2026 Pilot · Retrospective
How the 2026 pilot went
In the fall of 2026 we piloted a nonpartisan voter guide in two states — My Wisconsin Ballot and My Indiana Ballot — for the November 3 general election. One product, two editions, built entirely from official sources.
The guides have now closed. This page is the pilot's public record: what we built, the numbers we can verify, what worked, and what we learned the hard way. It is also the evidence base for what comes next — a redesigned Ballot Guide, planned for all 50 states in the 2028 cycle.
What we built
Enter your address and the guide assembled your ballot: every race you were actually voting on, the candidates side by side, official party marks where parties have them, why-they're-running statements, debate and town-hall highlights, and the deadlines and rules for voting in your state. No endorsements, no rankings, no recommended candidates — the homework done, the decision left to the voter.
The coverage, assembled race by race from official filings and state sources:
Wisconsin: 277 races, 428 candidates. Indiana: 168 races, 324 candidates.
The numbers
We counted with our own first-party counters — no advertising trackers deciding what a "visit" means. The counters ran from October 14, 2026 (the day the store was connected; no first-party data exists before that date) through November 13, 2026.
| Edition | Visits | Ballots viewed | Shares | Visit → ballot | Share rate |
|---|---|---|---|---|---|
| Wisconsin | 1,006 | 355 | 23 | 35.3% | 6.5% |
| Indiana | 1,069 | 508 | 21 | 47.5% | 4.1% |
| Combined | 2,075 | 863 | 44 | 41.6% | 5.1% |
Voters also left feedback: a Net Promoter Score of 37 (27 responses; two-thirds promoters). With a sample that small, treat the score as directional only — we do.
One day of counts (October 22) is missing from these totals — see the meter lesson below. The figures above are a floor, not an exact census.
What worked
The homework framing
About 4 in 10 visits turned into a viewed ballot — 47.5% in Indiana. For a tool with no advertising and no growth tricks, that conversion says the offer landed: voters don't want more politics; they want their specific ballot, explained plainly, in a few minutes.
Official sources, or nothing
Every race and candidate traced back to official state sources. That discipline is slower than scraping aggregators, and it is the whole point: a voter guide that is wrong about one race has failed at the only job it has.
A quality bar with teeth
Every reported bug got an ID, a fix, and an independent verification with screenshot proof before it was called resolved — the person who built a fix never got to grade it. The standard we held: the voter is never the regression tester.
What we learned
1. The quiz wasn't ready, so we removed it
A candidate-matching quiz was part of the original design. Mid-cycle, we retired it rather than keep a feature we couldn't stand behind — a quiz that matches voters to candidates has to be defensible in every answer, and ours wasn't there yet. A real voter quiz is a headline design problem for 2028, and it will be designed as one, not bolted on.
2. Watch your meters — someone is always spending them
Our metrics ran on a free-tier database with a command cap. Automated traffic burned through roughly half a million commands in about a week, the cap silently rejected writes, and one full day of counts (October 22) was lost while the dashboard cheerfully reported success. We fixed the rules — bot gate, leaner writes, a hard budget cap — but the lesson is permanent: meter everything, alert on the meter itself, and archive your final numbers before you switch anything off.
3. Bots are most of the traffic
In the last full hosting cycle we measured, these two small sites accounted for about three-quarters of all requests across our hosting account — some 846,000 requests in under a month — and the overwhelming majority were automated, not voters. Any 50-state version needs caching and bot discipline designed in from day one, not discovered in week three.
4. Two states by hand doesn't scale to fifty
The 2026 data operation was human-supervised, state by state, sweep by sweep. That craft is why the data was trustworthy — and it does not multiply by 25. The 2028 premise follows directly: a fully automated pipeline from official sources to verified guides, holding the same quality bar, or it doesn't ship.
Why there are no winners on this page
Deliberately. We prepared winner-marking under a strict rule — only races a recognized authority had called — but the guides closed without publishing winners. Results belong to the officials who certify them, and both closing pages link straight to them: the Wisconsin Elections Commission and the Indiana Secretary of State.
What comes next
The 2026 pilot added real value with hand-tended tooling — two states, assembled race by race. A 2027 design project is on the books: a 50-state, fully automated Ballot Guide for the 2028 cycle, with the real voter quiz the pilot had to cut. This page is its evidence base — the numbers it has to beat, and the lessons it doesn't get to relearn.