Every modern game has a random drop system, and almost none of them tell you the true probability. The number in the tooltip is usually the displayed rate, which is a different thing from the actual rate — and the gap is where the interesting design lives.
The Mechanics Designers Actually Use
Weighted probability. Almost nothing is a flat 50/50. Rarity tiers are weighted so a common item appears often and a legendary item appears rarely. The mechanism is the same one the Weighted Raffle uses: pick a uniform number, find which band it landed in, take that item.
| Rarity | Typical drop rate | How it feels |
|---|---|---|
| Common | 60–70% | Expected |
| Uncommon | 20–30% | Pleasant |
| Rare | 5–10% | Worth mentioning |
| Epic | 1–5% | A moment |
| Legendary | well under 1% | Something to post about |
The reason for tiers rather than a continuous distribution is perceptual. Players read a rarity label instantly, and the label is what makes the drop memorable.
Pity counters. After a hard threshold, the system guarantees a drop. This is the single most important mechanic for player retention, and it is the one most often left unmentioned in the UI.
A pity counter is not generosity — it is variance reduction. Over 100 draws at a 1% rate, some players see nothing and others see three. Pity converts a distribution with huge spread into one with a ceiling. Players notice the difference without knowing why.
Bad-luck protection. A variant that scales probability upward as your streak lengthens. A 1% drop might quietly become 3% on your twentieth consecutive miss. Again: variance reduction, and usually invisible.
Duplicate conversion. When the same item drops twice, the second becomes currency or a reroll. This solves the problem of the specific unfun outcome — not a bad item, but the same item — and it is arguably the most well-designed part of a modern drop system.
Why Players Are Right to Be Skeptical
The gambler’s fallacy is not always wrong. Everyone assumes “my luck is bad tonight” is irrational. In a system with a pity counter, it is a correct description of your position: you are genuinely closer to a guaranteed drop than someone who just started. The intuition that feels wrong is actually the model.
“Bad luck” clusters are real and unavoidable. A player opening 100 boxes with no rare item will believe the system is broken even at exactly the advertised rate. This is not a bug in their reasoning — it is what variance looks like from inside one sample. The problem is that no amount of statistical truth corrects the experience.
Near-misses are a design choice. A mechanic that frequently displays “so close” at a distance from the trigger is optimising for engagement, not honesty. Some games are explicit about this. Many are not.
Random is not the same as fair. Two problems here, worth separating. A weighted system is legitimate — weights are a published rule, and weighted selection is a real tool for real purposes. What is not legitimate is a system where the displayed probability does not match the implemented probability. That is a deception, not a design choice.
The Design Tension
Every loot system sits between two things that pull against each other.
Player expectation wants a return. Over enough attempts, the player should feel they are progressing.
House economics wants a stable drop rate. Revenue models depend on a predictable distribution, and generous variance destroys the predictability they are built on.
The resolution is pity protection: it lets the player experience a return while keeping the long-run average exactly where the economy needs it. Both sides get what they want, and the player is usually not told how it works.
Where Randomness Genuinely Improves Games
Procedural generation. Roguelikes and sandbox games use randomness to produce layouts nobody designed and nobody replayed identically. This is randomness creating content rather than dispensing it.
Matchmaking. Random or skill-based pairing reduces the effect of matchmaking order, which is a real and documented source of skill inflation in competitive games.
Asymmetric drafting. Random order in a draft removes first-pick advantage, which is a genuine fairness win — the same reason a real sports league randomises who chooses first.
Using Random Selection at a Game Night
The tools here are useful for the parts of a game night that are not the game: who starts, who plays with whom, which game to open with, and how a campaign decision gets made.
Seeding a tournament. The Tournament Bracket handles elimination rounds including byes. Alphabetical seeding gives the first team a real advantage, which is worth avoiding in anything you intend to win at.
Forming parties and teams. Group Maker with pot balance keeps party sizes equal. A random party is not a balanced one — if you have a strong player, seat them before drawing.
Rolling. The Dice Roller covers d4 through d20, custom faces from 2 to 100, up to 12 dice at once, and modifiers from −99 to +99. It draws from the browser’s cryptographic generator rather than Math.random(), so unlike a worn physical die it stays uniform indefinitely.
Deciding between equal options. A coin flip or the Wheel of Names settles it in two seconds instead of a discussion nobody wanted.
The Fairness Question, Answered Directly
If you want to run a competition among friends where the outcome will matter, the requirements are narrow and specific:
- The pool must be fixed and published before anyone draws
- The weights must be published, if any exist
- The source must not be predictable. A game engine’s drop system can be, in principle, predicted; a draw from
crypto.getRandomValues()cannot. The fairness page explains why that distinction is real - The result must be recorded with the pool visible
That is a complete answer. What you do not need is a complicated certification scheme — you need the list to be visible at the moment of the draw.
The Useful Takeaway
Games teach an intuition that transfers badly: that randomness has a memory. It does not, and a run of bad results means nothing about the next draw.
The one exception worth remembering is pity protection, where the system does remember — and does not tell you. Where a system has one, a “bad streak” is genuinely informative; where it does not, it is a sample you happened to draw.
Either way, the same rule holds: over enough attempts the average is what the weights say. Expect it to take far longer than feels reasonable.