RANKING

How the Top Rated list is sorted, and why a 5.0 with 2 reviews doesn't always beat a 4.8 with 10.

The problem

A simple average rewards luck. Three perfect 5★ reviews give a 5.0, but that's a tiny sample — one bad review would tank it. A game with a 4.8 average over 40 reviews is a stronger signal of quality, even though the raw number is lower.

Sorting by raw average puts low-volume games on top by accident. Sorting by review count alone ignores quality. We want both.

Bayesian average

Top Rated uses a Bayesian average — every game is treated as if it had a few extra phantom votes at a neutral rating. Games with many real reviews barely feel the prior; games with very few are pulled toward the neutral score.

score = (n · avg + k · m) / (n + k) n = number of reviews avg = actual average rating k = prior weight (2 votes) m = prior mean (4.0)

With k = 2 and m = 4.0, the prior is light. It only matters when review counts are small. Once a game has ~10+ reviews, the score is essentially the real average.

Live example

The current top 5 reviewed games on the site, scored under the formula above:

#	Game	Avg	Reviews	Score
1	Floppy Brawler	4.7	6	4.53
2	EmojiBeats	4.6	12	4.51
3	Almost Surgery	5.0	2	4.50
4	Joe Must Drive	5.0	2	4.50
5	ECO Guardian	4.5	22	4.46

The displayed star average stays the real number — only the internal ranking score is weighted.

One author, one vote

Anyone can post as many reviews as they want — but for the rating that matters (the average shown on the game page), all reviews from the same author are collapsed into a single vote using their own internal average. So six 5★ reviews from one person count the same as one 5★ review.

This means the rating is resistant to spam: posting the same review ten times doesn't move the needle. It also means a single determined troll can't tank a game by spamming 1★ — they get one 1★ vote, not ten.

per_game_average = average over distinct authors of (each author's own average)

On the game page, multi-review authors show up as a single card with the average of their own ratings up top, the most recent review below, and a "Show N previous reviews" expander that opens a tree view of the rest.

Identifying authors

Anonymity is a product requirement — there's no login. To identify the same person across reviews without an account, three signals are combined, in order of strength:

Device fingerprint — a SHA-256 hash of canvas rendering, WebGL renderer/vendor, OfflineAudioContext output, installed font probes, screen, timezone and user-agent. Stable across cookie clears, browser restarts, and most VPN switches.
Anonymous client cookie — a random UUID stored in an httpOnly cookie (vr_uid) for one year. Survives IP changes; cleared when the user wipes cookies.
IP hash — SHA-256 of the request IP plus a server secret. Raw IPs are never stored. Last-resort signal — shared networks (campus Wi-Fi, mobile carriers) collapse onto one hash.

For grouping, the strongest available signal wins: fingerprint → cookie → ip. Each review carries an opaque, truncated fingerprint id (e.g. #a1b2c3d4) shown on the card so repeat authors are visible at a glance.

Tiebreakers

Higher Bayesian score wins.
If scores tie, more reviews wins.
If review counts also tie, higher raw average wins.

Other sorts

Newest— sorts by the game's submission timestamp. Reviews don't affect this.
Most Reviewed — pure review count, with average as the tiebreaker.

Star display

Stars use fractional fills — a 4.5 average shows four-and-a-half stars filled, not five. The numeric score (e.g. 4.8 (4)) is shown next to the stars so close averages are easy to tell apart.