Base62 short codes and how URL shorteners make keys
Updated Nov 8, 2026 · 3 min read
A short link needs a code that is short, safe to put in a URL, and never handed out twice. Base62 covers the first two. It writes numbers using the 62 characters 0–9, a–z and A–Z, all of which are allowed in a URL path without escaping. The third part, uniqueness, depends on where the numbers come from.
How long does a code need to be?
Each extra character multiplies the number of possible codes by 62.
| Length | Possible codes | Roughly |
|---|---|---|
| 1 | 62 | |
| 2 | 3,844 | ~3.8 thousand |
| 3 | 238,328 | ~240 thousand |
| 4 | 14,776,336 | ~15 million |
| 5 | 916,132,832 | ~920 million |
| 6 | 56,800,235,584 | ~57 billion |
| 7 | 3,521,614,606,208 | ~3.5 trillion |
| 8 | 218,340,105,584,896 | ~220 trillion |
To pick a length, estimate how many links you will ever create (links per day × days you keep them) and take the shortest row that covers it. At 10 million links a day for 5 years you need about 18 billion codes. 6 characters cover 57 billion, so 6 works, with room to spare. At 100 million a day you need about 180 billion, and only 7 characters are enough.
Encoding a number
Base62 works like any other base. Divide by 62 repeatedly and use each remainder as a digit, least significant first.
const ALPHABET = '0123456789abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ';
function toBase62(n: number): string {
let code = '';
do {
code = ALPHABET[n % 62] + code;
n = Math.floor(n / 62);
} while (n > 0);
return code;
}
toBase62(125); // '21' (2 × 62 + 1)
toBase62(1_000_000_000); // '15FTGg'Decoding runs the other way. Start at zero, and for each character multiply the total by 62 and add that character’s position in the alphabet. For ids above 2⁵³ use bigint, since JavaScript numbers lose precision past that point.
Three ways to get the number
| Approach | Collisions? | Trade-off |
|---|---|---|
| Hash the long URL and keep the first 7 characters | Yes | Truncating throws away most of the hash, so different URLs can share a prefix. You must check the database and retry on every clash. |
| 7 random characters | Yes, eventually | Simple and unguessable, but you still have to check before saving. |
| A counter, encoded in base62 | Never | Unique by construction. The work moves to keeping the counter consistent across servers. |
Why random codes still collide
With 3.5 trillion possible 7-character codes it feels like clashes should never happen. The birthday problem says otherwise. The chance that some pair of codes matches passes 50% after only about 2.2 million codes, roughly the square root of the space. Any single new code is still very unlikely to clash (at a billion links, about 0.03%), so retrying is cheap, but the check before saving can never be skipped.
Running a counter across many servers
One shared counter that every server increments on every new link becomes a bottleneck and a single point of failure. The usual fix is to hand out ranges instead of single numbers:
- A small key-generation service owns the counter, backed by a database row or a coordination service.
- Each app server asks it for a block, say the next 10,000 numbers, and gets back a start value.
- The app server hands out numbers from its block in memory, with no network call per link.
- When the block runs low it asks for another one.
If a server crashes, the unused part of its block is simply skipped. Gaps in the sequence cost nothing, since nobody needs the codes to be dense.
An alternative is a time-based 64-bit id (timestamp, machine number and a per-machine sequence packed together), which needs no coordination at all. The cost is length, since a 64-bit number can need up to 11 base62 characters.
Practical details
- Sequential codes are guessable, so anyone can walk through
abc1,abc2and so on. If that matters, scramble the number with a reversible step before encoding, for example multiplying by a large constant that shares no factors with 62⁷ and taking the result modulo 62⁷. - Codes are case sensitive. If links will be read aloud or typed from print, consider base58, which drops look-alikes such as
0,O,Iandl, or base36 if the codes must survive being lowercased. - Reserve words you do not want to hand out (
api,admin, offensive strings) and skip them when they come up.