Skip to main content

How TasteRay Works

Methodology, data sources, and the research behind our claims.

TasteRay is built on a different premise than mainstream streaming recommendations: we optimize for the movies and TV series you'll remember, not the ones that maximize watch time. This page documents how that works in practice.

The 92% hit rate

We claim 92% of users discover a new personal favorite within their first month. Here's how that's measured: in opt-in user research between 2025-09 and 2026-04, we asked users to rate each TasteRay-recommended title they watched on a 1–10 scale. A "personal favorite" is defined as any title rated ≥ 8/10 by the user. The metric counts the share of monthly active users with at least one ≥8/10 rating during the period. Sample size: 4,200+ active users. We refresh this number quarterly.

Cinephile-grade AI

Our recommendation engine is trained on more than 500,000 long-form reviews from professional critics, festival programmers, and verified cinephiles — not just star ratings or watch-history. The model learns which films map onto which emotional and aesthetic registers (e.g. "slow contemplative reflection" vs "kinetic catharsis"), and matches them to the user's stated mood, occasion, and prior loves.

Evaluation: held-out user studies measure recommendation quality against three baselines (Netflix's algorithm, IMDb top-rated by genre, and a random control). TasteRay's average user-rated quality (1–10) currently exceeds the next-best baseline by 1.8 points.

Citations and source research

Several landing pages reference psychological or behavioral research. The full bibliography:

Data sources

  • The Movie Database (TMDB) — title metadata, posters, cast, crew, runtime, certification, streaming availability.
  • Critic and audience reviews — long-form reviews aggregated under fair-use research provisions; not redistributed.
  • User-provided signal — ratings, mood/occasion descriptors, and free-text feedback from TasteRay users.
  • Letterboxd public lists and tags — for community-curated taste signals.

Last updated: 2026-05-10.