A thumbnail cannot be judged in isolation, and that is why the decision is so often revisited after a video is already published rather than settled beforehand.
The image competes against unknown neighbours
A thumbnail is never seen alone. It appears in a grid alongside other videos, and whether it attracts a click depends entirely on how it differs from whatever happens to surround it.
That surrounding set is personalised and constantly changing, so the same image performs differently across viewers and across weeks without anything about the video itself changing.
Judging an image before publication therefore evaluates it against an imagined context. The only realistic test is placing it into circulation and observing what happens.
Click-through rate is a relative measurement
The figure reported is the share of times a video was shown that resulted in a play, which means it moves when the audience being shown the video changes.
As a video expands beyond subscribers into broader recommendation, it is shown to progressively less interested viewers, so the rate falls even while total views rise sharply.
This makes raw comparison between videos misleading, and it is why a declining rate on a growing video is a sign of expansion rather than of a failing thumbnail.
Changes reset what the system has learned
Distribution systems test a video with small audiences and expand based on response. Altering the thumbnail changes what is being tested, so the accumulated evidence is no longer reliable.
Frequent changes leave a video permanently in early testing, never accumulating enough consistent data to justify wider distribution, which produces exactly the stagnation the changes were meant to fix.
The practical rule most creators arrive at is to change once, decisively, after enough time has passed for the first version to be measured properly.
What actually varies between images
Legibility at small sizes dominates. Most impressions occur at a fraction of the full resolution, which is why detailed compositions and long text perform worse than their designers expect.
Contrast against the interface matters as well, since a dark image on a dark background and a bright one on a light background both recede rather than attract.
The thumbnail sets an expectation it must meet
An image that wins clicks by implying something the video does not deliver produces immediate departures, and departure is measured more heavily than the click that preceded it.
This is why extreme thumbnails frequently reduce distribution rather than increasing it. The additional clicks arrive from people who leave, and the system reads that as a poor match.