An AOI is a counting rule, not a neutral outline
An area of interest is a region drawn over a scene so that eye movements falling inside it can be summarized as attention to that region. Fixation duration and dwell time within each AOI are the typical outputs [5]. Because those statistics are computed only from fixations that land inside the drawn boundary, the boundary itself decides what counts. Move an edge, merge two regions, or split one, and the same eye movements produce different totals.
This is why AOI definition is described as the most important metric in eye-tracking analysis: AOI statistics are what make raw eye-movement data interpretable in the first place [4]. The choice is not cosmetic bookkeeping layered on top of the data; it is part of how the attention measure is constructed.
Where the numbers diverge: fixed versus data-driven regions
One clear source of divergence is whether AOIs are fixed in advance or derived from the data. In a study of visual attention in dynamic social scenes involving joint attention, researchers compared a priori AOIs, defined before analysis, with a posteriori AOIs, defined from the observed gaze patterns, and analyzed how attention was distributed across them [2]. The comparison matters because a posteriori regions are shaped by where people actually looked, so they can redistribute attention across categories relative to a pre-set layout.
A second source is the model used to define regions. In an ADHD study using naturalistic eye-tracking, two saliency models were used to define AOIs, and attention statistics were reported as scanpath saliency, a measure of how closely gaze follows visually salient locations [1]. Using two models rather than one implies the resulting attention statistics depend on which model's regions are used. The study frames this in terms of bottom-up versus top-down attention, so the AOI choice is tied to which mechanism the statistic is meant to reflect [1].
Why a differently shaped AOI changes what the measure means
Some proposals change the AOI geometry itself. The Peripheral Attention Index introduces area-weighted distal polygonal AOIs as an alternative to conventional AOIs for fixation-duration analysis, and is presented as applicable to general AOI-based eye tracking [3]. The key word is area-weighted: weighting by region size means a fixation's contribution depends on how large the region is, not just whether the gaze fell inside it. That is a different statistic from a plain within-AOI fixation duration, even when both are computed from the same recording.
The practical implication is that two studies can report "fixation duration in the AOI" and mean different quantities. A conventional AOI counts time inside a boundary; an area-weighted polygonal scheme adjusts that count by region area [3]. Neither is wrong, but the numbers are not directly comparable unless the AOI scheme is stated.
What this means when you read an attention result
The safest reading is that an attention statistic is only as specific as its AOI definition. A study that reports dwell time across five predefined AOIs, as in the nicotine-packaging work, is reporting attention to those five experimenter-chosen regions, not attention to packaging in general [5]. That study did not test how alternative AOI choices would change its statistics, so its numbers should be read as tied to its own layout.
Where AOI choice has been examined directly, the finding is that the definition is a live variable rather than a fixed constant: a priori and a posteriori definitions produce different attention distributions in dynamic social scenes [2], and different saliency models produce different attention statistics in naturalistic viewing [1]. Systems that support both manual and automatic AOI definition make the same point operationally, since the two routes can produce different regions from the same video [4]. The honest summary is that AOI choices can shift attention statistics enough to matter for interpretation, and that reporting the AOI scheme alongside the number is what makes the number meaningful.
Sources used in this answer
Saliency models reveal reduced top-down attention in attention-deficit/hyperactivity disorder: a naturalistic eye-tracking study
Using two saliency models to define AOIs in naturalistic eye-tracking of ADHD produced attention statistics based on scanpath saliency, linking AOI/model choice to the resulting attention measure.
Comparing a Priori and a Posteriori Areas of Interest in Dynamic Social Scenes: An Eye-Tracking Study of Visual Attention in Autism Spectrum Disorder
Comparing a priori and a posteriori AOI definitions in dynamic social scenes involving joint attention showed that the choice of AOI definition affects the analyzed distribution of visual attention.
Peripheral Attention Index: Area-Weighted Distal Polygonal Areas Of Interest
The Peripheral Attention Index proposes area-weighted distal polygonal AOIs as an alternative to conventional AOIs for fixation-duration analysis in general AOI-based eye tracking.
A complete system for analysis of video lecture based on eye tracking
A video-lecture eye-tracking system supports both manual and automatic AOI definition and provides statistical analysis of eye-movement data, treating AOI as the central metric for interpreting that data.
Visual attention to nicotine product packaging: Evidence from an eye-tracking study among university students
In a nicotine-packaging eye-tracking study, fixation duration and dwell time were measured within five predefined AOIs, without testing how alternative AOI definitions would change those statistics.