TL;DR
- Telling the engine a partisan identity shifted baseline exposure by up to 11.6 points before any rewriting. The best single rewrite added about 2 points.
- Rewriting gains stayed positive under almost every one of 59 personas, and no persona consistently amplified or dampened them.
- There are two levers on what a voter sees. The engine’s response to the asker is the larger one; a targeted rewrite is the smaller but deliberate one. Both need safeguards.
The first two experiments, in Part 2 and Part 3, treated the person asking as anonymous. Real voters are not. They tell the engine, directly or through context, who they are. This final part of VoteGEO-Bench asks two things. Does the voter’s persona change the engine’s baseline answer, and does it change how much a targeted rewrite can shift that answer?
How the personas were built
For the United States we used 15 personas of three kinds: demographic, explicitly partisan (Democrat or Republican), and questionnaire-derived profiles built from the pre-treatment questionnaires of 2,306 participants in an earlier AI persuasion study. For Japan we used 44 demographic personas. Each persona was added to the generation prompt, and the rewrites from the earlier experiments were run again under each one.
Finding 1: a stated party identity moves the baseline a lot
Before any rewriting, the share-of-voice gap between the Democrat persona and the Republican persona reached 11.6 points on gpt-4o-mini, 3.4 points on gpt-5-mini and 3.0 points on gemini-2.5-flash. The engine was already answering differently depending on who said they were asking, and the size of that difference depended heavily on the model.
Finding 2: one rewritten page moves it much less
Under the Democrat persona, the strongest strategy, Comparative Distinction, added only 1.76 to 2.20 points of share of voice. Against a baseline gap of up to 11.6 points, a single rewritten page is the smaller lever. The questionnaire-derived profiles told the same story: baseline shifts reached 7.4 points on gpt-4o-mini but only 2.5 and 1.3 on the two newer models.
Finding 3: the rewrites still work under almost every persona
Comparative Distinction produced significant positive share-of-voice gains under all 15 United States personas on gpt-4o-mini and gemini-2.5-flash. On gpt-5-mini, 29 of 30 persona-strategy gains were positive. Across the questionnaire-derived profiles, rewriting gains were positive in 17 of 18 engine-profile-strategy combinations. Knowing who is asking does not switch the effect off.
Finding 4: personas do not consistently amplify or dampen the effect
We tested whether any persona made rewriting gains systematically larger or smaller. Under the conservative inference, with standard errors clustered by query and false-discovery correction, none of the 90 persona contrasts remained significant, down from 7 under the simpler test. We conclude that persona conditioning shows no consistent tendency to increase or decrease rewriting gains.
What this says about the risk
There are two levers on what a voter sees, and they are of different sizes. The larger one is the engine’s own response to the voter’s stated identity, which shifted exposure by up to 11.6 points with no rewriting at all. The smaller one is VoteGEO, a targeted rewrite of one page, which added about 2 points and did so regardless of who was asking. A safeguard that only watches for manipulated pages would miss the larger lever, and one that only audits persona effects would miss the smaller but deliberate one. Both need attention, and the newer models in our runs already showed much smaller persona gaps than the older one.
What this part does not show
The persona analysis used previously collected questionnaire data and reports only aggregate results. Exposure counts mentions and does not say whether a persona received a more favorable or more accurate answer. The runs were offline, against fixed page pools, with retrieval disabled.
This completes the series. Start again from Part 1: the risk and what engines cite, or read the full paper on arXiv.