TL;DR
- Does rewriting a page get a business more room in an AI answer? In one 2026 benchmark, a little, and only when the rewrite was told which name on the page it was working for. With no name given, the rewrites produced a significant gain in 1 of 36 test settings. With the name given, 24 of 36 (18 of 36 under the paper's stricter test).
- The source is a National University of Singapore preprint (arXiv 2610.21330, version 1, 15 October 2026, not peer reviewed). Its test bed is election questions in English and Japanese, with political parties as the competing names, three small AI models answering with web search switched off, and gpt-4o-mini writing every rewrite. No business page and no commercial question was tested. Reading it for a business is our extrapolation, and it is not proof of a commercial lift.
- Reorganizing the page, and stating concretely what sets one option apart from the alternatives, finished in the top three most often. Adding statistics or quotes sat mid-table. Adding source attributions and rewording filled the bottom five places.
- The gains were small: by rewrite family, from a fraction of a point to about 4 percentage points, and fewer than half of the family-level results reached 1 point (our count: 14 of 36 in the paper's Table 3, 20 of 48 in its Table 4). Generating the same answer twice with no change at all moved the same measures by up to 1.4 points in the paper's main table. Changing one line that said who was asking to its opposite moved a name's share of voice by 11.6 points on one engine, and by 3.4 and 3.0 on the other two.
- What we take from it is a method: tell the rewrite who it is for, fix the structure before adding decoration, rerun the question with no change before crediting any uplift, and do all of it after checking that the engine retrieves the page at all.
Many of the pages a business hopes to be named from are not about that business alone. A comparison page, a "best of" roundup, a directory listing, a supplier guide: each mentions several competitors, and an AI engine that reads one has to decide how much of its answer goes to each name. Our earlier posts looked at the steps on either side of that decision. GEO tricks, measured covered the tactics that push an item up a list an AI model ranks. What drives AI citations covered the search an engine runs before it answers. This post is about what an edit does once a page with several names on it is already in the engine's hands.
A word on the source first, because it is an unusual one for a business blog. The only benchmark we have read that tests rewrites on pages with several competing names is VoteGEO-Bench, by Shiqi Zhang and Xiaokui Xiao of the National University of Singapore. It is built on web pages and questions from the 2024 U.S. presidential election and the 2026 Japanese House of Representatives election, and the names competing for space in the answer are political parties. The authors study it as a risk: whether editing web pages could shift what an AI engine says before a vote. We do not do GEO for parties or candidates, and this post takes no position on the politics. We read the paper for its method, because election pages have a property that single-brand test pages lack: many of them mention more than one side, and in the U.S. set nearly all do.
The limits that follow from that are large, so here they are before any figure. The paper is a preprint, version 1, not peer reviewed. We read it in October 2026 and have not re-run it. The questions are election questions in English and Japanese. The engines are three small models reached through their APIs (gpt-4o-mini, gpt-5-mini and gemini-2.5-flash) with web search switched off, answering from a set of pages handed to them. One model, gpt-4o-mini, wrote every rewrite. Nothing in it was measured on a business, a product, a French-language question or a consumer app. Every figure below, except the ones we label as ours or as coming from another study, is the authors' own. Wherever we carry an idea over to a business page, that step is ours, it is untested, and we say so.
What did the benchmark test?
VoteGEO-Bench tested, on election questions, whether rewriting one page, among a fixed set of pages given to an AI model, changes how much of the model's answer goes to one name mentioned on that page. The unit is what the authors call a query cell: one question, on one engine, on one day, with the set of pages collected for it. For the U.S. election those are pages Google returned at the time, about 9 per query cell. For Japan they are pages that web-enabled AI engines had cited for the question, about 63 per query cell. The rewriting test covers 2,939 query cells: 2,289 built from 109 U.S. questions and 650 from 25 Japanese ones.
The procedure is the same each time:
- One page is replaced. The engine answers the question from the full set of pages. Then one page that mentions the target is swapped for a rewritten version, every other page stays as it was, and the engine answers again.
- Exposure is counted in the answer. There are three measures. Mention coverage is the share of answers that name the target at all. Share of voice is the target's share of the words the answer spends on the competing names. Win rate is the share of answers where the target gets the most words. A mention counts whether it is praise or criticism.
- Eleven rewrites in four families are compared. Structure (reorganize the page), evidence (add quotes, statistics or source attributions), style (reword) and argumentation (make the case against the alternatives). The last family is new in this paper.
- A no-change rerun sits beside every result. The authors generate the answer a second time with no page changed and report how far the measures move anyway. They call it the placebo. Each test is repeated three times, with fresh rewrites and the pages supplied in a different order.
Does a rewrite help if it is not told who it is for?
No, not in the one benchmark that tested it. In VoteGEO-Bench, a rewrite that was not told which name to serve almost never raised that name's exposure, and sometimes lowered it. The authors ran the rewrites two ways: once with no target given, and once with the target named in the rewriting instruction and nothing else changed.
| Result across the 36 test settings | Target not named | Target named |
|---|---|---|
| Settings with a significant gain in exposure | 1 of 36 | 24 of 36 |
| Settings with a significant loss | 4 of 36 | not stated |
| Significant results left under the stricter test | 0 of 36 | 18 of 36 |
The reason is mechanical. In the U.S. page sets, 92 percent of pages mention both major parties, so an instruction to improve the page does not say whose exposure should rise. When the authors ran a failure analysis of the untargeted rewrites of one strategy in the Japanese set (the one that sharpens a comparison), about a quarter had strengthened an opposing party, and a quarter (25 percent) had removed the target's name from the page. The paper does not say whether the two groups overlap.
Our reading for a business, which the paper did not test: this is the situation of any page where a company shares the space. A rewrite of a roundup (the kind of list we looked at in reading AI visibility scores and top 10 lists), a directory entry or a "which one should I choose" article, briefed as "optimize this for AI", has the same gap: it never says who the edit is for. The brief has to name the business. On a page the business owns and that is only about the business, the target is already obvious and this finding adds nothing. On a page someone else publishes, the business does not control the edit, and the rewrite is a suggestion to that publisher.
Which edits moved a name's share of the answer?
In VoteGEO-Bench, an election-question benchmark, the rewrites that were told their target and that reorganized the page or argued its case gained more than the ones that added evidence or changed the wording. The table gives that result by family, with the no-change rerun in the first row.
| Rewrite family (target named) | Mention coverage | Share of voice | Win rate | Significant in |
|---|---|---|---|---|
| No-change rerun (placebo) | -0.44 to +0.29 | -0.26 to +0.16 | -1.13 to +0.18 | reference row |
| Structure | +1.76 to +2.81 | +0.89 to +1.47 | +0.73 to +2.14 | 9 of 9 settings |
| Argumentation | +0.74 to +1.94 | +0.53 to +1.48 | +0.52 to +2.57 | 7 of 9 settings |
| Evidence | +0.28 to +1.72 | +0.44 to +0.88 | +0.14 to +0.91 | 6 of 9 settings |
| Style | +0.17 to +0.38 | +0.03 to +0.94 | -0.21 to +0.52 | 2 of 9 settings |
Read the first row before the others. On win rate, the no-change rerun moved by as much as 1.13 points in these columns. In the paper's Table 3, the columns for rewrites with no target named, which we do not reproduce here, show the rerun moving by up to 1.32 points on win rate and 1.40 on mention coverage. On gpt-5-mini, the engine where the gains were smallest, the rerun's win-rate swing (0.87 points) was larger than any family's win-rate gain on that engine (the largest was 0.73). The authors note that this model refuses more often on questions that ask it to judge candidates, which leaves fewer answers for a rewrite to change.
The authors also ranked the eleven individual rewrites in each of 18 settings (three engines, two elections, three measures: a different grid from the 36 above).
| Rewrite | Family | What it does to the page | Mean rank (of 11) | Top-three finishes (of 18) |
|---|---|---|---|---|
| Content Improvement | Structure | Clearer sections, more logical order, key information emphasized, redundancy removed. Same facts, same length. | 2.61 | 13 |
| Comparative Distinction | Argumentation | States concretely what makes one option stand out against plausible alternatives: positions, commitments, timelines, mechanisms. Same facts, similar length. | 2.67 | 14 |
| "LLMs.txt" page format | Structure | Adds a block addressed to the model (topic, key facts, citation priority), Markdown headings and summaries, and a closing section on why the page should be cited. | 3.33 | 10 |
| Refutational | Argumentation | States substantive opposing arguments and rebuts them. | 5.00 | 6 |
| Statistics | Evidence | Inserts numerical evidence at relevant points. | 5.17 | 6 |
| Quotes | Evidence | Inserts quotations attributed to credible figures. | 6.28 | 4 |
| Fluency | Style | Improves sentence flow and readability. | 7.50 | 1 |
| Authoritative | Evidence | Adds five or six attributions to credible sources. | 7.61 | 0 |
| Simple Language | Style | Paraphrases in simpler language. | 8.17 | 0 |
| Unique Words | Style | Uses more distinctive vocabulary. | 8.39 | 0 |
| Technical Terms | Style | Rephrases with more technical terminology. | 9.28 | 0 |
Four things in that table matter to anyone buying or writing GEO content.
- The two leaders were told to change no facts. One reorganizes the page. The other spells out the differences between the options the page already describes. Both rewriting prompts told the model to keep the original facts and about the same length. The paper did not verify that the rewrites complied.
- Adding evidence did less. Statistics, quotes and attributions are the edits that the 2023 GEO paper (arXiv 2311.09735) made standard advice, and our own GEO overview cites that paper. Here they finish fifth, sixth and eighth of eleven. The two results are not on the same scale: the 2023 paper measured how visible a source was inside an answer, and this one measures one name's share of an answer when several names share a page. It is still the second benchmark we have covered where reorganizing did at least as well as adding. In GEO-Bench, the content improvement rewrite was the strongest of the legitimate (white-hat) rewrites: 0.125 on that benchmark's rank-gain scale, where 1 is a full climb from last place to first, against 0.083 to 0.105 for quotes, citations, statistics and authoritative tone, on one openly available model (Llama-3.1-8B), as we reported in GEO tricks, measured.
- The evidence rewrites can add things the page never said. The authors did not systematically verify the rewritten pages against the originals, and they note that the evidence strategies may introduce supporting material that was not on the page. A statistic or a quote that a model inserts to lift a score, with no source behind it, is a fabrication. A real figure with its source is a different thing, and this benchmark did not test it.
- The "LLMs.txt" rewrite is not the llms.txt file. In this paper it is a block of text written into the body of the page and addressed to the model, ending with a section that argues why the page should be cited. It ranked third. That is a result on three small models with search off, and it says nothing about the llms.txt file a website publishes at its root. On this evidence we would not write text aimed at the model, instead of the reader, into a client's page.
The family order is a pooled result, and it has exceptions. The family table above does not split the two elections. The paper's appendix (Table 8) does, which is why the next figures are larger than the table's. On gemini-2.5-flash in the U.S. set, the statistics rewrite produced the largest share-of-voice gain of any strategy (+3.71 points), and all four style rewrites gained between 1.41 and 2.80 points. The order does not hold on every engine.
How large are the gains, and how much is noise?
The gains in VoteGEO-Bench are small: by rewrite family, from a fraction of a point to about 4 percentage points, on measures where a rerun with no change can move by a point or more. Structure and argumentation mostly gained 1 to 3 points per engine. Evidence and style mostly gained under 1. Three checks from the paper belong beside any number quoted from it.
- The no-change rerun. It moved by up to 1.40 points in the main table, and by up to 3.24 points in the full appendix table, where results are split by engine and by group. Several individual gains in that appendix are smaller than the rerun in the same column.
- The stricter test. Many of the daily query cells reuse the same question. When the authors account for that and correct for the number of comparisons, the significant target-named settings fall from 24 to 18 of 36, and the significant family results by election group (4 families, 4 groups, 3 measures) from 31 to 22 of 48. The order of the families on share of voice (structure at least as good as argumentation, then evidence, then style) stays the same in the U.S. set and in each Japanese tier.
- Absolute against relative. The same gain in points is a much larger percentage for a name that starts low.
| Measure | Starting exposure of the target | Gain in points | The same gain, relative |
|---|---|---|---|
| Mention coverage, high-exposure target | 72.41% | +4.28 | +5.9% |
| Mention coverage, low-exposure target | 28.55% | +4.12 | +14.4% |
| Win rate, high-exposure target | 13.72% | +1.48 | +10.8% |
| Win rate, low-exposure target | 2.34% | +1.20 | +51.1% |
The authors make that last point themselves: comparable gains in points are larger relative gains for parties that start with little exposure. It is the honest form of the claim that GEO helps the small player most, and it comes with a condition. Plus 51.1 percent and plus 1.20 points on a base of 2.34 percent are the same result, and anyone who quotes one should quote the other.
The rule we take from the placebo row applies well beyond this paper. Before crediting a change with an uplift, run the same question again with nothing changed and see how far the number moves. A before-and-after report that leaves that out gives no way to tell the change from the noise. Live engines vary too: in the University of Toronto preprint we covered in what drives AI citations, two same-day ChatGPT runs of the same question shared 15 percent of their cited pages. Our post on reading AI visibility scores lists what else a measurement should state.
What moved exposure more than the edit?
What the engine was told about who was asking moved a name's starting exposure more than the rewrite did, by a wide margin on one engine and a narrow one on the other two. In the U.S. set, the authors added one line before the question saying who was asking: a demographic identity, a party identity, or a short voter profile. They then measured the target party's share of voice before any rewrite, and the gain from the two leading rewrites.
| Share of voice of the target, in points | gpt-4o-mini | gpt-5-mini | gemini-2.5-flash |
|---|---|---|---|
| Gap in starting exposure between the two opposite stated party identities | 11.6 | 3.4 | 3.0 |
| Gain from the leading comparison rewrite, under the identity aligned with the target | +2.20 | +1.76 | +2.16 |
| No-change rerun in the same setting | +0.22 | +0.38 | +0.07 |
On gpt-4o-mini, changing the asker's stated party identity from one side to the other moved the target's starting share of voice by 11.6 points. The leading rewrite added 2.20. On the other two engines the gap was 3.4 and 3.0 points against 1.76 and 2.16 for the rewrite: larger, though not by much. A stated demographic identity moved the starting point by -0.8 to +2.5 points.
Those gaps compare two opposite identities. Measured from no stated identity, one party-identity line moved the starting point by -7.0 to +4.6 points (the authors' range). On gemini-2.5-flash the shift was under 2 points either way, less than the rewrite's 2.16 (our subtraction from Table 5).
The rewrites kept working whoever was said to be asking. The comparison rewrite produced a significant share-of-voice gain under all 15 asker descriptions (the paper calls them personas) on gpt-4o-mini and on gemini-2.5-flash. On gpt-5-mini, 29 of the 30 gains from the two leading rewrites were positive, and the paper's text does not say how many were significant. The personas did not reliably make a rewrite stronger or weaker: of 90 such differences, none stayed significant under the stricter test. The context moved the floor, and the rewrite added its point or two on top of it.
What the paper measured here is who is asking, not how well known a name is. The link to brand familiarity is ours, and it rests on other measurements. In the Toronto preprint cited above, ChatGPT wrote a brand into its own searches 67 percent of the time when the brand was already in the model's memory for that question, against 11 percent otherwise. When we tested what models know about Quebec roofing with search off in July 2026, 5 of 70 responses named a local business we could verify (what that means for a late arrival is in the rich get richer, the fit get rich). These are different studies measuring different quantities, and they point one way: what the engine brings to the question, its memory of a name or its reading of the asker, can weigh more than an edit to one page. On this evidence the page edit looks like the smaller lever, and none of it has been measured on a business page.
What does this not tell you?
This benchmark does not tell you that rewriting a business page will raise its visibility in ChatGPT, Gemini or any other engine.
- Anything about a commercial question. Every question is about an election and every target is a political party. Carrying the pattern to a trade, a product or an agency is an extrapolation. We think the mechanism is general: a page with several names on it needs a rewrite that knows which one it serves. That is our opinion. The paper did not test it.
- Whether a rewritten page gets retrieved or cited. Web search was off and the page sets were fixed. The authors name this as their first limitation, and they chose it on purpose, to avoid publishing rewritten political content. In the Toronto preprint, ChatGPT's own search explained far more of the citation differences than the page did.
- How a consumer engine behaves today. Three small models answered through their APIs. No top-tier model, no ChatGPT or Gemini app and no Google AI Overview was tested.
- Anything in French. The questions are in English and Japanese.
- Whether the mention is good for you. Exposure counts criticism the same as praise, and the paper reports no human evaluation of the answers.
- Whether the rewritten pages stayed accurate. The authors did not systematically verify them.
- What a different rewriter would do. One small model wrote every rewrite.
The paper also measured which kinds of sources the engines cited on election questions and how quickly those sources turned over. We leave those figures out: they describe election coverage and we see no sound way to carry them to a business. For all these reasons we do not cite this paper to a client as evidence that a rewrite will lift their visibility. A GEO pitch that does is overreading it.
How would we rewrite a comparison or roundup page for AI answers?
We take four working rules from this paper for our GEO work. They are method. None has been tested on a business page, and none promises a result.
- Retrieval still comes first. Run the buyers' questions through the engines and record whether the page is retrieved, cited or passed over, as laid out in what drives AI citations. This paper is about the case that is left over: the page was in the engine's hands and the business was not named.
- Every rewrite brief names its target. Which business, on which page, for which question. On a shared page the client does not own, the brief is a proposal to the publisher: a corrected or clearer description of the client's own entry, with the facts that set it apart.
- Structure and distinctions before decoration. Reorganize the page: clear sections, a more logical order, the key facts emphasized, repetition removed, the same length. Ordering it around the buyer's question is our practice, not something the benchmark tested. Where the page compares options, say concretely what sets the business apart from the alternatives it is compared with: what it commits to, by when, and how it does the work. An invented example, not a client's: "reliable, experienced team" is a pair of adjectives, and "written quote within 48 hours, start date fixed in the contract, one named project lead from first visit to final inspection" is a distinction. Change no facts. Statistics and quotes go in when they are real and sourced, because they make the page more useful to a reader. We do not add them as a visibility tactic.
- No uplift without its rerun. The same question, run again with no change, before the edit and after it. By hand, that means asking the same question several times, in fresh conversations on the same engine, and counting how often each business is named. The spread between those runs is the noise any gain has to clear. It is rougher than the paper's control, since a live engine does not hold its pages fixed. State the result in points with its base, and never as a percentage alone.
If someone shows you a before-and-after chart for a content rewrite, three questions sort it quickly. Which name on the page was the rewrite written for? What did the same question return when it was run twice with nothing changed? And how many points is the gain, on what base? We sell this work ourselves, so we have an interest in how it is bought. Our own measurements start from the questions buyers ask, and the page work follows from what they show (see our SEO audits).