Why build a number for Election Day?
Election Day was Tuesday, November 3. When people talk about the economy, they often mean the stock market or GDP. I think the economy that matters is what a typical household can spend and what each dollar buys. The night before Election Day, I asked if a number like that existed.
I set the rules and made the key calls. AI found the sources, read them, built the number, tested it and checked it. At least 54 AI agents worked on the project. An agent is one AI worker given one job. My first question and my OK on a draft of the site were about 26 hours apart.
What did we find?
The result is the Citizen Economic Index. It counts how many weeks a typical U.S. household must work to pay five basic bills for a year. The bills are rent, groceries, gas, a car and health care. Typical means the household in the middle: half earn more, and half earn less.
We use take-home pay, which is pay after federal taxes. Taxes are figured for a married couple who both work and have no kids.
- In 2025, the five bills took 25.5 weeks of take-home pay. If a household spent every take-home dollar on them from January 1, it would finish on June 28.
- In 2000, the same bills took 23.2 weeks.
- In 1980, they took 28.2 weeks, but that comparison is less certain.
- In 2025, a year of the rent people actually paid took 13.0 weeks. Since 1980, only 2024 was higher, at 13.1.
We found no clear sign that political party moved the number.
What did the AI find first?
My first rule was to stop if the idea already existed. To check, 14 AI agents went to work. Five of them searched government data, research papers and published indexes. They logged 149 entries and scored how close each one came. The core idea did exist.
The Census Bureau tracks it for every year back to 1967. It is the middle household's income, adjusted for rising prices. So the AI paused and brought me five ideas that went further. I asked for one number people could point to, and the AI built it as weeks of pay.
The search also found a close cousin, American Compass's Cost-of-Thriving Index. It counts weeks of a full-time man's pay before taxes. Ours counts weeks of a household's take-home pay.
How did it research the answer?
Five agents gathered the evidence, and three more tried to poke holes in it. We dropped 20 claims that did not hold up. One was "people can't buy homes." The share of households that own their home was 64.2% in 1970 and 65.3% in 2025.
The number uses public data from 12 groups. Nine are federal agencies, like the Census Bureau. The AI had to read some health cost figures off chart images. A second, separate reading matched every value.
What did it build?
The first version counted what middle-income households actually spent. People spent less in 2020, so that year needed the fewest weeks of pay. I pushed back: "Did they even have money to spend?" So the AI rebuilt the number around a fixed shopping list. The list is what middle-income households bought in 2019.
It prices that same list every year, so cutting back can't pull the number down. COVID relief checks still make 2020 dip on the chart, because we count them as pay.
Then it built the site, with a 3D map of 867 counties and 681 cities and towns. Each has at least 65,000 people.
How did it test the party question?
We were not testing a hypothesis, just compiling information with no outcome in mind. So we wrote the test plan and locked it before we ran any party numbers. It is like agreeing on the rules of a game before anyone plays. Two AI analysts then ran the plan's three main tests separately.
Both found no clear sign for the president's party, each governor's party or how each county voted for president. Mayors were not tested. Only about 50 of the 681 cities have a mayor's party on record.
How did it check its own work?
Separate AI reviewers checked the first version of the site and confirmed 51 problems. New reviewers rechecked all 51 after the fixes. The AI also caught bad data, like a fictional Florida governor from the TV show The West Wing on a public list of governors.
Who made the calls, and what is next?
The map colors each city by its mayor's party. When that party was not on record, the AI offered its best guess. Where we could check, its guesses were right 15 of 17 times. I still chose gray, because I did not want to show a wrong party on a real city.
Next, we will look in good faith at other things that may move the number. We also plan for the site to look for new government data every day. It will publish new numbers only when they pass fixed checks run by a simple program with no AI in it.
To me, this project shows how much data work AI can do when a person asks clear questions and guides it. A business could take the same steps with its own records.
See the map at the Citizen Economic Index site. The PDF report has the full method, every year and every state.
