The squint test for designers: value study vs model
The squint test helps designers see large light and dark shapes without getting distracted by detail. Use it as a painter’s value study for a page. A saliency model adds learned gaze patterns, but neither method proves that visitors understand the page or will click.
Start now: open one design at its intended size and identify its main action.
1. Look for the large value groups
Narrow your eyes gently until fine detail becomes harder to resolve. Notice which blocks remain distinct: a dark headline, a light panel, a face or a saturated action area. Ask whether the broad arrangement supports your intended order. Stop if it feels uncomfortable; a reduced-size view can provide another quick way to inspect large shapes.
In painting, a value study separates the composition into light and dark masses before detail is finished. Applied to a layout, the useful question is whether those masses form a readable hierarchy. It is a visual inspection, not a calibrated measurement of gaze.
2. Name one competing shape
Pick one element that remains more prominent than intended. For a landing page, a decorative panel may form a larger dark block than the headline and action together. Try one change: reduce the panel’s contrast, remove decoration or adjust spacing. Look again at the normal design as well as the simplified impression.
Keep the action label readable. A composition that works as abstract shapes can still have poor text contrast, unclear labels or inaccessible controls. The squint test does not evaluate those properties. Check the interface at normal size and with the input methods people need before calling the design ready.
3. Compare with a model on the same image
A blur or squint mainly suppresses detail. A learned saliency model estimates likely fixation density using patterns learned from eye-tracking data. That can include semantic cues such as faces and text rather than just the sizes of light and dark patches. The two methods can therefore disagree for useful reasons.
Itti, Koch and Niebur’s saliency model is a reference for combining visual features into an attention map. DeepGaze IIE describes a later learned approach. These are different modeling methods; neither paper turns a manual squint into an eye-tracking experiment. A model overlay is also a prediction rather than a measurement of your audience.
Use Glance’s browser view for an image check, or inspect the desktop availability for local workflows. Record which model produced the result. A manual value study and a model output are separate observations; do not average them into an invented accuracy score.
4. Investigate disagreement
If a portrait attracts the model while a banner dominates your value study, inspect both at normal size. Does the portrait support the message or pull attention away from it? Could the banner be visually dominant without communicating anything useful? Change one element and compare again using the same image crop and model.
For graphic layouts, Fosco and colleagues study visual importance across design types. That distinction matters: visual prominence, predicted gaze and useful information are related questions with different measurements. Do not call a decorative element successful merely because it wins attention.
5. Finish with a task check
Ask someone from the intended audience to explain the offer and find the next action. Observe the attempt before explaining your design. This reveals confusion that a simplified composition or saliency overlay cannot establish. A strong visual hierarchy is useful only when it helps someone do what they came to do.
Next: follow the landing-page review steps. To interpret benchmark scores, read what predictive accuracy means and the linked research credits.
Frequently asked questions
What does the squint test show?
Large light and dark masses and their visual hierarchy after fine detail becomes less visible.
Is squinting measured eye tracking?
No. It is a manual composition check without recorded gaze or a calibrated benchmark.
How does a model differ from a blur?
A learned model uses patterns from gaze data, including semantic cues; blur mainly removes detail.
What if the model and squint test disagree?
Inspect the original image, identify the competing element and compare one controlled revision.
Can either method prove usability?
No. Check accessibility, comprehension and task completion directly with users.