Can rewriting a page for AI search make it worse?
Yes, and the measured way it happens is not the one most people worry about. The risk is not a Google penalty. It is that a page rewritten to sound more quotable can lose its place in the set of pages an engine retrieves in the first place, and nothing that happens later in the pipeline can rescue a page that was never retrieved. Two benchmarks that include the retrieval step, rather than handing the model a fixed set of candidate documents, both find it: SAGEO Arena reports that existing optimisation approaches "often degrade retrieval and reranking performance", and C-SEO Bench, peer reviewed at NeurIPS, finds most of these methods ineffective and capable of hurting document ranking. The rewrite that works changes which question the page answers and leaves its structure intact. The rewrite that hurts restyles the prose and strips the scaffolding.
Can rewriting a page for AI search make it worse?
Rewriting a page for AI search can make it worse, and the clearest evidence comes from the benchmarks that bother to include retrieval. Sunghwan Kim, Wooseok Jeong, Serin Kim, Sangam Lee and Dongha Lee built SAGEO Arena to run optimisation strategies through a complete pipeline, retrieval then reranking then generation, over a corpus of 171,003 web documents and 2,700 queries that keeps the documents' structural markup intact. Their finding is that existing approaches are largely impractical under realistic conditions and "often degrade retrieval and reranking performance", that structural information helps mitigate that, and that effective optimisation has to be tailored to each pipeline stage (arXiv 2602.12187, submitted 12 February 2026).
The size of the effect in their results is worth knowing with its scope attached. Averaged across the ten body-text strategies they tested, rewriting the body alone moved top-20 presence at retrieval down about 9%, top-10 presence after reranking down about 16%, and the final citation rate down about 6%, according to the results table in the first version of the paper. Those are benchmark figures from one environment with fictional candidate sets, not a prediction about your site, and the direction is what matters: the losses are largest at the stage that decides whether a page is considered at all.
The peer-reviewed companion finding is blunter. C-SEO Bench, by Haritz Puerto and colleagues, accepted at NeurIPS Datasets and Benchmarks 2025, reports that most current conversational-SEO methods are "largely ineffective, contrary to reported results in the literature", that traditional work improving a source's ranking in the model's context is considerably more effective, and that these methods "can even have a significant negative impact on document ranking".
So the honest answer to a vendor promising a rewrite that makes your pages quotable is to ask which stage the rewrite is aimed at, and what the before-and-after reading will be.
Why would a rewrite aimed at being quoted cost you retrieval?
A rewrite costs retrieval because the pipeline has stages with different appetites, and prose tuned for the last stage can be worse food for the first. An engine's own search has to find your page among millions, a reranker has to keep it in a short list, and only then does a model decide what to quote. Each step scores the page differently.
Most of the familiar advice was measured after that hard part was over. Benchmarks that hand a model a fixed set of candidate documents and then rewrite them can only measure what happens once a page is already in the context window. Olivier Martinez of Sciences Po reviewed 45 studies published between 16 November 2023 and 14 July 2026 and grades the evidence by weight, concluding that "citation-oriented rewrites can impair retrieval" and that "no reviewed technique shows a stable, longitudinal, cross-platform causal effect" (arXiv 2607.14035, submitted 15 July 2026). His survey also notes that the widely quoted gains from the foundational 2024 study describe documents already present in the model's context rather than organic discoverability, which is exactly the distinction a buyer never hears in a pitch.
A passage can become easier to quote and harder to find. Compressing a page into tight answer blocks removes the surrounding language a retriever uses to judge what the page is about. Swapping plain nouns for the industry's preferred terminology removes the words people actually type. Both read as improvements in a document and subtract from the document's findability.
The practical reading is that a rewrite is not one decision but two, and they can pull apart. How the stages work in sequence is set out in our explanation of the RAG pipeline, and the sub-question expansion that happens before any of it is in what query fan-out is.
Can optimising for AI search hurt your Google rankings?
Optimising for AI search rarely hurts Google rankings when the work stays inside what Google documents, and the exposure comes from tactics outside it. Google's own guidance on AI features states that "The best practices for SEO remain relevant for AI features in Google Search", that "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary", and that "You don't need to create new machine readable files, AI text files, or markup to appear in these features" (read 12 November 2026).
That last sentence is the one with money attached. If a proposal's centrepiece is a file or a markup pattern Google says you do not need, the downside is wasted budget rather than a penalty. The genuine ranking risk sits elsewhere: C-SEO Bench's finding that these methods "can even have a significant negative impact on document ranking" came from measuring ranking, not from measuring citations, which means a rewrite can cost you position in the classic index while chasing a quote.
Two rewrites account for most of the real damage we see. Replacing explanatory prose with stacked answer blocks, which thins the page to the point where it no longer covers the topic in any depth. And chasing a phrasing nobody searches, usually the company's internal vocabulary, because the rewrite was briefed on what the business calls the thing rather than on what buyers call it.
Where the two disciplines genuinely conflict, and how to decide between them, is set out in AEO against SEO. The short version is that the conflicts are narrower and rarer than either side's marketing suggests, and almost none of them involve writing more clearly.
Which rewrites are the risky ones?
Four rewrites carry measured or documented risk, and all four are standard items on tactic lists sold as AEO. The table sets out what each is meant to do and what the evidence says happens.
| Rewrite | What it is meant to do | Documented risk | Verdict |
|---|---|---|---|
| Add statistics to the copy | Make claims look verifiable | In LIVE-GEO, a verifier flagged all 80 descriptions produced by this strategy for numerical faithfulness | Only with real, sourced numbers |
| Load the page with technical terminology | Signal expertise | The worst retrieval losses in SAGEO Arena's body-only table, around 14% at retrieval and 20% after reranking | Avoid unless buyers use those words |
| Simplify the language throughout | Be easier to quote | Also among the larger losses in the same table, about 10% at retrieval | Simplify answers, not the whole page |
| Strip prose into answer blocks and drop markup | Be machine readable | SAGEO Arena finds structural information mitigates the damage, so removing it removes the brace | Keep the structure, add the answers |
| Automated rewriting at volume | Cover more questions faster | Google's spam policies name generating many pages without adding value, and manipulating generative AI responses | Not worth the exposure |
| Rewrite the title and headings into the buyer's question | Match how the question arrives | The one page-level lever that survives domain controls in the largest citation study published | Do this first |
The first row is the one to watch in a proposal, because adding impressive figures is the most sellable tactic in the category and the only one with a verification result against it. Runze Kong and colleagues published rewritten pages as live web pages and let real engines choose between them; their verifier flagged 86 of 400 rewritten descriptions, including all 80 from the statistics-adding strategy (arXiv 2611.01126, submitted 31 October 2026).
Read the last row against the others. The rewrite with the best evidence behind it is also the least dramatic: say what the page answers in the words the question arrives in, and leave the rest of the page alone.
Does the gain disappear once everybody does the same thing?
The gain does shrink as adoption rises, and that is measured rather than theorised. C-SEO Bench describes the problem as congested and zero-sum, reporting that as the number of adopters increases, the overall gains decrease (arXiv 2506.11097, accepted at NeurIPS Datasets and Benchmarks 2025). Martinez's survey reaches the same place from a different direction, finding that generic heuristics transfer poorly and that competition erodes individual gains.
That has a practical consequence for how you read any published tactic list, including the ones on this site. A tactic's measured effect was observed in a web where few pages used it. By the time it appears in a conference talk and then in twenty agency blog posts, the comparison set has moved, and the effect that remains is smaller than the one that was published.
Durable advantages in this category therefore tend not to be formatting tricks at all. Covering a buying question nobody answers well is durable, because the absence is specific to your market. Publishing your own measurements is durable, because nobody else has them. Alignment between your page's language and the question is durable because it is not a trick: it is the page being about the thing.
The corollary is that a programme built on a tactic list needs a review date as well as a backlog. Where an engagement should stop spending on page work altogether is covered in how to know when to stop paying for AEO, and what to do when progress flattens is in why your AI visibility has stopped improving.
How do you tell whether your rewrite helped or hurt?
Tell the difference by baselining twice before the rewrite and reading the same prompts afterwards, because a single reading cannot be separated from the engines' own variance. The variance here is larger than most reporting admits.
LIVE-GEO's live-engine results give a sense of the scale. Across its three engines, the engines picked the same top product for only 48.8% to 66.3% of brands, and ChatGPT's choice moved with the order of the candidate URLs: the leading strategy scored 73.8% top-one under one fixed order and 93.8% under the reverse (arXiv 2611.01126, 31 October 2026). Eighty fictional brands in a shopping task is not your category, and the lesson travels: a before-and-after difference of a few points on one run is indistinguishable from the order the engine happened to see.
So read the rewrite at three levels rather than one. Retrieval, meaning does the page appear in the engine's cited sources at all. Selection, meaning is it chosen when it does appear. And accuracy, meaning does the answer describe you correctly. A rewrite can improve the second while damaging the first, which is the specific failure the benchmarks above describe, and a report that only counts citations cannot see it.
Keep a change log with the page, the date and what changed, run the same prompt set either side, and wait for the structural change to register before drawing conclusions. The metric definitions for reading movement across runs are in how to measure AI citation decay, and the reason a single check misleads is in why AI answers change every time you ask.
What does a rewrite that works actually change?
A rewrite that works changes what the page says it answers, in the words buyers use, and leaves the page's structure and markup where they are. That is the unglamorous version of this work, and it is the version with the strongest evidence behind it.
Discovered Labs analysed about 2 million AI citations over six months across four engines, crawling 10,000 cited pages with more than sixty features each, and found prompt-content alignment at a standardised beta of +0.37, with a 95% confidence interval of +0.33 to +0.41, significant at q below 1e-73 after false-discovery correction. One standard deviation more alignment corresponds to "roughly 30% more citations on average" (read 12 November 2026). Two qualifications belong in the same breath: Discovered Labs sells answer engine optimisation services, and in that study a feature had to survive at least four of nine robustness checks to be reported at all.
In practice that means a short list. Put the buyer's question in the title and the headings, in their phrasing rather than yours. Answer each question in the first sentence under it, so the answer can be lifted without the paragraph above. Add facts you can source, and add nothing you cannot. Keep the structured data, the headings and the internal links, since SAGEO Arena's own finding is that structural information reduces the cost of editing a page.
Our own worked example is the teardown of one of our own pages, which scored 28 out of 100 on our framework and reached 71 after an afternoon of exactly these changes, with no new page published. The abstain rule behind it is the same one that keeps the first row of the table above from happening: where a page is too thin to ground a fix, our scoring returns no fix rather than inventing a number to fill the gap. If you want to see which of your pages are being retrieved before anyone rewrites one, a free visibility assessment shows the raw answers and cited sources for your buying questions.
What else do people ask about rewriting pages for AI search?
Marketing and content teams ask these five questions about rewriting pages most often before committing to a rewrite programme.
Can optimising a page for AI search lose it Google rankings?
It can, though not as a penalty. C-SEO Bench, accepted at NeurIPS Datasets and Benchmarks 2025, found that conversational-SEO methods are largely ineffective and can even have a significant negative impact on document ranking. Google separately states that the best practices for SEO remain relevant for AI features and that no special markup or files are needed, so work that stays inside ordinary good practice carries little ranking risk.
Should we stop rewriting pages for AI search altogether?
No, because the one lever with strong evidence behind it is a rewrite: putting the buyer's question and phrasing into the title and headings, and answering it in the first sentence. What the evidence argues against is restyling a whole page into answer blocks, loading it with terminology, or adding numbers you cannot source.
Does adding statistics to a page help it get cited?
Only when the statistics are real and carry a named source. In the LIVE-GEO benchmark published on 31 October 2026, a verifier flagged 86 of 400 rewritten product descriptions for numerical faithfulness, including all 80 produced by the statistics-adding strategy. Unsupported figures are the one tactic in this category with a verification result against it.
Why do published AEO gains not reproduce on our site?
Most published gains were measured on documents already supplied to the model, which skips the retrieval step that decides whether your page is considered at all. A survey of 45 studies published in July 2026 found no reviewed technique with a stable, longitudinal, cross-platform causal effect, and that gains shrink as more sites adopt the same tactic.
How long should we wait before judging a rewrite?
Long enough for the change to be recrawled and for more than one reading to exist. Judge retrieval first, meaning whether the page appears in cited sources at all, then selection. A difference of a few points on a single run is indistinguishable from the engines' own variance, which is why baselining twice before the rewrite is worth the extra run.