How it works

What we measure is bylined opinion content. Content where a person states their view in writing or speech and stands behind it by name. What we do not measure: wire copy, unbylined news, corporate announcements and advertising. The reason is simple: nobody defends an unsigned text, so its score cannot be written to anyone. If a measurement cannot be tied to a person, that measurement has no counterpart either. What we do or do not trust is not content but the people who produce it and take responsibility for it with their byline.

How we find and ingest it

We keep a curated list of outlets. We scan their feeds regularly. In every item we look for a byline: under the text, in the page masthead, in the channel details. We match the byline we find against our list and we keep the evidence, which page and which field it came from. If there is more than one byline, we count the attribution to a single creator as weak. Weakly attributed content enters the outlet score but not the creator score. We do not ingest what we cannot measure. The number of items we can take in each day is derived from the number that can actually be measured that day. Content that is ingested but never measured tells nobody anything.

Anyone with a verified email address. No expertise required, no application, no appointment, no hierarchy. Why we ask for verification. To make rating with fake accounts harder. An anonymous user cannot rate; we treat this not as a barrier but as the price of the ratings being trustworthy. But whether a verified user's judgement is objective is not our business. We believe everyone can and should express their judgement through a rating, as they see fit. Tiers. Your account has a tier and it affects how much your rating counts. Two tiers operate today: an unverified account cannot rate, a verified account can. Higher tiers with third-party identity verification will open later. What we do not keep. We never publish your username or your ratings anywhere, and never give them to any outlet. We show how many people rated next to the score; we do not show who they were.

When you rate a panel, who are you rating

In a television programme or a podcast, more than one person speaks. When you rate, we ask who you are rating: the host, or one of the guests. The default is the host, because they run the session. If you want to move quickly, your rating goes to them. But if you want to rate a guest, you can pick one. This question appears only in sessions whose participants can be identified. In a single-byline column there is no such choice; the rating goes straight to the author.

This product makes one claim: an item's score rises not because it got many votes, but because it won approval at the same time from people who disagree with each other. The system never asks you "which side are you on". It only looks at who approved what and draws a map. A map that places people who like similar things near each other, and people who like different things far apart. This map is not a political label. We do not call anyone anything; we only see who resembles whom. Then we split every rating into three: Is this person generally generous? Some people rate everything highly. That tendency is subtracted from the score. A teacher who gives everyone a hundred is not really grading. Did this person like it because it is from their own side? The shared approval of people who sit near each other on the map shows not the quality of the content but the internal cohesion of that segment. That is subtracted too. Members of the same family admiring each other says nothing about how admired they are on the street. What is left? Approval that cannot be explained by segment effects. That is the only thing we publish as a score. If an item collects approval only from its own segment, all of that approval is explained by the second term and none of it reaches the score. If it wins approval from two segments at once, something "real" that segment effects cannot explain is present, and that does reach the score.

Where this method comes from

Let us start by saying we did not discover the method. X (formerly Twitter) began using it in Community Notes in 2021, and Meta moved to the same model in 2025. In X's own test it was measured to reduce decisions to like and share an unfounded post by 25-34%. The University of Rochester ran an independent measurement over 264,600 posts. We use the same mathematics but produce a different output: they flag individual posts, we keep a lasting record, because trustworthiness is something that accumulates over time.

%20 of your feed comes from outside your usual reading habit. These items carry a "different perspective" label. This is not a recommendation, it is a precondition of measurement. We measured the reason: when we reduce this share to zero, the system never learns the opinion map and scores turn the wrong way. Content supported by a single segment ends up scoring higher than content supported by two. That is the exact opposite of what the product is trying to do. So removing that share does not merely make the system unable to measure; it makes it measure wrongly. The rest is selected by your interests and reading habits: the topics you picked at first launch and the sources of the items you have rated since.

Two conditions must hold together. Enough people. At least 10 distinct people must have rated. You can change your rating later. If you changed your mind, or think differently after reading the item, your latest rating counts and it does not increase this number. Enough diversity. The ratings must have come from readers of different segments. If they all come from one segment, we do not publish the score. If these two are not met, we write on the card what is going on: "Under evaluation". Not enough ratings yet. We show how many people have rated. "Narrow segment". There are enough ratings but they all came from one reader segment. We show the score, but with that label. If there is something the system cannot measure, we say so. We do not pretend to have measured what we could not.

Score (on a 0-100 scale). The approval left after segment effects are removed. A high score does not mean "this item is true". It means "readers from different segments found it trustworthy". Range (for example 65 - 82). Where the score might actually lie. This is a measure of uncertainty: with few raters the range is wide, with many it is narrow. If an item scored 74 with a range of 65-82, the score could also be 68 or 79. We do not hide this, because hiding it would manufacture false precision. How many people rated. How many people the score came from. The larger this number, the narrower the range. Coverage ratio on a creator's page. How much of that creator's published output we were able to measure. If coverage is low, the score should be read more cautiously, and we say so. Bridged scores cluster in a narrow band and do not reach the ends of the scale. If readers from different groups rate an item, then however low they rate it there is a shared judgement there. Where the band sits depends on how the raters rate, and that will be measured during the beta. When there is no shared judgement the item is not bridged; it is labelled "Divided" or "Single group" and shown separately.

Before a creator's aggregate score is published for the first time, we try to reach them. This is not asking permission, it is giving notice. We do not publish without notice, and this rule is enforced at the database level. The attempt follows a defined ladder: the author page at the outlet, the outlet masthead, the person's own public channel. Every attempt is recorded with its evidence. For people we cannot reach, there is a permanent public notice page: who is measured, how, and what can be done.

Three rights

Right of reply. They can write a text to be published next to their score. The system does not edit that text for its content: we do not shorten, soften or hide a defence we dislike. But our obligation to remove unlawful content stands, and there is a visible route for it. Both conditions are stated in the notification: the reply must concern the measured content, and must not contain insults or third-party data. The length limit is 1,000 characters. Correction objection. Factual objections such as "this piece is not mine" or "this is an interview, not my opinion" are reviewed. Objection to automated analysis. Under article 11 of Turkish data protection law. But there is a limit and we state it plainly: the score cannot be objected to, the input can. "My score is low, raise it" is rejected. "This item is not mine" is reviewed. The crowd's judgement remains sovereign; an objection only ensures that judgement was passed on the right thing.

We take no money from end users. Our revenue comes from corporate data services: advertisers, AI companies, media organisations, researchers. A person or an organisation can buy reporting on content published about them. But three things are not for sale, and this is bound by contract: The payer cannot choose which content is measured. They buy a period, and all content from that period is measured. The payer cannot buy priority. Priority would mean taking someone else's place in the measurement queue. Payment does not affect the score. The code that computes scores does not read payment records. This is not a promise, it is a constraint at the database level.

This may be the most important section on this page. We do not yet know how well the method works with real readers. We measured it in simulation: when an opinion-independent quality signal exists, the algorithm finds it with a correlation of 0.82, and when the signal is absent it does not find one. (That is a correlation coefficient, not a percentage.) But we built that simulation. How strong an opinion-independent quality dimension is in real readers' judgement is something we will only learn from real data. We do not know how many people will actually rate. We have estimates, but our real-world measurement is still very limited. We can only measure some panel guests. In television programmes and podcasts we measure the host and the guest separately, because these are different jobs and ranking them in one list would be meaningless. But to measure a guest, their name must appear in that broadcast's description; if it does not, we cannot see that participation and cannot offer it as an option when you rate. Today we can identify fewer than one in ten panel guests. We measure the ones we can identify, because measuring none of them just because we cannot measure all would be worse. But we have to say two things. The guests we measure are not a random sample. Named guests come from broadcasts that habitually write participants into the description. So guest scores may skew towards a particular type of broadcast. A guest's coverage ratio may be low. If their profile says "appeared on twenty programmes, measured on three", the score should be read accordingly. That is exactly why we show the coverage ratio. We write these down because a measure that does not know what it does not know is more dangerous than one that claims to know.