YTThumbFetcher asks the model for a fixed visual review, not a free-form conversation. The service converts a public YouTube URL into a known thumbnail CDN address, sends that image to Cloudflare Workers AI through AI Gateway, and validates the response against a set report structure. A report that does not match the structure is not shown as a partial result.

What the model receives

The model receives one image URL hosted at i.ytimg.com and a fixed review prompt. It does not receive the video's title, channel name, publication date, views, likes, comments, or any private Studio metric. The interface keeps YouTube Data API fields in a separate section.

Important: the report is produced by YTThumbFetcher. It is not YouTube data, a YouTube recommendation, or evidence of future performance.
Fictional camera cleaning thumbnail with six visual inspection detail windows
A fictional inspection board. The six windows isolate the subject edge, headline shape, local contrast, object order, empty space, and reduced preview. Original illustration created for YTThumbFetcher.

The six dimensions

1. Focus

Focus asks whether the dominant subject is identifiable before secondary details. The review looks at subject size, competing faces or objects, edge clarity, and the likely starting point for the eye.

2. Text readability

Text is judged as part of the image, not as a transcript. The review looks for a phrase that survives reduction, enough stroke weight, and separation from the background. Correct spelling cannot rescue a line that becomes visual noise at recommendation size.

3. Contrast

Contrast includes brightness, color temperature, saturation, and edge separation. A loud palette can still have weak contrast when every region has the same visual weight.

4. Visual hierarchy

Hierarchy asks whether the subject, message, and supporting cue have a deliberate order. A frame weakens when badges, arrows, logos, and text demand attention at the same time.

5. Information density

Density is not an object count. It describes how much sorting the image requires before it makes sense. A detailed product can remain clear when the background and labels stay restrained.

6. Mobile legibility

The image is evaluated for compact placements where fine outlines, small type, subtle expressions, and thin separators disappear. This is a visual accessibility check rather than a device-specific YouTube rendering simulation.

How scores should be read

Each dimension uses a five-level heuristic scale. Level one means the visible evidence conflicts strongly with the rubric. Level five means the visual decision remains clear after reduction. The levels are not percentages, and the site does not combine them into a performance number.

Model and validation

The AI model runs with medium reasoning effort and must return JSON that matches the published report structure. If the first answer is structurally invalid, the same model gets one repair request. A second invalid answer ends the job with a retryable error.

Retention and control

Thumbnail bytes are not uploaded or stored by YTThumbFetcher. Cloudflare Workers AI reads the selected public CDN image. The job record and report expire after 24 hours, and the interface includes an immediate delete command. Rate-limit identifiers are HMAC-protected and expire after 48 hours.

Known limitations

  • The model can misread stylized text, subtle expressions, or cultural references.
  • A strong visual diagnosis cannot establish audience relevance or viewer intent.
  • The report does not compare channels, topics, or historical performance.
  • Creators remain responsible for copyright, accuracy, and platform policy compliance.