The House left Washington on September 16th and will not return before the November 3rd election. The Senate returns September 22nd but is expected to recess after finishing its late-September business. Campaign news will replace most congressional action, much of it built around polls.

That is a reason to understand polls, not a reason to trust them blindly.

Skepticism is reasonable. Since 2016, polls have missed important states, understated Donald Trump's support and produced confident-looking numbers that did not reflect the actual outcome. The mixed record has made all polling feel unreliable, even though some polls came close.

Before judging that record, it helps to know what a poll is and why political institutions need one.

A poll surveys a small group to estimate the views of a much larger group. A pollster might interview 800 likely voters, then adjust the sample to better reflect the expected voting population by age, race, education and past voting behavior.

The result is not a count of votes. It is an estimate of opinion during the dates when interviews took place. The American Association for Public Opinion Research calls an election poll a snapshot and warns against reporting it as a declaration that someone has won.

The familiar "plus or minus three points" is the margin of sampling error. It covers uncertainty from interviewing a sample, but not every problem. It does not fully cover people who refuse to answer, a bad turnout estimate, misleading wording or a late change in opinion.

Campaigns use private polls to learn which issues matter, which messages move undecided voters and which supporters need a reminder to vote. Polling guides candidate travel, advertising and spending in their attempt to win the election.

Parties use polls to compare races and place limited money where it might change a result. Donors use them to judge whether a campaign can win.

Newsrooms use polls to show what voters care about, which groups are moving and which races deserve reporting. Trouble begins when a small lead becomes a settled outcome or every new survey becomes a new story.

Polling is a tool, not a crystal ball.

Understanding why a campaign releases a poll helps a reader understand what the number is meant to do.

Campaigns keep most research private. They release selected results to attract donors, win party spending or convince reporters that a race is competitive. A two-point deficit becomes "within reach." A two-point lead becomes "surging."

An unfavorable poll can fuel a fundraising warning. A close poll can encourage turnout by telling supporters that their vote matters.

That does not make a campaign poll false. It means the release has a purpose. Ask who paid, what methods were disclosed and why the sponsor wanted it public. This can provide valuable information about the motivation behind the poll and it

We wanted to review accuracy: how close late polls came to the final margin and whether they identified the winner. That differs from measuring issues or movement during a campaign.

In 2016, national polls came close to the popular vote, which Hillary Clinton won by 2.1 points. Several decisive state polls understated Donald Trump, so national accuracy did not mean Electoral College accuracy.

In 2020, polling error grew. The AAPOR review of the election found that national presidential polls overstated Joe Biden's margin by 3.9 points on average and state polls overstated it by 4.3 points. The average absolute error was 4.5 points nationally and 5.1 points in state presidential polling.

Polling improved in 2024. Across 611 late presidential, Senate, governor and House polls, the average absolute error was 3.3 points, compared with 5.3 points in 2020 and 5.2 points in 2016. Yet the 2024 polls still leaned toward Democrats by an average of 2.7 points.

That record explains the distrust. Polls can identify the leader but make the race look safer than it is. Many polls can miss together because they use similar turnout assumptions or fail to reach the same voters.

Even a wrong poll can identify the issues driving voters or show that a safe race has become competitive. Ask what it measured well and what it missed.

1. Start with averages. They reduce the power of one unusual sample and work best when independent pollsters use different methods.

2. Put leads below five points in the uncertain column. Five is safer than a common three-point sampling margin, allows for uncertainty in both candidates' shares and leaves room for recent Democratic polling bias. It is a caution line, not a guarantee.

3. Check the interview dates. The field dates show when people answered. A poll released after a debate may have been conducted before it.

4. Ask who was surveyed. Adults, registered voters and likely voters are different groups. A likely-voter poll depends on correctly guessing who will participate.

5. Check who paid and what was disclosed. Campaigns choose which results to release. Missing wording, sample details or dates should lower confidence.

6. Match the poll to the election. National presidential polls do not measure the Electoral College. Senate control comes from states and House control from 435 districts, though the generic ballot can show national momentum.

7. When the polling is close, judge the candidates, the state and campaign resources. Experience, past performance, public image and the state's voting habits matter most when surveys do not point clearly in one direction. Fundraising shows whether a candidate can attract support. Cash on hand shows how much the campaign still has available for advertising, staff and voter turnout. Money does not guarantee victory, but a clear financial advantage can break a polling tie.

Our forecast starts with averages. Below five points, we consider the candidates, state history, election rules, fundraising and cash on hand. Available cash matters more than money already spent. We also allow several states to miss in the same direction.

On September 18th, the FiftyPlusOne generic-ballot average showed Democrats ahead 50.5% to 43.2%. That measures national momentum but does not assign seats.

We use the Democratic, Republican and toss-up totals from Cook, Inside Elections, Sabato, Decision Desk, Fox and Silver Bulletin. The six ratings ranged widely. We gave Democrats 53% of each source's toss-ups, calculated each seat total and averaged them. The result was 222.46 Democratic seats, rounded to 222. This is independent of the CBS estimate of 228.

Republicans began with 53 seats. We predict Democrats gain North Carolina, Ohio and Texas but lose Michigan, a net gain of two seats.

Georgia: Jon Ossoff, Democrat. Incumbent Ossoff faces Rep. Mike Collins. The average is Ossoff 49.5% to Collins 43%. Ossoff also held about $42.5 million on June 30th, while Collins held about $2.1 million in cash on hand. Polling and money agree.

Michigan: Mike Rogers, Republican. Physician Abdul El-Sayed, a former Detroit health director, faces Rogers, a former FBI agent and seven-term congressman who lost the 2024 Senate race by 0.34 points. El-Sayed leads 47.2% to 45%, inside our caution line. On March 31st, Rogers held $4.2 million to El-Sayed's $2.5 million. Rogers's experience, discipline, previous statewide performance and earlier cash advantage decide the call.

New Hampshire: Chris Pappas, Democrat. Four-term Rep. Pappas faces former Sen. John Sununu, who lost in 2008. Pappas leads 46.3% to 43%. Both have federal experience, but Pappas's recent wins in a competitive district and fundraising lead give him the edge: Pappas has raised $15.4 million to Sununu's $4.2 million.

North Carolina: Roy Cooper, Democrat. The former two-term governor and four-term attorney general faces former Republican National Committee chair Michael Whatley, who has not held office. Cooper leads 48% to 39.2% and held $18.5 million to Whatley's $2.5 million on March 31st.

Alaska: Dan Sullivan, Republican. Former Rep. Mary Peltola won statewide in 2022 but lost in 2024. Sullivan is a two-term senator. The average has Peltola ahead 46% to 44.5%, but the polls disagree. Unreliable polling, ranked-choice voting and another Dan Sullivan add uncertainty. Sullivan's two wins and campaign cash decide the call.

Ohio: Sherrod Brown, Democrat. Brown, a former three-term senator who lost in 2024, faces Sen. Jon Husted, who was appointed in 2025 after serving as lieutenant governor. Brown leads 47% to 43% and raised $38.6 million to Husted's $14.3 million. Brown's lead, money and statewide record outweigh Ohio's Republican lean.

Maine: Susan Collins, Republican. Five-term Sen. Collins faces former Maine Senate president Troy Jackson. Jackson leads 46.7% to 45.3%, but Collins had more than $11 million while Jackson entered late. Collins won by nine points in 2020 after trailing late polls. Her money and history decide the call.

Iowa: Ashley Hinson, Republican. Rep. Hinson faces state Rep. Josh Turek, a Paralympic gold medalist. Hinson leads 46.2% to 44.2%. On June 30th, Hinson held $6.75 million while Turek held $2.36 million. Her federal experience, polling edge and cash advantage align.

Nebraska: Pete Ricketts, Republican. Ricketts, the incumbent and former governor, faces independent Dan Osborn, who lost a 2024 Senate race by about six points. The two-poll average is tied at 47%. Ricketts held $6.91 million on June 30, several times Osborn's earlier cash. Experience, money and Nebraska's Republican record decide the call.

Texas: James Talarico, Democrat. State Rep. Talarico faces Attorney General Ken Paxton, a three-time statewide winner impeached by the Texas House and acquitted by the Senate. Talarico leads 47.4% to 44.4%, inside our caution line. Paxton has more experience, but Talarico raised about $68.6 million and held $21.5 million in cash by June 30, compared with $9.2 million raised and $1.8 million in cash for Paxton. Polling, cash and Paxton's liabilities decide the call.

Prediction: Each party wins five of the ten races examined. Democrats gain North Carolina, Ohio and Texas while Republicans gain Michigan, moving the Senate from 53-47 to 51-49 Republican. Michigan and Texas show why a close poll is only the start of the analysis.

Keep an eye out for an analysis of our predictions after the election. Once the results are in, we will transparently explain what we got wrong.

Readers do not have to choose between believing every poll and ignoring all of them. A poll can be useful evidence without being a verdict.

The practical approach is to demand transparency, use averages, allow at least five points of caution and examine the candidates, state and campaign resources when a race remains tied. Then return to the results. After every race is final, we will publish our score, identify every miss and explain whether the problem came from the polling, our judgment or both.