covers most on-screen research: digital ads, web pages, photographed packs, simulated shelves. In three situations, screen-based eye tracking in the lab or with glasses remains the right tool:

• Fine precision. Small type, dense menus or elements a few millimeters apart call for dedicated hardware.
• Physical stimuli. A real pack on a real shelf is studied with glasses, in store or in a test room.
• Additional signals. When the question involves emotion or cognitive load measured with EEG, skin conductance or facial coding, the lab captures everything in one session.

In practice, many teams combine both: a broad remote phase to choose between versions, then a short lab phase to refine the winner.