Vasundra Srinivasan measures when browser assistants’ planned clicks become stale. On 31 public pages, screenshots captured during loading cause much more trouble than a long thinking interval on a settled page.
A click counts as stale when its intended target has moved, disappeared or become covered. The paper separates direct target tracking, commercial-site layout measurements and synthetic tests of verification methods.
- Loading dominates the measured risk. On fully loaded pages, 0.4% of planned clicks become stale within thirty seconds. On newly opened pages, 17% become stale within the first second.
- Waiting helps on the study pages. Waiting for network quiet leaves 0.46% stale clicks, at a median cost of 1.24 seconds. Reimplemented wait rules from eight of ten assistants leave more than 1% stale.
- Commercial homepages keep moving. In a desktop crawl, 52% of 65 booking and buying homepages shift after a network-idle wait capped at three seconds.
- Pixels check position cheaply. On synthetic moving pages, a pixel check detects 82% of stale targets with 4% false alarms in 312 milliseconds. Opus detects 83% with 12% false alarms in 2,366 milliseconds.
The released test generator distinguishes target position, obstruction and whether an action succeeded. A failed position check should trigger a new plan from a fresh screenshot, not an adjusted click.
Caveat: Commercial layout shifts do not measure checkout failures; the attempted transfer to stale-click rates fails validation. Check accuracy comes from synthetic pages, and the assistants themselves were not tested end to end.