Looks: six frames from every ad, read by an image model; ads are cut into 45 looks of 16 to 146 alike ads, the same groups as the dashboard. Messages and problems: every transcript translated to English, then grouped by its core message, or by problem, cause, promise and angle. On the map each look is drawn as its own cluster; nearby clusters look more alike. These are wider groups than the dashboard's tight looks, which the crowding findings use.

Each ad's frames are turned into vectors with Meta's PE-Core image model. Two ads count as one creative when their frames match above a cut-off fitted on Traya's own launches: new ads at least that alike to an ad already running got about 18% less spend in their first 30 days (2,016 launches, Feb–Oct 2026). "First 2 seconds" compares the opening frames; "Whole video look" compares six frames across the ad. Positions on the map are a 3D projection for viewing; groups are worked out on the full vectors, so two ads can sit close on the map without being grouped.