Our summary
A web page contains more than the words people read: scripts, styling and long links can take up most of its size. Aldo Morese compares ways to clean pages before sending them to Jev, which judges supplied text rather than fetching websites itself.
The study requested 25 pages. Two timed out and three returned access blocks, leaving 20 content pages for size comparisons. Converting website code into readable Markdown greatly reduced the estimated input size. Removing link addresses reduced it further, though links should stay when the question depends on them.
Small does not mean useful: some pages return an almost empty shell until browser code fills it. Check for missing content before asking Jev to judge it. The study measured input size, not decision accuracy, and used another model's text counter rather than Jev's own.
Key takeaways
- Remove page decoration and unnecessary link addresses before asking a model about the page's meaning.
- Check for blocked pages and empty shells; fetching or rendering usable content is a separate step.
- Leave room for questions as well as page text. A character cap does not guarantee a token limit.