GreatReads
Rebuilding Goodreads' recommendation engine around transparency, preference-first matching, and recency. A product case study.
Goodreads Tracks Reading. It Doesn't Shape It.
Goodreads has 150M+ registered users, 15+ years of structured rating data, and is owned by Amazon — one of the world's most sophisticated recommendation companies. Yet its recommendation engine has barely improved since 2013.
Users consistently report that recommendations feel generic, disconnected from their actual preferences, and driven by popularity rather than personal taste. Many describe Goodreads as a log for their reading history — useful for tracking, but not for discovery.
The opportunity: Goodreads has a moat competitors can't match — a 15-year deep graph of structured rating data. A preference-first recommendation engine would convert that latent data asset into genuine user value for the first time.
300+ Complaints. One Pattern.
Research began with a qualitative analysis of 300+ user complaints about Goodreads recommendations sourced from Reddit, Twitter, app store reviews, and user forums. The pattern was consistent:
- Recommendations don't reflect actual reading preferences
- Suggestions are driven by popularity, not personal taste
- Likes, dislikes, and shelves have limited visible impact on what's shown
- Even active users with 500+ ratings receive generic lists
- Users ignore recommendations and rely on external sources (TikTok, friends, newsletters)
A common phrase across complaints: "Even after rating hundreds of books, I still get the same generic lists." The data signal is there. The system isn't using it.
"When recommendations fail, Goodreads becomes a passive book log — not a guide for discovery. Smart recommendations unlock the positive feedback loop: better signals → better recommendations → better engagement."
Research synthesisThree Principles. Explicit Trade-offs.
The redesign was constrained by three explicit design principles, each with articulated trade-offs:
Principle 1 — Transparency Users should always understand why they're seeing a recommendation. Show the reasoning. Trade-off: limits black-box ML, but explainability builds trust faster than accuracy alone.
Principle 2 — Preference-First Not all user actions signal liking. Optimise for 4–5★ ratings (clear preference signal), not shelves (intent is ambiguous). Trade-off: sacrifices some recency for signal clarity.
Principle 3 — Recency Reader taste evolves. Recent high ratings matter more than decade-old favourites. Trade-off: new users get less personalised recommendations initially — but cold start is temporary.
Preference-First Recommendations
The redesign moves Goodreads from activity-based to preference-first recommendations through a three-layer matching approach:
- Layer 1 — Preference extraction: Only 4–5★ ratings from the last 24 months power the recommendation engine. Recent ratings are weighted higher.
- Layer 2 — Semantic similarity: Books are matched on genre, themes, writing style, and pacing — not just category tags.
- Layer 3 — Social graph: Recommendations are boosted when readers with highly aligned taste (same books rated similarly) also loved the suggested title.
Each recommended book shows a % match score and expands to explain the reasoning: "Because you loved X, Y, and Z — and 847 readers with your taste gave this 4.8 stars." This directly addresses the transparency principle.
10–20% of recommendations are intentionally "exploration" titles — books outside the user's core cluster, but surfaced transparently as "something different" to prevent taste narrowing.
Measuring What Matters
North star metric: % of books added to shelves from recommendations. Current estimated baseline <5%. Target: 25%+.
Supporting L1 metrics (trust and engagement signals):
- Recommendation click-through rate
- % of recommendations later rated 4–5★ (quality signal)
- Repeat visits to the Explore section
- "Not interested" / hide actions (guardrail — should decrease)
L2 metrics (business and ecosystem impact):
- Amazon click-out conversion rate
- Hidden gem discovery rate (books with <10K ratings)
- Retention of active readers month-over-month
Guardrail metrics:
- Diversity score — number of unique genres in recommendations (must not collapse)
- Over-personalisation rate — recommendations from 2 genres or fewer
- Cold start quality — % of new users adding ≥1 recommendation in first week
What Could Go Wrong
Risk
New users get weak recommendations without sufficient rating history
Mitigation
Guided onboarding collects 5 strong preference signals upfront (books you already loved)
Risk
Over-personalisation creates narrow taste bubbles
Mitigation
Inject 10–20% exploration titles. Diversity score as active guardrail metric.
Risk
Matching and similarity calculations increase infrastructure cost
Mitigation
Batch process recommendation generation. Incremental updates, not real-time recalculation.
The future of Goodreads is not social. It's intelligent. A preference-first engine makes Goodreads personal, trusted, and indispensable — rather than a database with a social veneer.