Claude ingests about an eighth of a page before citing, Rudzki finds

Tomek Rudzki published an analysis on LinkedIn on 29 October, saying Claude reads only around 12.5% of a given page before deciding what to cite. He reached the figure by combining Anthropic's official documentation and engineering blog posts with his own tests on how Claude selects sources.

Rudzki also writes that the evidence Claude relies on Brave for retrieval is stronger than before. He did not publish the underlying query set in the LinkedIn post, and the full write-up sits in a linked article.

What the 12.5% figure implies for page structure

If Claude ingests roughly an eighth of a page, the material near the top carries most of the weight for citation. Content buried below that threshold is unlikely to reach the model at all, regardless of its quality.

We think the practical read is that the opening of a page needs to answer the question the page is about, in plain terms, with the specifics a model would quote. That is good practice for readers too, which is the test that matters.

Checking the finding on your own content

Rudzki's number comes from one practitioner's tests. Confirm it on your own pages before restructuring anything.

Prompt Claude with questions your pages answer. Look at what it quotes. Note where that text sits on the page. Clustering near the top gives you a local version of Rudzki's result to work from.

The Brave claim deserves the same check. Run the same queries through Claude and through Brave. Compare the URLs that surface. Repeat across several topics before concluding anything about the dependency.

If the Brave link holds

Claude's visibility would then track Brave's index in part. That makes Brave's crawling and ranking behaviour a second place to look when a page you expect to be cited is not. It does not replace the work of being the best answer on the page itself.