Passage rewriting so a model can lift the paragraph

An assistant does not cite your page. It lifts a span of roughly sixty to a hundred and twenty words and attributes wherever that span came from. That is the unit this engagement works at. We take pages that already turn up in retrieval and lose the citation to somebody else, and rewrite them so a passage cut out of its page still names its subject, carries its number, says when the number was true and says where it came from. What actually changes in the text: Pronouns get resolved. Every passage restates its subject, because "it supports both" is worthless the moment it is extracted from under a heading the retriever discarded. Back-references go, since "as noted above" makes a span depend on a span nobody took. Claim, number, date and source are pulled into the same few sentences instead of being distributed across a section for rhythm. Headings become the question a buyer types rather than a keyword phrase. The first paragraph under each heading answers before it contextualises, because the chunker takes the top of the section and your throat-clearing sentence is what it takes. Most of the gain is deletion: The hedged summary paragraph, the transition sentence, the restated introduction, the second example that repeats the first. These do not only waste a reader's time. They occupy the part of the chunk that gets scored. On a typical page the rewrite is shorter than what it replaces, which is usually the first argument we have with a content team measured on word count. Formats that quietly cost you citations: Comparison data as a screenshot. Specifications inside a JavaScript tab component that loads on click. A table whose header row is styled div elements. Code blocks that split a claim from its explanation, so the passage ends mid-snippet. In developer tools that last one is the common case, and the fix is that the prose claim above the snippet has to stand without the snippet. What your categories need inside the sentence: Crypto claims need a date and a network or version in the same sentence. An undated fee or throughput figure reads as unsafe to repeat and gets dropped in favour of a competitor's dated one. Regulated betting content needs the jurisdiction inside the claim rather than in a footer, or the model declines to restate it at all. Developer tooling needs the error string as a literal heading, because the error string is what people paste. Not included: No new pages. No keyword research, no content plan, no topical map. We do not invent statistics. If a page carries a claim with no first-party evidence behind it, we ask you for a number from your own product data, and if there is not one, the claim comes out. Several clients have found that the most useful part of the engagement. Who this is not for: Sites whose pages say what ten other sites say. Rewriting is a mechanical improvement to retrievability. It cannot make a generic page worth quoting, and we would rather tell you that in the first week than bill you for the pass. How it runs: You send the pages and the evidence you hold. We rewrite in markdown, one pull request per page, with the reasoning for each structural change in the description so your editors can argue with it rather than accept it. Your team merges what it agrees with. Six weeks later we re-run the retrieval check on the same prompts and show what moved, including the pages where the rewrite changed nothing, because those are the ones telling you the problem was never the page.