Is predictive eye tracking accurate?
Predictive eye tracking can help estimate early visual attention. A headline such as “94%” does not mean 94 out of 100 people will look at your button. Check the metric, benchmark and human-agreement ceiling before using that number to make a design decision.
Try one image, then use this checklist to interpret the result.
What can “94%” mean?
For a ceiling-normalized benchmark claim, their headline % is a ratio to the human-agreement ceiling on a public benchmark. The denominator describes how well one group of observers predicts another under the benchmark’s conditions. People do not look at every image in exactly the same way, so agreement between people is itself imperfect.
To read such a ratio, ask for the raw model score and raw human-reference score. Dividing the former by the latter gives the reported fraction of the ceiling. That fraction is not the probability that a particular visitor will notice a particular element. Do not compare percentages calculated from different metrics or different human-reference procedures.
Brainsight’s pricing FAQ uses a 94% headline against live eye tracking. Its exact denominator is not independently established here. The explanation above describes ceiling normalization; it does not verify that vendor’s calculation. Request its validation method before interpreting the headline as that ratio.
Read raw scores before choosing a winner
AUC evaluates how well a saliency map ranks fixation locations against comparison locations. CC compares the similarity of two maps. NSS reads normalized map values at recorded fixations. These metrics ask different questions; a strong value on one is not interchangeable with a strong value on another. Bylinskii and colleagues explain the metric differences and why evaluation choices affect conclusions.
The MIT/Tübingen benchmark is a place to inspect named models and evaluation protocols. Read the dataset, metric and human reference together. Glance places raw research references beside the human ceiling and links its sources. A published reference model score does not establish a measured result for Glance’s exported model or runtime.
Where predictions stop
Free-viewing saliency estimates where attention may land without a specific task. A visitor searching for shipping information has a task, prior knowledge and expectations. Those can change gaze. A static screenshot also leaves out scrolling, interaction, loading order and the rest of a session. An attention map cannot establish comprehension, trust, purchase intent or conversion.
Domain matters too. A model evaluated on photographs may behave differently on dense interfaces. DeepGaze IIE studies calibrated predictions in and out of domain. The UEyes workshop derivative discusses eye-tracking data across interface types. Neither source makes an arbitrary landing-page prediction a substitute for observing its users.
Use this review checklist
- Write the design question: for example, whether the hero image competes with the primary action.
- Record the model, input crop, resolution and metric. Keep these fixed across variants.
- Ask whether the benchmark resembles your material and viewing task. Check whether reported scores use the same protocol.
- Change one element. Inspect the new map and test the resulting hypothesis with people.
For a real landing page, combine the map with usability sessions or a properly designed experiment. A clearer predicted hierarchy can still accompany confusing copy or a broken form. Glance belongs to the same broad family of research saliency methods as predictive attention tools; no direct accuracy ranking is established here.
Next: run a landing-page review. Research and asset terms are listed in research credits.
Frequently asked questions
Does predictive eye tracking record gaze?
No. It predicts likely attention from an image; measured eye tracking records where participants look.
Does 94% mean my CTA will be noticed?
No. A ceiling-normalized ratio describes a benchmark score relative to human agreement, not an individual element’s notice probability.
Can I compare two vendor percentages?
Only after checking the same dataset, metric, viewing protocol and normalization. Their denominators may differ.
Does a benchmark validate Glance?
A published research result does not validate a Glance model export or runtime. No direct Glance benchmark is established here.
What should I test next?
Use the prediction to choose a design question, then observe users completing the relevant task.