How PressRank Measures Credibility
PressRank uses a blind, community-driven, algorithmically computed framework to rank news channels. By stripping outlets of their branding and letting people judge anonymized coverage, the platform aims to produce an objective, manipulation-resistant credibility rating.
1. The Five Quality Dimensions
Rather than compiling a single "good/bad" score, we rate channels across five key dimensions. In the Arena, voters answer low-temperature, comparative prompts:
“Which is most precise about what is actually known vs. speculation?” Measures factual grounding over speculation.
“Which of these is worded most neutrally?” Identifies tone objectivity and the absence of loaded words or editorial slants.
“Which is best sourced / most specific about where it comes from?” Evaluates if claims cite verifiable entities or anonymous rumors.
“Which statement is most independent of government narrative or establishment propaganda?” Assesses how independent the reporting is from state narratives.
“Which is the least sensational / clickbait-driven?” Rates the objective, measured editorial delivery of stories over high-alert headline grabbing.
2. Statement-Level Quality (Chance-Adjusted Shrinkage)
Each vote is a partial ranking: on a given slate, the statements the voter selected rank above the ones they did not. The raw signal is a selection rate — selections divided by exposures.
A raw rate is not comparable across slates, though. Being picked once out of a head-to-head pair is a coin flip; being picked once out of a seven-statement slate where the voter chose three is a 43% shot. So for every impression we record what a random voter making the same number of picks would have scored, and each statement is measured against its own chance baseline rather than one global constant.
The shrinkage term (5 pseudo-observations) stops a statement shown twice from outranking one shown 2,000 times. The odds ratio then rescales the result so that performing exactly at chance always lands on 0.50, whatever mix of slate sizes the statement happened to appear in. Above 0.50 means voters picked it more often than random selection would.
Votes where the voter selected nothing are discarded rather than counted as exposures — an abstention carries no ranking information.
3. Rolling Up to Channel Ratings
A channel's rating pools the evidence from all of its statements rather than averaging their scores. Averaging would let a statement judged twice count as much as one judged forty times; pooling the underlying counts weights each statement by how much evidence it actually carries.
The per-statement cap stops one excerpt that went viral in the Arena from becoming the channel's whole rating. 50 is chance; the scale runs 0–100.
The uncertainty band (±σ) is a real standard error, not a function of the statement count alone. It combines the binomial error on the pooled rate with how much the channel's own statements disagree with each other:
The leaderboard is ordered by rating − 1.96σ, the conservative estimate, so a channel judged thirty times outranks one that got lucky twice. The figure displayed is still the rating itself.
Ranking thresholds. A channel qualifies for the public ranking once it clears a minimum statement count and a minimum number of judgements. Those bars are deliberately low while the database is young — otherwise the leaderboard would simply be empty — and tighten automatically as vote volume grows:
- Under 2,500 votes — 1 statement, 2 judgements.
- 2,500 to 10,000 votes — 2 statements, 4 judgements.
- Above 10,000 votes — 3 statements, 10 judgements.
Channels with real votes behind them that have not yet cleared the current bar are still shown, listed below the qualified ones and clearly marked not yet qualified. Hiding them would misrepresent how much the community has actually judged.
4. Vote Weighting (Anti-Brigading)
To resist coordinated inauthentic behavior, vote farms, and brigading fanbases, votes are weighted, not simply counted. Every cast ballot has its weight computed server-side:
- Identity Trust: Derived from account verification, hardware attestation layers, and account age. Freshly spawned accounts start with a weight near zero.
- Behavioral Authenticity: Collusion models monitor lockstep voting patterns, temporal burst activity, and network IP clustering to identify coordinate bot behavior.
- Recency: Emphasizes recent evaluations, ensuring stale historical ratings decay as channels alter their content styles.