Skip to content

Methodology

The science and architecture behind MirrorSource.

The Science Behind It

MirrorSource draws on established frameworks in media analysis and journalism studies:

Narrative Analysis

Based on Matt Taibbi's work on how different outlets frame the same events using different narrative structures—hero/villain, crisis/resolution, us/them.

See: Taibbi, M. (2019). Hate Inc.

Churnalism Research

Building on Nick Davies' research showing how press releases and wire copy get recycled across outlets with minimal original reporting.

See: Davies, N. (2009). Flat Earth News

Structural Analysis

Informed by Herman & Chomsky's propaganda model, which examines how ownership, funding, and institutional pressures shape news coverage.

See: Herman, E. & Chomsky, N. (1988). Manufacturing Consent

Bias Rating Methodology

Political lean classifications sourced from AllSides (blind bias surveys) and Ad Fontes Media (content analysis methodology).

AllSides | Ad Fontes

Our Architecture

Decoupled "Eyes + Brain" Design

MirrorSource separates search from synthesis to reduce latency and minimize AI hallucination.

👁️ Eyes (Search Layer)

  • • Brave Search API for real-time news discovery
  • • No content scraping or storage
  • • Returns source metadata only

🧠 Brain (Synthesis Layer)

  • • Google Gemini for summarization
  • • Grounded in search results only
  • • Structured output (Common Ground, Key Differences)

This architecture means the AI never "makes up" sources—it can only work with what the search layer returns. Trade-off: if a source isn't indexed by the search API, we won't find it.

Benchmarks

Performance scores measured using Google Lighthouse. Results may vary by device and connection.

View detailed benchmark results

Limitations and Safety

MirrorSource summarizes and compares coverage

It does not fact-check. We show you how different sources are reporting a story, not which version is "true."

AI-generated summaries may contain errors

Always click through to original sources to verify information. Our summaries are starting points, not endpoints.

We show confidence indicators where data is uncertain

Story provenance, divergence levels, and coverage gaps are shown with appropriate uncertainty markers.

Coverage limitations

  • • We analyze 187+ sources, but can't cover every outlet
  • • Breaking news may have limited alternative coverage initially
  • • Paywalled articles may require keyword search fallback
  • • Political lean ratings are approximations, not absolute truths

Transparency

No Tracking

We use Vercel Analytics for aggregate page views only. No cookies, no user profiles, no reading history. Your searches are processed in real-time and not stored.

Changelog

Major updates are documented in our changelog.

View changelog

We may publish aggregate pilot learnings (e.g., "users found X feature helpful"). No personal data is ever shared.

Questions about our methodology?

We're always looking to improve. Join the pilot and share your feedback.