Methodology

The Breakout Score is a 0–100 measure of how strongly an experience is accelerating right now. It combines six signals, each weighted by how well it has predicted a real breakout in our own data:

• CCU acceleration: how fast concurrent players are growing over the last 24 hours.
• Sustained trend: the 7-day CCU trajectory, so a single viral spike doesn’t dominate.
• Visits & favorites velocity: the rate of change of visits and favorites.
• Like-ratio vs. genre cohort: quality relative to comparable games, not the whole platform.
• Update cadence: actively maintained games score higher than abandoned ones.

Each signal is converted to a z-score against the game’s genre cohort, winsorized at ±3 so a single extreme outlier cannot dominate the score, then passed through a weighted sum and a sigmoid to produce a 0–100 value. Acceleration carries the most weight and update cadence the least. The exact coefficients are the one part of this we keep to ourselves, because they are what the product is; everything they operate on is on this page.

score = σ( Σ wᵢ · z(signalᵢ) ) × 100

z(x)  signal vs its genre cohort, winsorized at ±3
wᵢ    fixed weights, largest on 24h acceleration
σ     logistic squash onto 0–100

Four adjustments apply on top:

• A scale factor dampens acceleration on a tiny base: the multiplier ramps from ×0.5 at ≈10 average CCU to ×1.0 at 1,000+ (log scale). A 20-player game doubling overnight is a lead worth watching, not a chart-topper above games holding tens of thousands of players.
• Games younger than 72 hours are multiplied by ×0.85 and marked provisional. Early data is noisy and the score may still move sharply.
• A day we did not observe enough of is not rescored. When too few readings land for a game on a given day, the last confident score is carried forward and marked provisional rather than a fresh number computed from a fraction of the day. A score that moved because our crawler was thin, not because the game was, is the one failure a breakout score cannot afford.
• Games with suspected botted CCU are capped at a score of 35 (see Bot-CCU flags).

Roblox does not publish per-game sales, so revenue must be estimated. We build the estimate in three steps and always report it with a confidence interval.

1. DAU proxy. We estimate daily active users as the geometric mean of two independent sources: the daily award count of the game’s most-awarded badge (pastDayAwardedCount), and CCU-derived turnover (CCU × 86,400 / average session length). Using two uncorrelated proxies lets us detect when either one is unreliable.

2. Daily revenue. The DAU proxy is multiplied by a payer-conversion rate and the expected gamepass value across the game’s live pass ladder, where a pass’s purchase weight is proportional to price^−1.2 (cheaper passes sell disproportionately more). We apply the 0.7 creator revenue share, then add a Premium Payouts term of roughly CCU-hours × 1.4 R$.

daily  = DAU × payer_conversion × E[gamepass value]  × 0.7
         + CCU_hours × 1.4 R$            (Premium Payouts)
monthly = daily × 30.44

3. Confidence interval. The point estimate is wrapped in a multiplicative band whose width depends on how much we trust the inputs:

A “≈2.4M R$/mo, high confidence” estimate therefore means: we believe the true figure most plausibly lies between ≈1.7M and ≈3.5M R$/month.

Each of our tracked niches is scored on three axes:

• Demand: combined CCU of the niche’s top-20 games and its 28-day growth.
• Clone pressure: how many new entrants shipped into the niche over the last 90 days.
• Incumbent quality: the like ratio of the games currently holding the niche.

These combine into two 0–100 outputs: a saturation score (how crowded the niche already is) and an opportunity score (how much unmet demand remains). Verdicts are then assigned against absolute demand thresholds, not against each other: a niche has to clear an absolute demand floor before it can be called an opportunity at all, and a higher bar again before a clone is called a business rather than a side project. Above that second bar, low crowding earns the “Clonable window” verdict: the demand is proven and the incumbents are beatable. Real demand with heavy crowding is marked “Saturated”, and a niche below the floor is called subscale however empty it looks. Those floors are derived from the live distribution of demand across every niche we track and are re-fit as it moves. We used to rank niches against each other instead, which produced the one failure this change exists to prevent: a niche with 50k CCU of total demand and two competitors outranked one with 456k, because ranking can see that a pond is empty but not that it is small.

UGC items are scored by favorites growth, z-scored within their category , so hair accessories are compared with hair accessories, not with shoulder pets. Two more terms complete the score: a price sweet-spot fit (how well the item’s price sits in the band where its category actually converts) and, for limiteds, resale acceleration, meaning how quickly resale activity is picking up.

Some experiences inflate their concurrent player counts with bots. We flag suspected botting with an engagement-incoherence heuristic: when a game’s CCU is high but the engagement that should accompany it (visits, favorites, likes) does not move coherently with it, the pattern is flagged.

Flagged games are labeled throughout the product and their Breakout Score is capped at 35, so botted titles can’t crowd genuine breakouts off the radar. Flags are heuristics, not proof. We mark suspicion, we don’t accuse.

• All data comes from public Roblox endpoints . The same information visible on any game or catalog page. We use no private APIs and never ask for access to your Roblox account.
• Collection is batched and rate-respectful, on a two-speed schedule: hot games are polled every 15 minutes, the long tail less frequently.
• Raw snapshots are kept for a maximum of 90 days. Daily aggregates (the basis of charts and scores) are kept forever.

• Every figure is an estimate, not ground truth. Roblox exposes no per-game sales or revenue data, so no third party, including us, can measure earnings directly.
• Confidence bands quantify our uncertainty but do not eliminate it; a low-confidence estimate can be wrong by a large factor.
• Calibration is ongoing. As our tracking history deepens, weights and bands are re-fit. Expect scores to improve over time, and see the changelog for material changes.