The "Recommended for You" section currently sorts by rating DESC. No other signals. A 4.9-star restaurant 30 minutes away shows above the 4.1-star spot next door. Users report recommendations don't feel relevant.
Weighted sum scoring. Each restaurant gets scored against four signals: distance, rating, review count, and visit recency. Each signal is normalized to a 0-1 range, multiplied by a configurable weight, and summed into a final score. The weights let product tune what "relevant" means without changing code.
The distance gate (toggle it above) shows what happens with a hard cutoff: restaurants past the threshold get their score crushed to 10%. A 4.9-star place at 5.2 km gets buried while a 3.2-star place at 4.9 km stays untouched. Whether that's the right tradeoff is a product call.
| Question |
|---|
| How much should distance matter vs rating? The "quick lunch nearby" and "date night" scenarios produce wildly different rankings from the same data. |
| Should we penalize recently visited places for variety? Current recency signal treats "visited yesterday" as a negative, but some users have a regular spot. |
| Do we trust ratings from places with very few reviews? The Green Fork (4.0, 42 reviews) and El Fogon (3.5, 28 reviews) score differently than their ratings suggest once review count factors in. |
| Should there be a hard distance cutoff or a gradual penalty? Toggle the distance gate above to see the cliff effect. A smooth decay curve is another option. |