How we read the races

What Meantly measures, and what it does not

Meantly reads how public figures argue: the rhetorical moves in their words, the likely effect on the audience they target, and how their tone compares with their own usual. It does not check whether a claim is true, does not rate honesty or character, and does not predict who will win. Every candidate in every race is read with the same method and the same rules.

Where the statements come from

Every statement on the site links to its original source. For the 2026 midterms we collect from:

• Official records. The Congressional Record (floor speeches and House extensions of remarks, via the Government Publishing Office's GovInfo API), the White House, and the official press releases of 43 offices held by candidates in these races (U.S. House and Senate offices, governors, a state attorney general).
• Campaigns. The news or press pages of 28 campaign websites, for candidates who publish statements there.
• Candidates' own posts. Bluesky accounts that Bluesky has verified or that use an official .gov handle. Unverified look-alike accounts are never read.
• News reports. 104 local and state outlets covering every state with a race (public radio, TV newsrooms, newspapers, nonprofit newsrooms), national and international outlets, and a daily search of the GDELT news index, which covers thousands of outlets, for articles naming each candidate.

How a quote is credited to a candidate

• Named speaker only. From news reports we keep only direct quotes where the candidate's name sits right next to a speech verb ("said Collins", "Lawler told reporters"). A quote merely near a name is not enough. In the Congressional Record every paragraph names its speaker, so attribution there is exact.
• No namesakes. For candidates, a news quote counts only if the article uses their full name, no one else in it shares their surname, and their state is named. This is what stops, for example, a rugby player called Barrett, or Alabama's Rep. Mike Rogers, being credited to a candidate.
• An automatic second check. Recent news quotes are checked by an AI model that asks one narrow question: did this person say these words? A quote it is sure belongs to someone else is hidden.
• Counted once. When several outlets carry the same quote, it counts once, and the other outlets are credited as "also reported by" (same person, 85% of the words in common, within 14 days).
• Posts that are not statements are left out. Fundraising appeals, posts signed by campaign staff, and pure courtesy posts (greetings, congratulations) are skipped.

What we calculate

• Rhetorical moves. A transparent word-and-phrase engine tags eleven moves: threat and fear, us versus them, labeling, deflection, absolutes, leverage, vague promise, self-promotion, phantom sourcing, deniability hedge and sincerity insistence. They are counted per 100 words, and the exact words that triggered each tag are shown. The engine reads English only.
• Audience impact, from −5 to +5. The likely effect of a statement on the audience it is aimed at: negative when it stirs fear, anger, blame or division; positive when it reassures, unites or offers credible progress; 0 when neutral. It is scored by an AI model with written instructions, and by the word engine when no AI score exists. It describes effect, not truth. On the tracker it is averaged over each candidate's last 30 days.
• Hostility, from 0 to 10. 65% from the density of hostile moves (fear, us versus them and labeling count fully; leverage 0.7; absolutes 0.4; deflection 0.3; short quotes are damped) and 35% from negative audience impact. It is compared with the same person's own norm, the average of their earlier statements over up to 180 days, shown only once there are at least 5 earlier statements spread over at least 3 days.
• Statement counts. "Reported statements" are news quotes, speeches and official releases from the last 30 days. Candidates' own social posts are shown separately ("+ own posts") and count a quarter when ordering races, so a candidate who posts a lot does not look more covered than one the press is quoting.
• Meantly read (Pro). An AI-written brief across everything a candidate said in the last 30 days: the lines they repeat, who they are addressing, how the tone is moving. It is labeled as AI-generated, quotes only words that appear in the statements, and follows the same neutral instructions for every party.

AI scores and briefs come from large language models accessed through OpenRouter, currently open-weight models. They can miss sarcasm, context and tone.

Coverage right now

63 of 66 races have statements from both sides.

Known gaps

• X (Twitter) is not collected. Its terms and paid access rule it out for us. Candidates who speak mainly on X, often Republicans, are covered only when others quote them. Bluesky, which we do read, is used more by Democrats; that is why own posts are shown and weighted separately.
• Some outlets block automated reading and are not included, for example Newsday, the Atlanta Journal-Constitution, PennLive and The Post and Courier. The States Newsroom network of statehouse sites is also blocked for now; we have asked for access.
• U.S. Senate websites block our server, so senators' press releases are missing. Their floor speeches still come in through the Congressional Record, but Congress sits little in October, so this source slows before the election.
• Candidates without an office or a news page depend on being quoted by reporters. They are the most likely to show few or no statements, and that says nothing about how much they campaign.
• Automatic attribution can still be wrong. We have found and removed misattributions caused by namesakes and tightened the rules each time. Every item links to its source so it can be checked, and we correct errors when told.
• Paraphrase and translation. Reworded versions of a quote may count twice, and the word engine reads English only.
• Race list. The 66 races are the competitive contests rated by the Cook Political Report in late September 2026.

Review, corrections and data

A person reviews the reads every week for verdict-like wording, disputed attribution, extreme scores and balance across parties, and corrects or removes what falls short. Meantly is independent and takes no money from parties, campaigns or governments. Journalists can ask for the underlying data for a race, or report an error, through the contact page.