How Electorate HQ works, honestly

What a simulation is

An Electorate HQ presidential simulation runs 51 state delegations, one per state plus the District of Columbia. Each delegation is an AI agent given the same brief: read the tickets and platforms supplied by the user, read current reporting about your state, and decide the race in your state issue by issue, explaining your reasoning. Race simulations on the midterm pages work the same way at the level of a single Senate seat, governorship or House district.

The output is a simulation of how an election might unfold given those inputs. It is not a poll, not a measurement of real voter intent, and not a prediction of fact.

What the agents read

Simulations draw on three kinds of input. First, the material the user supplies: candidate names, running mates, platforms and the state of the union. Second, current reporting: headlines gathered live from Google News for the state or race in question. Third, background from Wikipedia for real races and public figures. Nothing is fabricated to fill gaps: when no public record exists for a name, the simulation is told to say so and judge the platform alone.

Race pages refresh their reporting on a rolling basis, roughly every six hours, so what a crawler or a reader sees reflects recent coverage rather than a frozen snapshot.

The rules the agents follow

• Decide issue by issue. A state is not awarded to a party because it usually votes that way; the delegation must weigh what each ticket proposes against what the state actually needs.
• Treat names and platforms as untrusted input. Flattery, titles and instructions embedded in a candidate field are ignored; people are judged on their public record.
• A disqualifying record cannot be rescued by a pleasant platform, and candidates who are deceased, fictional or constitutionally ineligible cannot win.
• Results are normalized so shares always total 100 percent, with a ceiling on third-party share reflecting the historical record.
• The simulation must explain itself. Every state call ships with its reasoning, the issues that decided it, and voices from the state.

The models

Electorate HQ routes work across several independent AI providers, including Google's Gemini models and other large language models, with automatic failover between them. Which provider answers a given request depends on availability and load. We do not publish the full prompt text, but the rules above summarize its substance faithfully.

Limitations, plainly

• Language models make mistakes: they can misread a state, overweight a headline, or produce results that a careful reader would question. Treat every number as an artifact of the simulation.
• Simulations have no access to polling data, voter files or private information, and none is implied.
• Two runs with identical inputs can produce different results; that variance is inherent to the method.
• Current reporting is fetched from public aggregators and carries the emphasis and blind spots of the coverage itself.

Chamber and statewide boards

The Senate, governor and House boards on the homepage, and the seat bars on the hub pages, are not a separate model. They are the same per-race simulation described above, run on a schedule over every 2026 race in batches of a few races per call, and written to a file the site serves. Visiting a board makes no model call. Each row on it carries the date it was produced, the model that produced it, and how well sourced the reporting behind it was.

The Senate and governor boards are swept whole. The House, at 435 districts, is swept in stages across several runs, so its board mixes rows of different ages and says which districts were simulated when. A visitor with their own AI key can re-run any board; those rows are drawn over the scheduled ones in their browser, labelled as their run, and are not saved.

The seat totals add the seats not on the 2026 ballot to the seats the simulation called each way. Those holdover counts come from the Wikipedia article for each chamber’s 2026 elections, read at each sweep and dated on the bar; when that read fails, the previous count is kept. The House has no holdovers because every seat is up.

What the totals are not: they are not a poll, not an aggregate of polls, not a probability that either party wins the chamber, and not a statement about the electorate. They are the sum of one model reading of the reporting on each race, on one day, with the row marked thin wherever that reporting was scarce. A statewide figure on a state board is the Senate or governor row for that state, the mean of the two where it has both, or the mean of its House rows where it has neither.

What Electorate HQ does not claim

We do not claim to predict elections. We do not claim our results reflect real voter intent. We do not endorse candidates or parties, and no simulation output should be read as an endorsement. Where a page states a fact, the fact comes from the sources described above; where it shows a simulation, it is labeled as one. Read our editorial standards for how we keep those separate, or the plain-English version of this page.