I work in ML for search engines and have been part of very large cohort (n > 5000) user studies carried out by a former employer to measure eye saccade movements over search result pages to understand where user attention is drawn in different search modalities.

In list form results, users typically just click whatever is the first thing without much consideration for its relevance.

In gallery form results, users will scan in a ”backwards C” shape along the first row, down the far right column, then from right to left along the bottom row above the fold, pretty consistently, but still usually just click things in the first row.

The cohort was meticulously designed to allow controlling for factors like age, search interface familiarity (users completed tasks in sessions prior to the eye measurement portion), country, language, etc.