About AllPollsUSA.com

AllPollsUSA.com brings together United States voting-intention polling in one place: the generic congressional ballot (“if the election were held today, would you vote for the Democratic or the Republican candidate for Congress?”) on the home page, and the presidential head-to-head on its own page. Both are shown as vote share, in interactive charts, with every polling firm side by side.

AllPollsUSA.com is independent and not affiliated with any polling firm.

Messages and privacy. Messages you send us on social media (including direct messages on Bluesky) may be processed automatically so they can be answered with a standard reply. The text of your message may be passed to our AI processor (Anthropic) as part of that. Messages we cannot answer automatically are read personally by the site's administrator.

What this site adds

Our database holds 2,788 polls going back to 1998. We bring them together and work out what no single poll shows.

How the average is calculated

The average quoted on the home page and on every party page is the mean of the most recent poll from each polling firm that published in the 12 months up to the newest poll. Each firm counts once and equally, however often it polls. We do not weight by sample size, population or method, and we make no forecast.

For the charts, each firm’s polls are joined into a line: between two polls from the same firm less than 122 days apart we interpolate in a straight line; across longer gaps, and beyond a firm’s latest poll, we draw nothing. In the overview each party’s line runs midway between its lowest and highest firm, with that range shaded; on a party’s own page every firm has its own line. How to read the charts

How accurate each polling firm has been

How this is measured: for each past election we take each pollster’s last poll in the 30 days before polling day and compare it, party by party, with the result. The figure is the mean absolute difference across the parties that appear in both. Only parties the pollster actually reported are counted. A late shift in opinion is not the same thing as a polling error. For the generic congressional ballot, polls and results are both compared as two-party shares: each party’s share of the combined Democratic and Republican total. Undecided and third-party respondents therefore do not count as polling error, and the result figures in these comparisons are two-party shares too, not the raw popular vote. Elections assessed so far: 15. See the accuracy comparison

How often the data is updated

Our system re-reads its sources every 15 minutes. Newly added polls are picked up automatically, and the averages and charts are recalculated at the same time.

Polling firms covered

Polling firms with at least one poll in the last 12 months, each with its own page:

Every polling firm, including earlier ones, is listed in the overview. All polling firms

Sources and licenses

Two licenses apply to the data on this site, and each row in the table below names the material it covers. The presidential series is CC BY-SA 4.0 (Wikipedia). The generic-ballot series carries both: its history comes from CC BY 4.0 sources (the FiveThirtyEight archive and VoteHub), and polls from July 2026 on come from Wikipedia under CC BY-SA 4.0. Each license applies to its own material, a page that shows material from both credits both, and nothing here claims a single license over the whole site. There is no combined download.

Source	What it covers	License	Source link	License text
Wikipedia contributors	Generic congressional ballot, July 2026 onward	CC BY-SA 4.0	en.wikipedia.org	CC BY-SA 4.0
VoteHub	Generic congressional ballot, December 2024 to September 2026 (historical)	CC BY 4.0	votehub.com	CC BY 4.0
FiveThirtyEight / ABC News	Generic congressional ballot, 1998–2022	CC BY 4.0	github.com/fivethirtyeight/data	CC BY 4.0
FiveThirtyEight / ABC News	Generic congressional ballot, November 2022 to November 2024 (the 2024 cycle; archived copy of November 27, 2024)	CC BY 4.0	web.archive.org	CC BY 4.0
Wikipedia contributors	Presidential head-to-head polling, 2020 and 2024 cycles	CC BY-SA 4.0	en.wikipedia.org: 2020, 2024	CC BY-SA 4.0

How current the generic-ballot data is

New generic-ballot polls are read from the polling table on English Wikipedia’s 2026 United States elections page, which volunteer editors maintain and cite. The hand-over date is July 1, 2026: we add a poll from that table only when its fieldwork ended on or after July 1, 2026. Everything before that date comes from VoteHub and FiveThirtyEight alone, even where the Wikipedia table lists the same poll, and the few polls VoteHub delivered after that date stay as VoteHub recorded them. We check it every fifteen minutes, so a poll usually appears here within days of its release — as soon as an editor has added it, and not before. Because the source is a volunteer-edited encyclopedia, a later correction there is picked up here the same way. Polls we already held from our historical sources (VoteHub and FiveThirtyEight) keep the figures and the population label those sources recorded, even where the Wikipedia table now prints different figures or a different population for them. A Wikipedia row that matches one of those polls to within a point, up to four days apart, is taken to be the same poll and is not added a second time; each of those polls is matched with at most one Wikipedia row, the nearest in date.

Where a poll was commissioned by a news organization and fielded by a polling firm (The Economist/YouGov, Reuters/Ipsos, The New York Times/Siena and the like), we credit it to the firm that fielded it, so each firm’s page holds its polls whoever paid for them. When one firm publishes two polls ending on the same day, we keep the one with the larger sample (or, where a sample size is missing, the one whose listed source comes first alphabetically), so a firm’s page shows one poll for that day. If a larger poll by the same firm ending the same day is added to the table later, it replaces the one we showed. A poll run jointly by two firms keeps the single joint name it has always carried in our data (the Fox News poll, conducted by Beacon Research and Shaw & Company, is one example).

Likely voters, registered voters and adults

Polling firms report their results for a specific population: LV (likely voters), RV (registered voters) or A (all adults). Where a poll reports more than one, we keep a single headline figure per poll, preferring likely voters, then registered voters, then adults — the same order US poll aggregators use. That preference applies to the polls we read from Wikipedia; polls we took over from VoteHub or FiveThirtyEight keep the population those sources recorded. Where a firm reports the same population twice, with and without voters who lean towards a party, we keep the version the Wikipedia table marks as its headline (the one including leaners); if the table does not say which is which, we leave that poll out rather than guess. We record the label with the poll wherever the source states it, which means the generic-ballot polls: both their historical source and the Wikipedia table they now come from name the population. A handful of polls report only “voters”, which is none of the three; they are kept and stored without a label. The presidential tables on Wikipedia carry no population column at all, so the presidential polls are stored without one. The presidential page is built to print the label next to each polling firm and will start doing so for any presidential poll that arrives with one.

Two honest consequences. First, firms typically switch from registered voters to likely voters as an election approaches, and the two populations do not give the same number, so a series can show a small step at that switch. We show that step and explain it; we do not adjust it away. This site publishes polls, not models. Second, the labels exist for the November-2022-onward data only: the 1998–2022 archive we use records no population column at all, so those polls carry no label rather than a guessed one.

For the generic-ballot polls since December 2024, the mix is currently 306 registered-voter polls, 193 likely-voter polls, 27 all-adult polls, and 3 where the source states no population.

What the generic-ballot history looks like

The 1998–2022 archive is a late-campaign collection: it holds the polls taken in the roughly two months before each election and very little in between. The history is therefore a series of dense pre-election clusters with near-empty years between them — that is the shape of the source, not a gap in our processing.

The archive stops after the 2022 midterms. The two years that follow come from FiveThirtyEight’s own file of 2024-cycle generic-ballot polls, as archived on November 27, 2024 (CC BY 4.0), read with the same rules as the VoteHub polls: likely voters preferred over registered voters over adults, and partisan and internal polls left out. That file runs from November 2022 to the eve of the 2024 election, so there is one short, real gap: between the November 2024 election and December 11, 2024, when our VoteHub data begins, we have no generic-ballot polls. Nothing was lost or dropped; we would rather show the gap than fill it with data we cannot license.

What the presidential page covers

The presidential series covers the 2020 and 2024 cycles. Earlier cycles (2008, 2012, 2016) are not included: in the source tables for those years the candidate columns are not machine-identifiable the way later cycles are, and rather than guess which column belongs to which candidate we have left them out until that can be read reliably. They are deferred, not dropped.

For each cycle we show the head-to-head between the actual major-party nominees, from the date both had clinched their nomination onward, plus any third-party candidate appearing in the same table. Polling taken before that date is a hypothetical matchup rather than a choice voters could make, and primary polling is a different question entirely; neither is included. In 2024 the nominee changed mid-cycle, so the chart shows one line ending in July and another beginning — which is what happened.

One further exclusion worth naming. The 2024 source also holds separate five-way tables (Harris vs. Trump vs. Kennedy vs. West vs. Stein). We do not merge those into the series: most of their rows are the same polling firm on the same date as a row in the two-way table, and one reading per firm per day is all this site can store — merging them would make the stored figure flip between the two-way and five-way numbers from one update to the next. The presidential page therefore shows the two-way head-to-head, at the cost of the few polls that appear only in the five-way tables.

Polls we exclude

Our historical generic-ballot sources flag polls conducted for a party, campaign or candidate, and those partisan and internal polls are excluded from that history: VoteHub and FiveThirtyEight’s 2024-cycle file flag both “partisan” and “internal” polls, the 1998–2022 archive a partisan flag only.

Wikipedia, the source of new generic-ballot polls and of the presidential series, does not identify sponsored polls. The generic-ballot table there marks a polling firm’s own party affiliation with (D) or (R) next to its name, on every poll that firm runs, whoever commissioned it — that says who the firm usually works for, not who paid for a given poll. So we keep those polls, credit them to the firm without the suffix, and do not label any individual poll as partisan. A poll published as a campaign’s internal poll can therefore appear in the generic-ballot series from July 2026 on without being marked as one.

Polling averages and forecasts — RealClearPolitics, model-based averages, and similar — are also excluded. They are other people's analyses of the same polls, not polls.

Not covered

Presidential approval, state-level polling, Senate, gubernatorial and district races, primary and hypothetical-matchup polling, and aggregator averages are all out of scope. This site covers two national vote-intention questions and reports them carefully, rather than covering everything shallowly.