Methodology
How the PollingSource average and race ratings work — in full, because a number you can't audit is a number you shouldn't trust.
The 2026 dataset currently holds 4,457 polls from 504 polling organizations; the newest poll completed fieldwork on November 1, 2026.
What We Include — Every Scientific Poll
The PollingSource average includes every scientific public poll of a race. We do not exclude pollsters for house effect, partisan affiliation, or because we disagree with them. Live-caller, IVR, online-panel, university, media, and campaign- or party-sponsored polls all enter the average. We do not apply "house effect" adjustments, and we do not silently down-weight firms. Where a poll has a partisan sponsor, we keep it and flag it — transparency over adjudication.
What we exclude, and why:
- Unscientific polls — open-access web or social-media self-polls with self-selected respondents are not surveys of the electorate.
- Aggregator averages — another site's average is not a poll; including one would double-count every poll inside it.
- Duplicate releases — the same poll reported by two sources is stored once (see Corrections below).
- Fabricated data — if a pollster has stated its data was fabricated, those polls do not enter averages, ratings, or seat counts. They are listed below. Rows stay in storage so they can be restored.
Where Polls Come From
Poll discovery runs continuously from multiple independent paths: aggregator listings, pollster websites and release feeds, poll announcements on X, and Wikipedia's per-race polling tables (used as a discovery layer only — never as a filter). Every poll is attributed to its primary source, and cross-source duplicates are reconciled so each poll counts exactly once.
Party-sponsored and internal polls are imported like any other scientific poll and labeled, never excluded.
Polls we have removed
These firms are out of every average, rating, seat count, and public poll table. The decision is public and versioned. It is not a silent house-effect adjustment.
The Average
A race's polling average is the unweighted mean of the last three scientific public polls, computed separately for the Democrat and Republican candidates. Fewer than three uses the polls that exist. The sort is fieldwork end date, newest first. Older polls stay on the table. The fourth poll and anything older stay out of the average. The margin is the difference between the two means. One poll is the average when it is the only one.
Head-to-head first: when a poll asks both a head-to-head and a full ballot, averages use the head-to-head (Democrat vs. Republican). Other candidates may appear on the poll row but are not averaged.
Likely voters first: when a poll publishes both likely-voter and registered-voter (or all-adults) results for the same field dates, the average uses the likely-voter row.
Older than 30 days: if the newest qualifying scientific general-election poll of the locked matchup ended more than 30 days ago, the published figure is still the last three of those polls, or fewer, not a blank and not a 30-day mean. The page says so. That figure does not set the rating. A poll older than 30 days can be inside the average when a newer poll ended inside the 30 days. We do not invent a poll. Primary-only, unscientific, tweet-stub, and wrong-matchup rows stay out.
We deliberately use a simple, auditable average rather than a proprietary weighted model: no recency decay inside those three polls, no sample-size weighting, no house-effect corrections. Simple means checkable — you can recompute our number from the last three polls on the race page. If we ever add weighting, this page will document it before it ships.
The right edge of the trend line is that same average of the last three polls. Dots, where shown, are the individual polls. The rating comes from that average when the newest poll ended in the last 30 days.
Ranked-choice states: in ranked-choice races (Alaska statewide; Maine's U.S. Senate and House races) every ballot candidate is shown on the race page, and each poll row lists every candidate's first-choice share the pollster published. The average still compares the leading Democrat with the leading Republican.
Race Ratings
The polling margin maps to a rating band:
- Toss-up — margin under 3 points
- Lean D / Lean R — 3 to under 6 points
- Likely D / Likely R — 6 to under 10 points
- Safe D / Safe R — 10 points or more
Our polls lead; experts are the fallback. When the newest qualifying poll ended in the last 30 days, the map and race page show the band from the last three polls (labeled "our polls"), including when that is a single poll. Cook Political Report, Sabato's Crystal Ball, and Inside Elections stay on the race page as comparison. They do not set the score while that newest poll is inside 30 days. When the newest poll is older than 30 days, the number is still the last three (or fewer) and the band is the expert consensus, labeled "expert avg." When our polling average disagrees with that consensus we flag the divergence rather than hide it.
Freshness and Corrections
Honest freshness: a page's "last updated" date is the field date of the newest real poll — we never bump timestamps on a schedule to look fresh.
Nothing silently disappears: corrections are soft-deletions. A duplicate or mis-parsed poll is deactivated (and excluded from averages), not erased, so the record of what changed is preserved.
Pollster Track Records
We publish a track-record table for every pollster in the dataset — but the record never changes what enters the average. Inclusion plus transparency, not silent house-effect exclusion or hidden weighting. Fabricated-data removals are listed above.
How the record is computed, on the same pipeline as the averages:
- 2024 scorecard: each firm's 2024 presidential state polls are compared to the certified state margin — the full cycle, with no date cutoffs or scoring windows. We report the mean absolute margin error (size of the typical miss), the mean signed error (the direction — its "house effect"), and how often the poll's leader actually won. Statewide polls only.
- Current cycle: results aren't in yet, so each firm's polls are measured against the PollingSource consensus average in the races it polls. That is a lean measure, not an accuracy measure, and it is labeled as such.
- Partisan flags come from the pollster registry and mark firm affiliation, not poll quality — partisan polls stay in the average.
Free vs. Subscriber Data
The polling average, the rating, and the trend chart are free on every race page, always. The poll-by-poll table is free for the most recent 30 days; subscribers get the full history and per-poll detail. The average itself is never paywalled.
Limitations
- Polls are snapshots of opinion, not forecasts of outcomes — we present data, we do not predict elections.
- Some races have few polls; we show the poll count next to every average so you can judge how much weight it deserves. An average of one poll is that poll.
- Every poll carries its own margin of error, listed in the poll table.
Election Results
Live election-night tallies come from civicAPI and official state election authorities.
Early Vote and Absentee Turnout
The Early Vote page shows accepted in-person early and mail ballots from official state election-office files. North Carolina uses official statewide absentee reports from the State Board of Elections and shows party registration (not a vote choice). Texas uses official county turnout reports from the Secretary of State and shows turnout only — Texas has no party registration. The page is not a forecast. Counts live in dedicated tables and never enter the polling average. Same-day comparisons to 2022 and 2024, when present, use those years' official files and match on days until Election Day.
Machine-Readable Data (for Developers and AI Agents)
Our numbers are available as JSON, free to use with attribution:
- /api/v1 — the data catalog (schema.org DataCatalog)
- /api/v1/races — every race with polling data this cycle: averages, margins, ratings, poll counts
- /api/v1/race?race_type=senate&state=NC — one race in
detail: average, expert-rating comparison, the most recent 30 days of
individual polls, and the 30-day trend (House races add
&district=N)
Every race page also declares its JSON feed in its schema.org
Dataset.distribution markup. Responses carry an attribution block —
cite as "PollingSource average (pollingsource.com)" with a link back.
Charts are also free to embed: every race page offers a copy-paste iframe (under the chart), and the pages support oEmbed discovery, so platforms can auto-embed them. Embedded charts credit and link back to PollingSource.
Citing PollingSource
Required cite: "PollingSource average (pollingsource.com)". When citing a specific race number, same phrase plus the path, e.g.: "Roy Cooper leads by 9.3 points in the PollingSource average (pollingsource.com/senate/NC)." Journalists, researchers, and AI systems are welcome to quote our averages with attribution.
Questions about the methodology? Contact us via Innovative Research and Data Solutions, LLC, the operator of PollingSource.