How VoteCow counts comments
VoteCow counts public YouTube comments under news videos about 2026 races. It does not poll anyone, does not predict results, and never shows who wrote a comment.
What is counted
- YouTube's top-sorted comments, at most 300 per video, fetched through our scraper. Videos without a transcript never enter our data.
- A video counts for a race only if its title or transcript names a candidate or the race. Videos from channels we have not sourced are left out and listed as excluded.
- Comments on campaign-owned channels are never counted.
- A comment's place is the place of the video it sits under, never the commenter's location.
- One person can post many comments; we do not de-duplicate by person.
- Our first seed videos came from searches built around candidate names; later searches add race and station searches.
Candidate share
The share of counted comments our matcher links to any candidate. One algorithm for every candidate: full names match as phrases; surnames match as words (possessives stripped) and with small spelling slips (one letter for names up to 7 letters, two for longer, same first letter). Common-word surnames match only as full names. These words are never treated as a misspelled name: pardon, parton, pastor, patron, paxson, taliban. First names alone are not matched. We never publish a number for any single candidate.
Issue share
The share of counted comments that contain at least one word from our lists for the 10 issues Pew Research Center asked voters about (Pew, 2024). The lists contain no candidate or party words.
| Issue | Words |
|---|---|
| abortion | abortion, abortions, pro choice, pro life, pro-choice, pro-life, reproductive, roe |
| climate change | climate, climate change, emissions, global warming, renewable, renewables |
| economy | affordability, affordable, budget, cost of living, debt, deduction, deductions, deficit, diesel, economic, economics, economy, gas prices, groceries, grocery, housing, inflation, minimum wage, oil prices, price, prices, recession, rent, spending, tariff, tariffs, tax, taxes, trickle down, unemployment, wage, wages |
| foreign policy | china, foreign aid, foreign policy, gaza, iran, israel, nato, russia, troops, ukraine, venezuela |
| gun policy | 2nd amendment, firearms, gun, gun control, guns, second amendment, uvalde |
| health care | aca, health care, healthcare, hospital, hospitals, insurance, medicaid, medical, medicare, obamacare, prescription, prescriptions |
| immigration | asylum, border, deport, deportation, deportations, ice raids, illegal aliens, illegal immigration, illegals, immigrant, immigrants, immigration, migrant, migrants |
| racial and ethnic inequality | civil rights, dei, inequality, racial, racism, racist |
| supreme court | justices, scotus, supreme court |
| violent crime | crime rate, homicide, murder, murders, shooting, shootings, violent crime |
How well the definitions work
On 2026-10-06 we checked the definitions against comments labelled blind by a fresh-context AI annotator (Claude, 'sonnet' tier, Claude Code), not a human. The labeller never saw our word lists or matcher. These are AI labels, not human ones.
| Sample | Issue words: found / right | Candidate matcher: found / right |
|---|---|---|
| 200 comments (seed 20261007) | 44.4% [25–66%] / 80% [49–94%] | 85.1% [76–91%] / 100% [95–100%] |
| 172 comments (seed 20261006) | 55.6% [34–75%] / 62.5% [39–82%] | 91.2% [83–96%] / 100% [95–100%] |
"Found" is recall (of the comments the labeller marked, how many our definition caught); "right" is precision (of the comments our definition caught, how many the labeller agreed with). Brackets are 95% intervals. Labelled issue talk was 9% and 10.5% of the samples; the word lists caught 5% and 9.3%, so true issue talk is likely higher than our issue share. The matcher made no wrong match in either sample but missed some, so the true candidate share is higher than ours.
Polls
VoteCow runs no polls. When a pollster posts a new Senate or governor poll, we fetch the pollster's own release and read its numbers twice: once from the main question and once from the table of results by group. A poll's numbers appear only when both readings agree exactly. Until then the poll is listed as not yet verified, with no numbers. Each poll shows who ran it, when, how many people were asked, whether they were likely voters, registered voters or all adults, the margin of error, and who paid for it, or "not stated in the release".
Posts on X
We read public posts on X that name a race's candidates, as X's search shows them to our logged-in account, plus each candidate's own campaign posts. Posts are not people: one person can post many times, and bots post too. Likes on a candidate's own posts measure an audience on X, not votes. A post is featured only if it, or the post it quotes, names one of the race's candidates: X's Top search also returns popular posts that are merely related. "Most-engaged posts" ranks those posts by likes plus twice the reposts plus quotes, one post per account, and leaves out the candidates' own accounts. "What people are sharing" lists news links in those posts from the last day, ranked by how many different accounts shared each one; the headline is the publisher's own. Featured posts show the account's display name, the post's own text, the post it quotes, and its first photo or video still, loaded from X's image servers, with a link to the post on X. A post is never featured if its text, the post it quotes or the account name contains a slur, a claim that the election is rigged or stolen, or instructions about when, where or how to vote, or if X no longer shows it. Media X marks as possibly sensitive is not shown. We call a post "ratioed" only when it has at least twice as many replies as likes. "Posts per hour" is an estimate: our search reads X's newest posts in short passes, and each hour's figure is the posts we saw divided by the minutes our passes covered. An hour we couldn't measure is left blank, never shown as zero. "Busiest hours" are the three highest hours at least four hours apart, each with the most-liked post we found from inside that hour. Videos play from X's own servers, muted until you turn the sound on; to let them load, VoteCow pages send no referrer to any site. The videos page shows the most-liked featured posts with a video from the last three days. A story page is a news headline that at least eight accounts posted about; once a story page is published it keeps its address and headline, and later posts about the same moment join it.
What we log
To learn which posts readers find worth their time, the site sends a few anonymous events: the page you opened, the site that sent you (its name only), which featured posts were on your screen for a second or more and for how long, which ones you opened on X, and which busiest hour or tab you picked. Each event carries a random ID made by your browser and kept only in it. We store no IP address, no cookies, no account and nothing you type. If your browser sends Global Privacy Control or Do Not Track, nothing is sent at all. "Read here" ranks featured posts by how many readers stayed on them for three seconds or opened them, and appears only once at least 20 readers count. It never shows a number, and VoteCow never asks who you support.
How often the site updates
The polling average
The map and the race poll pages use VoteCow's own average of polls collected by VoteHub (votehub.com, licensed CC BY 4.0). For each race we take each pollster's latest poll from the last 30 days: the likely-voter version first, then registered voters, then adults. Campaign internal polls are left out, and a poll counts only if it asks about a candidate from each major party in the race. A candidate's average is the plain mean of their numbers in those polls. The lead decides the shading: under 2 points is a Toss-up, 2 to 6 Lean, 6 to 12 Likely, 12 or more Safe. The Senate seat count keeps the seats not up this year with their party, gives each race with a recent average to its leader unless it is a Toss-up, and counts a race without recent polls for the party that holds the seat now. It is a reading of the polls, not a forecast.
New posts on X about the Texas Senate race are read every 15 minutes, and the most-engaged posts about every other race every 2 hours. New YouTube videos and comments are added one race per hour. The site rebuilds about every half hour; each page shows when its data was last updated. Every rebuild is checked before it goes live, and nothing is published if a check fails.
When a page or group is shown
- A race page needs at least 5 videos with comments from at least 3 different owners, at least 1 local TV station video, and at least 300 counted comments.
- A group of channels (local stations, national, independent creators) is shown only with comments under at least 2 different owners and at least 150 comments. If one owner has more than two thirds of a group's comments, the group's name says so.
- All shares are raw counts. Each figure shows its comment count, video count and owner mix.