Which rewrites quietly make a page harder to retrieve?

Because two stages score different things and one edit moves both. The stage that finds your page scores overlap with the question somebody typed; the stage that writes the answer scores how cleanly your passage lifts. Replacing the page’s ordinary words improves the second and subtracts from the first, before any metric you watch exists. The 2026 critical survey puts it in a line: a rewrite “may therefore perform well once injected while making the document less retrievable or less competitive upstream”.²

The measurement under that sentence is SAGEO Arena, published at KDD 2026: 2,700 queries from nine public retrieval datasets, the 171,003 web documents behind them, and a real retriever and reranker where other benchmarks put a fixed candidate list.¹ Its verdict on body-only editing is flat: “optimizing body text alone consistently degrades visibility across all stages.” Averaged over ten strategies, top-20 presence at retrieval fell from 0.58 to 0.53, top-10 presence after reranking from 1.00 to 0.84, and final citation from 0.50 to 0.47.¹

The attribution names which edits carry risk. The authors put the retrieval loss down to “the lexical mismatch between optimized documents and user queries, which typically use common vocabulary”, and show its shape: “replacing terms like ‘eating’ with ‘alimentary routines’ or ‘sleeping’ with ‘somnolence’ directly reduces term overlap.”¹ What that does to a budget is the retrievability ceiling; why position outranks prose is relevance and ranking. This page is about the edit.

Read the diff, not the draft. A rewrite judged by reading the new version always looks better, because it was written to look better; the retrieval cost lives in what left. List the nouns and verbs the edit removed, then check that list against the queries the page already wins. Every word on both lists is one you are paying to delete, upstream, where no invoice reaches you.

The second check takes ten minutes. SAGEO tested its optimised documents against four rewritten forms of every query, expansion, simplification, rephrasing and abstraction, and found “consistent visibility degradation under all query variations, with a progressive drop as queries deviate further from the original phrasing”, worst under abstraction.¹ The explanation: because a rewrite happens without knowing the incoming query, it “tends to elaborate on the document’s existing content, making the document more specific to its original topic and vocabulary”. Ask the vague version of your question and see whether the page still answers it.

The pattern it surfaces most often is a house-style migration. A team moving support pages into one approved glossary voice finds, in the diff rather than any dashboard, that every page swapped the customer’s word for the product team’s, headings included. Nothing in the new copy is wrong. It now competes for a term that exists only inside the company.

Less, and not zero, and the gap changes how you may quote the numbers. SAGEO’s default pipeline uses BM25, a lexical retriever scoring on term overlap, the configuration most exposed to a vocabulary swap; the authors rebuilt the test on a dense retriever and a hybrid. Average top-20 presence under body-only rewriting fell 9% with the lexical retriever, 5% with the hybrid and 2% with the dense one, and average rank loss fell from 4.54 places to 2.81 to 0.95.¹ Their conclusion: the degradation “is not specific to BM25, but reflects a broader limitation of optimizing body text alone regardless of the underlying retriever”.

Google says a compatible thing from the other side. Its systems “have advanced even further and improved upon our ability to understand the relevance of pages, even when there is no exact match between the query and the page’s primary content” (updated 10 Jul 2026).³ Take both at face value: exact term overlap matters less each year and has not stopped mattering. Swapping the other components moved the size without moving the sign, costing 16% of post-rerank top-10 presence under one reranker and 19% under another, and final citation 6% under one generator and 12% under another.¹

None of which argues for keyword density, and the survey rules that out separately. It grades keyword stuffing “null or negative” across multiple benchmarks with a one-word instruction, avoid, and grades fluency and simplification only “weak to moderate”, saying to optimise for the user first.² The finding here is narrower and stranger than a keyword rule. Nobody is telling you to add words. You are being told not to delete the ones that already work.