By the Wavefield Research team · Published Oct 15, 2026
We followed the public citation trail for 141 poll documents covering 128 distinct polls across 26 Senate and governor races, and checked each one against eight basic disclosure elements drawn from professional polling standards. The short version: the basics are usually there, and the parts that matter most often are not.
of readable poll documents disclose all 8 basics
say where their sample came from, the weakest element
of public citations lead to documents a reader cannot open
Of the 141 cited documents, 82 could be read in full. Each was coded for whether it states the element itself; a number that exists only in secondhand coverage does not count. The percentages below are shares of those 82 readable documents.
who was surveyed (likely voters, registered voters, adults)
how many people answered
when the poll was in the field
a stated precision figure
how answers were collected (online, phone, text, mixed)
who paid for or commissioned the poll
whether and how the data were weighted
where respondents came from (voter file, named panel, RDD)
The median document discloses six of the eight. Twenty two documents disclose all eight. One discloses none: horse race numbers published with no dates, no sample size, and no population, by the polling firm itself.
The pattern across 82 documents is consistent: pollsters reliably publish the numbers a news story needs (population, sample size, dates, margin of error all land near or above 90 percent) and far less reliably publish the facts a reader needs to judge those numbers. Where the sample came from is disclosed in 43 percent of documents. Whether the data were weighted, 57 percent. Who paid for the poll, 60 percent.
The sampling frame is the sharpest gap, and it is not a technicality. A sample of likely voters drawn from a voter file, a sample from a named opt-in panel, and a sample from an unnamed source can produce the same headline with very different reliability, which is why professional standards treat the frame as a core disclosure. Most 2026 midterm poll documents leave it out. Our guide to sampling methods covers what each answer would mean.
The gaps have recognizable shapes. Strategy memos addressed “To: Interested Parties” state dates and sample sizes while omitting the client who paid, as a genre. One firm published 41 pages of crosstabs with no methodology statement at all; its data collection mode is discoverable only because the mode question appears in the crosstabs. And one release gives a sample size in its headline that differs from the one in its own methodology paragraph, 812 against 800.
Before a poll document can disclose anything, a reader has to reach it. For 59 of the 141 cited documents (42 percent), our automated readers could not: news paywalls, bot walls on the hosting sites, pages whose content only loads inside a browser, image-only PDFs, login walled tweets, and at least one citation that now points at a dead file sharing link. A few citations point at a bare homepage where the poll no longer appears at all.
Two fairness notes belong next to that number. First, a document our tools could not open is counted as not assessed, never as nondisclosure; the scorecard above covers only documents we read in full. Second, some of these sources open fine for a human with a browser, so 42 percent is a measure of friction in the public record, not of pollster secrecy. It still matters: the citation trail is how the public record of polling gets checked, and a trail that dead ends in paywalls and broken links cannot be checked. Archives rot fastest exactly when a cycle ends and nobody is looking.
Twenty two documents disclose all eight elements, and the firms behind them deserve the namecheck. Two stand out for doing it repeatedly:
5 of 5 releases
every readable release disclosed all 8 elements; its NetChoice topline names the panel exchange, the matched voter file, and per state margins
5 of its readable releases
all 8 elements in releases across Michigan, Texas, South Carolina, Florida, and Minnesota
2 of 2
both Maine releases fully disclosed
2 of 2
Michigan and Wisconsin releases fully disclosed
Single fully disclosed releases also came from Beacon Research with Shaw & Co., Zenith Research, Overton Insights (whose methodology note explicitly invokes the AAPOR Transparency Initiative), Tavern Research, Siena College, The Public Sentiment Institute, SurveyUSA, and SoCal Strategies.
Two of those are worth a sentence each. The SurveyUSA poll reached full disclosure through a television station's news article, proof that coverage can carry a complete methodology when the pollster provides one. And SoCal Strategies, a crowdfunded operation publishing on Substack, disclosed everything down to its voter file, its weighting variables, and its design effect. Disclosure is a choice, not a budget item.
The sharpest illustration in the dataset is one poll published twice. An Echelon Insights survey appears once as the pollster's topline PDF, which names the panel exchange, the matched voter file, the weighting dimensions, and per state margins of error, and once as the sponsor's press page, which reports the results and discloses almost nothing. Same poll, same numbers, opposite transparency. Which version a reader meets depends entirely on which link gets passed around.
Corpus. We collected every poll listed in the polling tables of the Wikipedia pages for 40 races in the 2026 United States midterm elections (20 Senate, 20 governor) on October 15, 2026, yielding 368 unique poll rows. We kept polls fielded in 2026 that carry a public source citation: 141 documents covering 128 distinct polls across 26 races. Polls absent from those pages are outside this sample, and national generic ballot polling was not included.
Unit and coding. We audited the document each public citation reaches, which is what a reader following the trail actually finds. Each readable document was coded 1 or 0 for eight elements adapted from AAPOR's Disclosure Standards: sponsor, field dates, sample size, population, mode, sampling frame, weighting, and a stated precision figure. A document we could not read in full was coded not assessed on every element and excluded from disclosure percentages.
Conventions. University polling centers publishing under their own name count as disclosing their sponsor; a commercial firm publishing with no client named does not. An unnamed “online panel” does not count as a disclosed frame; a named panel or voter file does. Where the same poll is cited through more than one document, each document is coded separately, since each is its own disclosure act.
Limits. This is an audit of the public record, read by automated tools with the reachability caveats stated above, at one point in time. It measures transparency, not polling accuracy, and the sample reflects which races and polls the public record tracks most closely. The full dataset, with every source link and per element code, is here: download the data (CSV).
No. This report measures transparency, not accuracy. A poll can disclose nothing and still be well executed, and a fully documented poll can still miss. The point of disclosure is that readers should not have to take that on faith: the elements here are what a reader needs to judge a poll for themselves, and professional standards say every published poll should provide them.
Because it captures the public record as an interested reader actually experiences it. Wikipedia's race pages are among the most visited poll listings this cycle, and each poll row cites a source. We audited the document that citation reaches, which is exactly what a voter, journalist, or researcher following the trail would find. The trade-off is coverage: polls never added to those pages are not in our sample, and we say so plainly in the methodology.
Yes. The full dataset, including every source link and per element code, is downloadable below. If we misread a document, or a release has moved, write to support@wavefieldresearch.com with the link and we will recheck and correct the record. Corrections will be noted on this page.
Sponsor, population, sampling frame and method, mode, field dates, sample size, weighting approach, and a precision statement that matches the design. Our survey methodology guide includes a fill-in template that covers all of them in four sentences.
Related: survey methodology, with a fill-in statement · sampling methods · political polling · what a political poll costs
Sample source, quotas, quality rules, and weighting are stated in every report we deliver, because that is what we would ask of anyone else. Describe your study and see how it works.
Start a study