2026 midterms · Methods

About the data

Sources, definitions and the rules behind every number

What it covers

The tracker follows the candidates for the U.S. House, the U.S. Senate and governor in the 2026 elections, from the primaries to the general election on Nov 3, 2026: who is running, what they say in public, the money they report and how forecasters rate their races.

House and Senate candidates are those who have filed with the Federal Election Commission and have either raised money or hold the seat. Sitting members of Congress are added from the lab’s list of officials when the FEC list does not carry them. Candidates for governor come from the lab’s own list. Once a primary is called, only the candidates recorded as advancing stay on the race’s list; a sitting member stays on their own page either way.

Where the data come from

Statements. For challengers, open-seat candidates and every candidate for governor, the posts on X the lab collects from their campaign accounts. For incumbent members of Congress, everything the lab collects from them in office: posts on X, press releases, newsletters and floor speeches. The two records are never compared with each other directly.
Money. Totals each House and Senate campaign reports to the FEC for the 2026 cycle: receipts, spending, cash on hand and debts, with receipts split by source. Candidates for governor report to their states, and no figures for them are on file.
Ratings. Cook Political Report ratings, entered by the lab (last updated by the lab on Jun 24, 2026); a forecaster’s judgement about the general election, not the lab’s. House and Senate list only the races rated lean or toss-up; every governor’s race is rated.
Results. The lab records each primary’s result once it is official.
Addresses. The address lookup asks the Census Bureau’s geocoder for the state and the congressional district under the maps for the 2026 election. The address is sent to the Census Bureau and not kept by the lab.

The five kinds of statement

An AI model reads each statement and marks which of five kinds it contains. A statement can carry several labels or none, so a candidate’s five rates do not add up to 100%.

Personal attacks: Overtly questions a named person’s character, integrity or loyalty, or calls them names. Criticising a record or a policy does not count.
Policy criticism: Objects to a specific policy, law or court ruling with fact-based arguments, without personal attacks or emotional appeals.
Policy discussion: Discusses a specific public policy or policy area (health care, taxes, defense, …).
Bipartisanship: Emphasises working across the aisle, compromise or common ground between the parties.
Credit claiming: Claims credit for passing legislation, securing funding or grants, or other accomplishments in office.

How the numbers are made

Rates. A candidate’s rate on a kind of statement is the share of their classified statements carrying that label. Rates are shown once a candidate has 10 classified statements and ranked from 50, so a handful of posts cannot top a list.

Rankings and the party comparison. The leaderboards rank every candidate with 50 or more statements, one list per party. The party comparison averages candidates’ rates, each candidate counted once, separately for incumbent members of Congress and for everyone else, because the two are measured on different records.

Headlines. Every headline is written from the data by a fixed rule, the same for every race and every candidate of every party. The overview leads with how many primaries are called and the days to the general election. A state leads with its closest-rated race, toss-ups before lean. A race leads with the candidates recorded as advancing, or with the primary date before a result. A candidate leads with the kind of statement on which they sit farthest from the middle of their peer group (same party, office and record; both parties when the group has fewer than 10), if it is outside the middle half (a score above 0.25, ties within 0.03); a high rate backed by fewer than 10 labelled statements never leads. Otherwise a plain count leads.

Awards. The award names are the lab’s. Each goes to the candidates at one end of a kind of statement among every candidate with 50 or more statements, all parties ranked together; the top and bottom 3% are rounded up to whole candidates.

Policy Discussion Leader: Top 3% in policy discussion among all ranked candidates.
Least Policy-Focused: Bottom 3% in policy discussion among all ranked candidates.
Policy Criticism Leader: Top 3% in policy criticism among all ranked candidates.
Least Policy-Critical: Bottom 3% in policy criticism among all ranked candidates.
Accomplishments Leader: Top 3% in credit claiming among all ranked candidates.
Fewest Accomplishment Claims: Bottom 3% in credit claiming among all ranked candidates.
Bipartisanship Leader: Top 3% in bipartisan language among all ranked candidates.
Least Bipartisan: Bottom 3% in bipartisan language among all ranked candidates.
Least Civil Candidate: Top 3% in personal attacks among all ranked candidates.
Most Civil Candidate: Bottom 3% in personal attacks among all ranked candidates.
Zero Personal Attacks: No statement marked as a personal attack, with 50 or more statements.

Known limits

Posts on X are one channel. A candidate who campaigns mostly in person, on television or on other platforms is measured on what they post here, which may be little.
The AI model makes mistakes on individual statements; rates over many statements are more reliable than any one label.
House districts follow the candidates’ FEC filings. Where a state redrew its map for 2026, a few candidates may still appear under their old district, and a sitting member is placed in the race for the seat they hold now.
The ratings are a forecaster’s judgement entered by hand (last updated by the lab on Jun 24, 2026); they are not refreshed automatically.
Campaign-finance totals are as filed and can be amended later by the campaigns.
A rate describes what was said, not whether it was right. Back to the overview.