TL;DR
- Of four categories of rewriting, Structure raised exposure in 9 of 9 settings and Argumentation in 7 of 9. Evidence managed 6 of 9 and Style only 2 of 9.
- The two best single strategies reorganize the page or state how the target differs from its rivals. Neither adds a single fact.
- The edits that move exposure are cheap and look legitimate, so fact-checking retrieved pages would miss most of the effect.
Part 2 showed that a rewrite raises a party’s exposure when it knows which party it is for. This part asks what kind of rewrite does it. If VoteGEO is a real risk, it matters whether the effective edits are the expensive ones, such as fabricating evidence, or the cheap ones, such as reorganizing a page. The experiment comes from VoteGEO-Bench and uses the target-specified rewrites throughout.
The eleven strategies and their four categories
We took nine strategies from earlier GEO benchmarks and added two argumentation strategies of our own. Structure strategies reorganize the page or add an llms.txt summary. Evidence strategies add quotes, statistics or authoritative references. Style strategies change fluency, simplify the language, or add technical or unusual words. Argumentation strategies either state plainly how the target differs from its rivals (Comparative Distinction) or anticipate and answer objections (Refutational). For each category we counted how many of the 9 engine-and-metric settings showed a significant gain.
Finding 1: structure and argument work, style does not
| Category | Strategies | Settings with a significant gain |
|---|---|---|
| Structure | Content Improvement, LLMs.txt | 9 of 9 |
| Argumentation | Comparative Distinction, Refutational | 7 of 9 |
| Evidence | Quotes, Statistics, Authoritative | 6 of 9 |
| Style | Fluency, Simple Language, Technical Terms, Unique Words | 2 of 9 |
Reorganizing the page raised the target’s exposure in every setting. Stating the target’s position and its difference from rivals raised it in most. Adding evidence helped less, and changing the wording barely helped at all.
Finding 2: two individual strategies lead everywhere
Content Improvement, which rewrites the page into a clearer structure, had the best mean rank at 2.61 and placed in the top three in 13 of 18 settings. Comparative Distinction, which says how the target differs from the alternatives, was close behind at 2.67 and placed in the top three in 14 of 18. No evidence or style strategy came near either.
Finding 3: the weakest parties gain the most in relative terms
In Japan, where the parties fall into tiers by their baseline exposure, Structure strategies raised mention coverage for the lowest tier by 4.12 points. Against that tier’s low starting point this is a 14.4% relative gain. A party that the engines rarely mention is the one a targeted rewrite can help most, in proportion.
Finding 4: the ordering survives a stricter test
Under clustered standard errors and false-discovery correction, 22 of the 48 category comparisons remained significant, and the ones that dropped out all had effects below 1.5 percentage points. The ordering Structure, then Argumentation, then Evidence, then Style held in the United States and in all three Japanese tiers.
What this says about the risk
The edits that move exposure are the cheap and legitimate-looking ones. Nobody has to invent a statistic or forge a quote. Reorganizing a page and stating a position clearly, both of which are ordinary editorial work, are what shift the engine. That makes VoteGEO harder to detect by content inspection, because the rewritten page is not false, only better organized around one party. It also means that safeguards built around fact-checking retrieved pages would miss most of the effect.
What this part does not show
The experiment measures exposure among a fixed set of pages and not retrieval. It does not test whether readers find the reorganized page more persuasive or more accurate. All rewrites came from gpt-4o-mini, and the original Grok models were unavailable, so gemini-2.5-flash was run on the Grok pools.
Continue with Part 4: does the voter’s persona change the effect?