RedBox Research, a polling operation affiliated with the political super PAC A Fight Worth Having, has published results from 30 surveys conducted between October 20 and 23, 2026. The release includes statewide polls and congressional district surveys across multiple states, with some statewide results incorporating respondents from district-level polls.

The organization’s methodology disclosure identifies A Fight Worth Having as the sponsor and RedBox Research as its polling unit. The super PAC is registered with the Federal Election Commission under identification number C00933085. That financial relationship is important context when evaluating the polling, particularly in closely contested races where small differences can attract significant attention.

RedBox says it drew participants from state voter files and invited them by text message to complete online surveys. However, only voters with usable phone numbers matched to their voter-file records could be contacted. People who could not be reached or who declined to participate are represented in the results through statistical weighting rather than direct interviews.

The most significant qualification appears in the organization’s own description of its likely-voter model. RedBox acknowledges that its model failed to meet a pre-registered accuracy standard when tested against past elections and explicitly labels the model “unvalidated.” Although the model was established in advance, that disclosure means the firm has not demonstrated that it meets its own stated standard for predicting likely voters.

RedBox also adjusts its results through multiple weighting formulas that incorporate factors such as party registration, age, gender, education, race and ethnicity, turnout estimates, and respondents’ recollections of their 2024 presidential votes. The firm compares as many as four weighting recipes, excludes those that fail specified checks, and averages the surviving approaches within two groups. It then uses 500 resampled versions of the respondents to determine how much influence each group has on the final estimate.

The method is designed to correct imbalances between survey respondents and the electorate the poll intends to represent. However, weighting can also substantially reduce a poll’s effective sample size, particularly when certain respondents receive considerably more weight than others.

Ohio’s statewide survey illustrates the issue. Although RedBox reports 2,998 completed interviews, its disclosed design effect of 6.48 reduces the effective sample to approximately 463 respondents. The reported margin of error is ±4.6 percentage points for a result near 50%. The organization’s published formula accounts for sampling error and the effects of weighting, but that margin does not capture every possible source of error, including nonresponse, coverage limitations or inaccuracies in the likely-voter model.

The surveys also use respondents’ recollections of their 2024 presidential votes as part of the weighting process, adjusting those reports toward certified election results. RedBox says its system compares alternative weighting approaches against voter-file information not used as weighting targets, including past turnout and registration history. Where available, it also evaluates early voting against official ballot-return records.

The firm says it publishes only polls with a margin of error of ±7 percentage points or less. It also acknowledges that the margin of error on the difference between two candidates is roughly twice the margin for an individual candidate’s share. Consequently, a reported lead of just a few points may not establish that one candidate is actually ahead.

Questions have also been raised about the authorship of RedBox’s methodology page. The page has reportedly received a 100% AI-written assessment from the Pangram detection tool. Such a score is not, by itself, proof that the material was generated by artificial intelligence, nor does it establish that survey responses or results were fabricated. RedBox’s published methodology says the surveys were completed by respondents themselves through online questionnaires, without an AI interviewer or chatbot.

The organization’s disclosure provides substantial technical detail, but that detail does not eliminate the need to examine its assumptions, sponsorship, sample composition and performance against actual election results. RedBox itself cautions readers to consider the sponsor’s role and acknowledges that sampling error is only one of several factors that can affect a poll.

With Election Day approaching on November 3, the central question is not simply how many people participated in these surveys, but how reliably the methods translate those responses into estimates of the electorate. The firm’s acknowledgment that its likely-voter model remains unvalidated is an important limitation that should accompany coverage of its polling results.