1. Voter-file sample

Each poll's sample is drawn from the state's voter file. We divide the file into groups and draw voters at random within each group, and we record each voter's chance of being drawn. That chance sets their starting weight: a voter who was less likely to be drawn counts for more.

We can invite a voter only through a usable phone number matched to their record. A voter without one cannot be reached through these invitations. A statewide poll can also count the respondents to our district polls in the same state. Each poll's Methodology tab lists the samples in it.

2. Text to web

We invite voters by text message to answer a survey on the web through their own link. They answer on their own. Each poll's Methodology tab lists how its interviews were done, the language, and its first invitation and last interview in Eastern time.

3. Likely voters

No question screened respondents out of these interviews. For weighting, each respondent starts with the voter file's estimated chance of voting, updated by their stated likelihood of voting. Someone who says they already voted counts as certain to vote. A missing likelihood answer keeps the file's score. A zero file score otherwise receives no weight. We multiply the starting weight by this estimated chance before adjusting it to the weighting targets.

Our likely-voter model was set in advance. In tests against past elections it did not clear its pre-registered accuracy bar, so we describe it as unvalidated. Each poll's Methodology tab records that status and explains where its electorate targets come from.

4. Weighting

We adjust weights to the poll's electorate targets and to the certified 2024 presidential result, using respondents' reports of their 2024 vote. Depending on the poll and recipe, targets include party registration, gender, age group, congressional district for statewide polls, four-year college education, race and ethnicity, and turnout-score groups. Each poll's Methodology tab lists its recorded targets and compares them with the weighted sample.

Other organizations' race polls and polling averages do not enter these estimates.

5. The house estimate

We compare up to four weighting recipes in two families. Both use the certified 2024 presidential result. One adjusts the balance between people who recall voting for Harris and Trump to their certified two-candidate shares, while keeping other and missing answers at their starting-weight share. The other also checks whether the voter file records a 2024 vote. Among those recorded voters, named candidate choices follow the certified Harris, Trump and other-candidate shares. Other or missing answers keep their starting-weight share within that group. People without a recorded 2024 vote are held to their share of the likely electorate.

A recipe is left out if its weighting does not converge, if any party-registration group is more than 3 percentage points from its target where that target exists, or if Harris's share of Harris-plus-Trump recall is more than 5 points from the certified two-candidate result. Only the surviving recipes enter the estimate. Sometimes only one family remains.

Within each family, we put the surviving recipes' weights on the same scale and average them. We then draw 500 samples, with replacement, from the respondents and repeat the weighting. Each family is scored against voter-file facts that were not weighting targets, such as past turnout, registration year and county. Where official returned-ballot records cover the poll, the score also checks early voting by party. A family's share of the final estimate is how often it scores best among the resamples in which every surviving family can be scored. Ties split the credit. Each poll's Methodology tab shows which recipes entered and each family's share.

6. Margin of error

The margin of error is given at 95 percent confidence for a result near 50 percent, and it includes the design effect, the precision lost to weighting: 1.96 × √(design effect × 0.25 ÷ interviews) × 100. The margin on the gap between two candidates is roughly twice the margin on one share. Each result says whether its lead is within the margin of error.

Our publishing standard. We publish a poll only when its margin of error is ±7.0 points or less. A poll above that limit is held back from the current releases.

7. Small groups

In the crosstabs, each group is judged by its own number of interviews. Below 50, its cells, interview count and margin of error show only “n<50”. From 50 to 99, the interview count carries a “†”, and the legend says “interpret with caution”. For groups with at least 50 interviews, the column shows its interview count and a margin of error calculated from that group's own weights. A subgroup's margin of error is usually larger than the poll's, but not necessarily.

Sponsor and limits

Sponsored by A Fight Worth Having, a super PAC (FEC C00933085). Conducted by RedBox Research, AFWH's polling unit. Weigh the results with that sponsorship in mind.

Sampling error is one source of error among several. Who we could reach, who chose to answer, question wording and order, and our likely-voter and weighting choices can also move results. People who can't or won't answer a web survey are represented only through weighting. Our polls are separate state and district surveys, not a national sample.

Reading our results

Lead words. Using the lead rounded to one decimal place, under 1 point is tied, 1 through 5 points is a narrow lead, and over 5 points leads.

Undecided stays undecided. Undecided answers remain in the headline results. We show leaning follow-up answers separately and do not add them to candidates' headline totals.

Party in subgroups is the party respondents said they identify with, or their voter-file registration where they didn't say.

Counts of polls, not pooled numbers. When we count polls, each poll counts once; we never average polls into one figure.

Earlier polls (Rhode Island, Delaware, NY-15, AZ-04, MO-01) used the methods printed on their own pages.