2. Pick the survivor
Before you upload, drop 2–4 thumbnail candidates. They are ranked by what survives the 168 px sidebar — not what looks best at full size. Winner first.
3. Read the score on a single pick → 4. Repair if it fails → download PNG + JSON.
Delivered, at true size
How the wall works
Every tile is the exact CSS pixel box that surface uses, Lanczos-downscaled the same way analysis measures, with YouTube's duration pill stamped on. Not a CSS shrink of a big image. Scroll sideways to keep true delivered size.
What was measured
Each row is a number with a stated threshold. Recompute it yourself from the source — same pixels always give the same answer.
What the tool saw
detected text lines with cap height attention peak duration-pill keep-out
Repair it
Two deterministic fixes, both things you would do by hand: solve the minimum scrim opacity that reaches WCAG 4.5:1, and recrop 16:9 tighter so the same type is delivered larger. No generative model touches your pixels. Use the Repair button above.
Then what should I change?
Most thumbnails fail for reasons no raster edit can fix — you wrote fourteen lines of copy, or the composition is too busy to survive a downscale. Every item below is arithmetic on the measurements: where it says move it 84 px left, 84 is the computed distance that clears the pill.
The column you actually compete in
The shelf test compares you against your own back catalogue. But your thumbnail appears in a suggested rail beside other channels, all rendered at 168 px, all fighting for the same glance. Paste the videos yours will sit next to and it ranks the column — including how far you sit from its average, because the same red that wins in a muted column loses in a red one.
The shelf test
A thumbnail is never seen alone — it sits in a column with your own back catalogue. Nailing a house style quietly destroys your distinctiveness, and nothing tells you. This compares structure (dHash), colour identity, and where the busy regions sit.