How the guess works.

Split the PGN into games and moves

A grammar-based PGN parser splits your export into games and strips comments, clock times, variations and annotations, leaving only the moves that were actually played. Each game is then replayed move by move to make sure every move is legal.

Ask Stockfish about every move — twice

For each position, Stockfish searches twice: once freely, to find the best move and its evaluation, and once restricted to the move you played. The gap between the two is your move's centipawn loss (100 centipawns = one pawn). Because both searches look at the very same position, quirks of the engine's evaluation cancel out — and if you played the engine's first choice, the loss is exactly zero.

After 1. e4 e5 2. Nf3 Nc6 — the engine prefers Bb5 (solid ring), you played Nc3 (dashed ring). Both fine moves; the difference is small.

Keep only the moves that say something

Two kinds of moves are excluded before averaging. The first 4 moves of each side are usually memorized opening theory, and any move played in an already-decided position (more than ±8 pawns) is garbage time — finding the fastest mate, or shuffling pieces in a lost cause, says little about skill.

One game's evaluation over time — shaded moves don't count toward the average.

Map average loss to a rating

Your average centipawn loss (ACPL) over all rated moves is looked up in a calibration curve fitted against real chess.com players of known rapid rating, measured with this exact pipeline. Give away ~15 cp per move and you play like a grandmaster; give away ~95 and you're around 900.

The actual lookup table used by the app (dots are anchor points; values in between are interpolated).

Adjust for the clock

Everyone plays worse with less time — so the same move quality means a stronger player in bullet than in classical or daily games. The estimate is shifted by the time control found in the PGN.

Say how sure we are

Each game also gets its own mini-estimate. How tightly those cluster — and how many games and rated moves went in — decides the ± range and the confidence label. One brilliant game proves little; twenty consistent ones prove a lot.

All together

And the honest fine print: this is a well-informed guess, not a measurement. Style matters — solid positional players measure a bit high, wild attackers a bit low — and site ratings sit on different scales. Expect the truth within a couple hundred points.