Recommendation Simulator
How different objective functions produce dramatically different recommendations from the same seed content.
The algorithm decides what you hear next
Most audio platforms optimize for one thing: engagement — keeping you listening as long as possible. This tends to recommend more of the same: the same speaker, the same style, the same narrow slice of content. Rejoice is exploring what happens when you optimize for something different.
This simulator runs three recommendation engines side by side on the same starting content, so you can see exactly how the results diverge:
Engagement Mode (Spotify-style) — Recommends the most similar content, biased toward popular items and the same content type. Tends to create echo chambers.
Discovery Mode (Rejoice's approach) — Actively diversifies across artists, content types, and scripture. Penalizes recommending the same speaker twice. Surfaces content you wouldn't find otherwise.
Theological Depth Mode — Prioritizes content with scripture references, shared biblical connections with the seed, and longer-form material. For users who want to go deeper.
The diversity metrics below each column count how many unique artists, content types, scripture books, and themes appear in the results — a simple way to measure how broad or narrow each approach is.
- Engagement mode heavily biases toward the same content type (1.15x boost) and popular items, often returning a single speaker repeatedly — the echo chamber effect
- Discovery mode penalizes repeat artists (0.65x) and rewards cross-type recommendations (1.12x), consistently surfacing more unique voices and formats from the same seed
- Depth mode prioritizes scripture-rich content (1.2x+ boost for items with references) and longer-form material, revealing theological connections that engagement-optimized systems miss
Loading featured items...
Select a content item above to see how each algorithm recommends differently.
The diversity gap, at scale
Single-seed demos are suggestive. Averages across 100 random seeds show how systematically the three modes differ.
Recommendation diversity, measured
To turn the side-by-side simulator above into a population-level claim, we sampled 100 random seed items from the 14,729-item catalog (seed 42, reproducible) and ran every one of them through all three engines. For each run we recorded the unique artists, content types, scripture books, and themes in the top-10, then averaged across all 100 runs.
We also measured how much the three modes agree: for every seed, how many items appear in more than one mode's top-10. If the modes were interchangeable, the overlap would be near 100%. It isn't.
- Discovery mode produces 1.39× more unique artists than Engagement mode (9.20 vs 6.62 per top-10 on average) — the clearest quantitative signal of the "echo chamber" effect that engagement-style ranking introduces
- Depth mode surfaces ~1,450% more scripture references than Engagement mode (77.5 vs 5.0 refs per top-10) and covers ~7× more distinct biblical books (26.8 vs 3.4) — a massive shift in what "related content" means when the objective changes
- Across 100 seeds, only 2.8% of recommendations appear in all three modes' top-10 on average — the three engines are producing largely non-overlapping lists from the same underlying embedding space
- The pairwise overlap confirms the same story: Engagement ∩ Discovery share 19.9%, Engagement ∩ Depth 11.5%, and Discovery ∩ Depth just 5.0% — Discovery and Depth are optimizing for almost entirely different things
Methodology & Limitations
Algorithm: All three engines start from the same cosine similarity scores in the 384-dimensional embedding space, then apply different weighting functions. Engagement boosts same-type content (1.15x) and popularity; Discovery penalizes repeat artists (0.65x) and rewards cross-type results (1.12x); Depth prioritizes scripture references (1.2x+) and longer content. These multipliers were set heuristically, not optimized.
Metrics: Diversity counts (unique artists, types, scripture books, themes) measure structural diversity, not recommendation quality. No user satisfaction data, A/B testing, or relevance judgments inform these results.
Limitations: Results depend heavily on the seed item. The "echo chamber" effect in Engagement mode varies by content type — it's most pronounced for popular speakers with large catalogs. No collaborative filtering signal is used (no user behavior data), so all recommendations are content-based.