What We Include — Every Scientific Poll

The PollingSource average includes every scientific public poll of a race — we do not exclude pollsters. 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 we might disagree with. 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).

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.

Generic-ballot, governor, approval, and favorability polls include data from VoteHub, licensed under CC BY 4.0. Party-sponsored and internal polls are imported like any other scientific poll and labeled, never excluded.

The Average

A race's polling average is the unweighted mean of the most recent week of included polls whose fieldwork ended within the last 30 days, computed separately for the Democrat and Republican candidates. The most recent week is every included poll that ended within six days of the newest poll in that 30-day window. Older polls inside the 30 days stay on the table and stay out of the average. The margin is the difference between the two means. One poll in the window is the average.

Empty 30-day window: if no qualifying scientific general-election poll of the locked matchup ended in the last 30 days, the published figure is the newest qualifying poll of that matchup (n=1), not a blank and not a 30-day mean. The page says so. That figure does not set the rating. 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 the week, no sample-size weighting, no house-effect corrections. Simple means checkable — you can recompute our number from the polls in that week on the race page. If we ever add weighting, this page will document it before it ships.

The trend chart plots each poll. The published average, and the rating that comes from it, is the newest week described above.

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 at least one qualifying poll ended in the last 30 days, the map and race page show the band from that week's average (labeled "our polls"), including when the week 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 a poll from the last 30 days is in the average. When the 30-day window is empty, the number may still be the newest qualifying poll (see above) 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 exclusion or hidden weighting.

How the record is computed (ratings/compute_pollster_records.py, run 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.

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.