Net Trust Score shows how much verified voters on this site trust each AI company. It is not a representative poll of any population: people choose to take part, so every score carries its voter count.
The question
Each voter answers one question per company: "How much do you trust {company}?", from 0 "Not at all" to 10 "Completely". The question names each company's best-known AI product in parentheses. Voters are asked to think about the company and its AI products, and they can rate any companies they choose.
The formula
Scores of 9 or 10 count as trust, 7 or 8 as neutral, and 0 to 6 as distrust. Neutral ratings count in the total, so they pull a score toward zero without counting for either side.
Net Trust Score = % trust − % distrust, from −100 to +100
Example: 200 ratings with 90 at 9 or 10, 50 at 7 or 8 and 60 at 0 to 6 give 45% − 30% = +15. Every score is shown with the average rating and the full 0 to 10 distribution.
The method is inspired by promoter-style customer scoring, applied here to trust. Net Trust Score is not affiliated with any owner of those scoring trademarks.
When a score appears
A company's score, average and distribution stay hidden until it has 30 ratings. Until then the page shows how many it has, such as "Collecting ratings: 12 of 30". The split between people who use a company's product and people who don't appears once each group has 30 ratings. A score is recorded every night, and the trend compares today with 30 days earlier.
One person, one vote
- Voting requires signing in with a Google account, or an email address confirmed with a one-time code. Every voter has a verified email, and one email is one voter.
- Each voter has one current rating per company. A rating can be changed after 30 days, and the earlier rating is kept in the history.
- The rating page hides current scores and shuffles the order of companies, so neither nudges a vote.
- A bot check, rate limits and a review of unusual spikes protect the count. Ratings that look like manipulation can be set aside without being deleted, and they are excluded from scores while set aside.
- Someone with several email addresses could vote more than once. No sign-in system fully prevents that, and we say so here.
Reasons and comments
A voter may add one reason, up to 500 characters, to each rating. Reasons are published without a name and show only the score given, whether the voter uses the product, and the month.
Before a reason is published, an AI model checks it against the comment policy. Harsh criticism and praise of a company pass. Spam, advertising, harassment, hate, personal data, accusations against named private individuals, threats and off-topic text do not. Uncertain cases go to a person. The AI never affects a score: a rating counts whether or not its reason is published. Anyone signed in can report a published reason, and three reports send it back for review.
The AI we use
Comment checks run on Claude Haiku 5.5, made by Anthropic. Anthropic is one of the rated companies, so we track and will publish the rejection rate for reasons about each company. The model receives only the reason's text and the company name, never a name, email or account id.
Independence
Net Trust Score takes no money from the companies it rates. Any future sponsor will be named on this page. Scores are never edited by hand, and every moderation action is logged.
Corrections
If you represent a rated company, or you think something on the site is wrong, write to us. We correct errors of fact in our pages and in this method. We do not change voters' ratings.
Change log
- 12 October 2026: clearer wording on neutral ratings (7 and 8). They count in the total and pull a score toward zero. The formula is unchanged.
- 11 October 2026: method published for the alpha.