Methodology
A transparent breakdown of how we built the blended ranking — from raw data to final score.
Why Blended?
Customer feedback (Google + Tripadvisor + sentiment) is noisy: review counts vary 10x across courses, resort umbrellas distort individual courses, and some clubs have no public listings at all. The SA Top 100 expert ranking provides a stable anchor that reflects course quality. Blending the two smooths out feedback noise while still letting customer experience pull a course up or down.
Components (each on a 0–10 scale)
Expert Score
Linear from rank 1 → 10.0 down to rank 100 → 7.0. (Floor isn't 0 because being top-100 is itself a quality signal.)
Google Score
Star rating × 2, Bayesian-shrunk (prior weight m=100) toward the all-courses mean.
Tripadvisor Score
Star rating × 2, Bayesian-shrunk (prior weight m=50).
Sentiment Score
Qualitative 1-10 from 2-5 recent review snippets.
Weights
40%
Expert Rank
anchors the ranking, smooths kinks
30%
20%
Tripadvisor
10%
Sentiment
Missing Data Handling
Expert weight is always 40%. If a customer-feedback source is missing for a course, its weight is redistributed proportionally across the remaining customer sources, so the customer block always sums to 60%.
Example: Google missing → TA scaled from 20% to 40%, Sentiment from 10% to 20%.
Bayesian Shrinkage
In Plain English
Imagine two restaurants. One has a perfect 5-star rating from a single review. The other has 4.5 stars from 2,000 reviews. Which would you actually trust more? Probably the 4.5 — because one review barely tells you anything, while 2,000 reviews are hard to fake.
The Problem
If we took star ratings at face value, a brand-new course with one glowing review would beat a famous course with thousands of reviews averaging slightly lower. That's clearly wrong — the single rating just hasn't been tested enough.
The Fix
Bayesian shrinkage pulls every rating partway toward the average of all courses. How much it gets pulled depends on how many reviews back it up: few reviews → pulled a lot toward the average (we don't trust it yet). Many reviews → barely pulled (we trust what we see).
The Formula
Shrunk = (v / (v + m)) × R + (m / (v + m)) × C
- RCourse's own rating (0-10 scale)
- vNumber of reviews on that source
- mPrior weight (Google m=100, TA m=50)
- CAll-courses mean (Google ≈ 8.98, TA ≈ 8.48, on 0-10 scale)
Data Quality Note
- •Courses 1–25: Verified line-by-line against Google Maps via Wanderlog and direct Tripadvisor page fetches.
- •Courses 26–100: Star ratings from search results / Wanderlog. Review COUNTS for many of these are estimates (flagged per row). Estimated counts affect the Bayesian shrinkage modestly, not the underlying star ratings.