Methodology

Roblox Pulse infers engagement signals from public, keyless Roblox endpoints. Both headline metrics are proxies, not ground-truth retention or session length. Treat them as smell tests, not as substitutes for in-game telemetry.

Retention proxy (v2)

retentionProxy =
    (upVotes / (upVotes + downVotes))
  × log10(favoritedCount + 1)
  × log10(playing + 1)

Like-ratio rewards games with positive reception. Returns 0 when a game has no votes (e.g. brand-new releases).
log10(favourites + 1) measures depth of fandom on a log scale — the difference between 100 and 1,000 favourites matters more than between 10M and 100M.
log10(playing + 1) measures current concurrent engagement, also log-scaled so a 600k-player giant doesn’t completely drown out a 5k-player niche darling.

Why v2: the v1 formula divided by visits / 24, which collapsed established games like Brookhaven (82B lifetime visits) to a near-zero score — backwards from the business meaning of retention. v2 drops the visits-as-denominator term entirely and rewards quality × fandom × current activity. Lifetime visits no longer penalise the score; if anything, they correlate with all three terms going up.

Session-length proxy

estimatedSessionMinutes =
    later.playing / ((later.visits - earlier.visits) / intervalMinutes)

Two snapshots of playing and visits taken at different times let us compute visit velocity (visits per minute). Dividing concurrency by that velocity yields an estimated average session in minutes — under the assumption that, in steady state, average concurrent players ≈ session-length × arrival rate (Little’s law).

Buckets: <5m, 5–15m, 15–30m, 30m+.

What this doesn’t measure

True retention. D1/D7/D30 retention requires per-user data Roblox doesn’t expose publicly.
True session length. The proxy assumes steady state; live events, server outages and time-of-day swings will skew it.
Bots and idle servers. playing includes AFK accounts and bots. There is no public way to net these out.
Game age. Lifetime visits are heavily biased towards older games; the visit-velocity normalisation in the retention proxy partially — but not fully — corrects for this.
Discovery sampling bias. Seeds come from the explore-api sorts (Top Trending, Top Playing Now, etc.), so the dataset over-represents currently-popular games.

Sources & refresh cadence

Discovery seed: apis.roblox.com/explore-api/v1/get-sorts — ~230 unique universes per call across five sort rails.

Game enrichment: games.roblox.com/v1/games?universeIds=… (50 IDs per batch).

Votes: games.roblox.com/v1/games/votes?universeIds=… (50 IDs per batch).

Run pnpm refresh-data to regenerate the snapshot. Each refresh appends a (ts, playing, visits) sample per game; the cohort proxy uses the most recent two samples.