How Trust Turtl works

The short version

Trust Turtl builds a credibility scorecard for the page you're on. It gathers independent signals about the domain, the writing, the images, the claims, and the page's technical behavior. Each signal gets its own score. Together they produce a single 1 to 10 rating, with the evidence laid out for you to see.

AI detection is one of those signals, not the verdict. A page can be AI-written and still factual. It can also be human-written and still a scam. Trust Turtl is built for the second question: should you trust this?

The pipeline

Collect visible content

When you click the icon, the extension reads the visible text of the active tab plus a sample of its images and embedded videos. Background tabs and pages you haven't opened it on are off limits.
Gather signals in parallel

The backend looks up domain history via RDAP, checks the URL against Google Safe Browsing, cross-references testable claims against the Google Fact Check database, and verifies image provenance through C2PA. Slow-changing lookups are cached so repeated visits stay fast.
Score each section

A language model reviews the text, bias, and consistency alongside the raw signals. Each of the five sections gets its own 1 to 10 score, a short explanation, and the evidence behind it.
Assemble the scorecard

Sections combine into a weighted overall score. Serious technical red flags can cap the result. The popup shows a headline score, a one-line summary, and five collapsible cards you can expand for the details.

What the ratings mean

Untrustworthy

Score 0 to 2

Strong red flags. Phishing patterns, known-scam domain, or a headline claim that already has a "false" fact-check.

Suspect

Score 3 to 4

Multiple weak signals. New domain, AI copy with no citations, missing provenance on images, or shady technical behavior.

Mixed

Score 5 to 7

Some parts read as solid, others are vague, opinion-driven, or hard to verify. Worth a closer look before sharing.

Trustworthy

Score 8 to 10

Established source, specific sourcing, coherent claims, and images with real provenance. Cross-check still encouraged.

The five signals

language

Domain & source

Domain age, registrar reputation, whether the site is on a known-publisher list, URL hygiene (typosquats, lookalikes), and whether an about or contact page even exists.

text_fields

Text content

Factual accuracy of testable claims, citation quality, bias and loaded language, internal consistency, and AI authorship. AI writing is flagged but never punished on its own.

image

Images & media

Ratio of AI-generated visuals, manipulation or deepfake indicators, stock versus original imagery, and C2PA content credentials where a camera or editing tool embedded them.

fact_check

Claims & fact-check

Direct matches in the Google Fact Check database, corroboration against Wikipedia, and how novel the claim is. A "false" match weighs heavily; an unmatched claim lowers confidence, not the score.

shield

Privacy & safety

Misleading URLs or link cloaking, Google Safe Browsing flags, hidden SEO text, affiliate density, and tracker count. Serious issues here can cap the overall score regardless of the other sections.

How AI content fits in

AI-generated text and images are a signal, not a verdict. A well-sourced explainer assembled with AI help can still be accurate, and a hand-typed page can still be a scam. Trust Turtl always shows an "AI-generated content" chip when it detects it, so you know, but it only drags the score down when AI copy shows up alongside weak citations or factually wrong claims.

The point is to help you navigate a web where most content has some AI in it. The question worth answering is no longer "did a human type this?" but "is any of this true, and who's on the hook if it isn't?"

Privacy and data

Trust Turtl only reads a page when you click the icon on it. Background tabs, private windows you have not interacted with, and the new-tab screen stay untouched.

The page content and URL go to the Trust Turtl backend so it can be analyzed. Requests are not tied to an account, and the extension does not keep a history of your browsing.

info

A note on limits. Language models can be wrong, especially on niche topics or a story that broke ten minutes ago. Treat the rating as a nudge to look closer. For anything that matters, read the source.