On November 3, ABC, CBS, CNN, Fox News, MS Now, NBC and the Associated Press all described the electorate with the same numbers. They came from the Voter Poll by SSRS, which the seven organizations use together for the first general election in a decade.¹ The poll's crosstabs are free to read on any network site. The respondent-level file is not, and when past exit-poll files have been released, it has taken years and a fee.

Both facts have the same source. The exit poll is built for one customer group, a handful of news organizations that need a fast, defensible answer on one night, and it is priced and governed accordingly.

How the pool formed

Networks used to run competing exit polls. Competition was expensive, and from 1990 they pooled the work. Pooling saved money but left one point of failure, and that failure came twice.

The pattern is that the buyers, not a dominant seller, keep the market at one supplier. Few organizations need the product, each needs all 50 states on the same night, and none wants to pay for a full operation alone. When Fox left the pool in 2017, the open question was whether the remaining members could afford a survey that costs millions without it.⁶

What the “exit poll” is now

Very little of it is exit interviews. In 2016, Edison interviewed about 85,000 voters in person at roughly 1,000 polling places plus 16,000 early and absentee voters by phone.⁷ In 2026 the Voter Poll interviewed roughly 75,000 voters before Election Day, through mailed invitations, phone, text and online panels, and added in-person interviews at polling places in 15 states.⁸ Its samples come from the voter file, which SSRS licenses in full from the data vendor L2, and from its probability-based Opinion Panel. They also include an opt-in online panel, Verasight, whose members are partly recruited through internet advertising.^9,8

The data are then weighted to an estimate of each state's electorate and reweighted on election night as counts arrive, so the published crosstabs keep changing for days.⁷

Why a mobile startup can't simply do it better

Collecting answers on phones is not the hard part; the incumbent already does much of its fieldwork by text and online panel. The barriers are elsewhere.

The buyer. The product's value is concentrated in a few hours on one night, and the customers are the same seven organizations that have just signed with SSRS. A startup has no one to sell to unless it persuades a network to put an untested survey on air.

The track record. The networks have been burned before, in 2000 and 2002. In 2016 a startup, VoteCastr, published real-time Election Day estimates with Slate and Vice News in seven battleground states. It missed the winner in five of the seven.¹⁰ Edison's Joe Lenski has described the exit poll as a survey that gets graded within two hours of polls closing, against precinct returns from the places it sampled.⁷ A new entrant gets one public test, on the most watched night of the cycle.

The sampling frame. A credible voter survey needs a sample drawn from the voter file in every state, with enough respondents to report Black, Hispanic, Asian and young voters separately, in English and Spanish. A sample recruited through an app overrepresents the people who use that app. Older and offline voters are hard to reach, and the fix, licensing voter files and running probability-based recruitment, is the expensive part of the incumbent's operation.

Verification. On election night no one can confirm that a respondent voted. Vote history in the voter file arrives weeks or months later. Until then the survey relies on self-reports and weighting, and a mobile panel is more exposed to fraudulent and duplicate respondents than an interviewer at a polling place.

The weighting inputs. The final numbers depend on reweighting to the actual vote, which requires fast precinct and county counts. The AP, a pool member, also runs the main vote-counting operation.

Is the microdata really never released?

It is released, but late, for a fee, and not for the current election. The Roper Center at Cornell archives respondent-level exit polls back to 1972 and sells them: a national file costs $1,625 to $2,165 and a single state $605 to $810, depending on membership.¹¹ The 2016 and 2018 pool files became available in January 2020.¹² Access once required an application and a signed data protection plan covering all data derived from the files; that requirement has since been dropped.¹³ AP VoteCast public-use files for 2018 through 2024 are posted by AP-NORC.⁵ I found no announcement on when or whether the 2026 Voter Poll file will be released.

The open alternative already exists, just not on election night. The Cooperative Election Study, run by a consortium of universities and fielded by YouGov, surveys about 60,000 Americans each federal election. It matches respondents to the voter file to validate who actually voted and posts the full file free on the Harvard Dataverse.¹⁴ Universities fund it by buying 1,000-person modules for $14,000 each. The data reach teams in March, with validated vote added later.¹⁵

VoteCast sample size is the 2018 plan.¹⁶ Voter Poll figures are for 2026.⁸

Should someone build it?

A startup trying to beat SSRS at election-night calls faces the barriers above and a market of seven buyers. There is a gap, but it is in a different place: no survey of the electorate is both fast and open. The networks get speed, academics get openness a few months later, and everyone between those points reads crosstabs off news sites.

The pooling logic that produced the networks' single exit poll works the other way too. A consortium of the people who want the data afterward could fund a survey whose main product is the respondent file, released fast enough to matter while the election is still being explained.