Predicted attention on 93 website captures

What the dataset contains

The published dataset contains 93 successful page analyses, with a mean composite score of 44.1 and a median of 41. These are outputs of a model and scoring formula, not measurements of conversion or participant gaze.

The dataset is historical. It is not a current review of each website, and a named site's design may have changed since capture. The sample was not randomly drawn to represent all websites or all AI-built pages.

Correcting the earlier hero claim

The aggregate contains 75 dominant zones outside the top row of a three-by-three image grid, rounded to 81%. Earlier copy called this ‘outside the hero’. That interpretation was too strong and has been corrected.

The aggregation did not identify hero component boundaries. Outside the top third is not equivalent to below the fold, outside the hero, or outside the reader's attention. No scroll behaviour was measured. The stored legacy field pct_off_hero is retained in the downloadable data for compatibility, but must be read with this corrected definition.

How the summary was calculated

Each map was summarised using normalised intensity, focus, coverage and the dominant cell of a three-by-three grid. The composite combines scaled attention, hierarchy and focus. Scores from other tools or other formulas are not directly comparable.

Only successful analyses entered this aggregate. Capture failures, varying page states, consent notices, responsive rendering and model bias can all affect the result. A low score cannot establish that a site is badly designed, and a high score cannot establish that it converts well.

References: Current metric definitions

Data and reproducibility limits

The download contains the same aggregates and per-domain scores displayed here. It does not contain all original screenshots, per-image maps, capture timestamps or a complete record of the engine configuration. It supports inspecting the published summary, not independently reproducing every prediction.

A stronger follow-up study would preserve those inputs and settings, define the sample in advance and evaluate predictions against an outcome appropriate to the research question. This page does not invent missing records to make the historical study look more rigorous.

References: Download the published dataset (JSON) · Download the reported domain scores (CSV)

What to do with the observation

Use the result as a reason to inspect a design's hierarchy, not as a rule that every hotspot belongs at the top. A product image or relevant information lower in a capture may appropriately attract stronger predicted attention.

For your own page, define the intended message, capture a realistic viewport and inspect the original with the overlay. Test a specific revision and validate important decisions with people.