Chess accuracy calculator
Paste a Chess.com or Lichess game link, or a PGN, to calculate both players’ accuracy. Inspect every move and see where the score changed.
Updated
Review your game now
How accuracy is calculated
Accuracy is not a measure of how many moves matched the engine. It is a measure of how much you gave away, and it is computed in three steps.
- Each position's evaluation in centipawns is converted to a winning chance between 0 and 100, using the public Lichess logistic model. This keeps the number fair: throwing away a pawn in a level position matters, and throwing away a pawn when you are a rook up does not.
- Each move's cost is the winning chance lost against the engine's own choice, from the mover's side of the board.
- That cost becomes a per-move score on the public Lichess exponential curve, and a side's accuracy is the mean of its own move scores. The model is an interpretation of the engine’s evaluation, not your personal probability of winning.
Because it averages per-move scores, a game can hold one catastrophic blunder and still read in the eighties. That is intended. Accuracy describes the whole game, not its worst moment. The move labels and the graph are where the worst moment shows up.
A worked calculation
Suppose White’s evaluation changes from +1.00 to 0.00. This is an arithmetic example, not a recorded game. The model converts +100 centipawns to 59.10% and 0 centipawns to 50.00%, a loss of 9.10 percentage points. Applying103.1668 × exp(−0.04354 × loss) − 3.1669 gives a per-move accuracy of 66.24%. If the same side’s only other move scored 100%, the two-move mean would be 83.12%.
The equations come from Lichess’s published accuracy method. RV Chess clamps its per-move values between 0 and 100 and uses a simple mean for the game. Lichess’s own game score combines weighted and harmonic means, so sharing the per-move curve does not make the final game scores identical.
What counts as a good accuracy
Position type matters. A forced sequence or a lopsided position can produce a high score without requiring difficult choices. Compare the errors and threats in your own games, rather than using an accuracy percentage as a rating or a cheating test. Even the same game can score differently with another engine build, search depth or aggregation formula.
The performance rating, and its real precision
Each side also gets a rating estimate for the game. It is fitted to real data rather than guessed: 50 rated games spanning 785 to 2978 were run through this exact pipeline and regressed against the players' actual ratings.
That small calibration had residual error on the order of 600 rating points per game. It is not a confidence interval or a guarantee for a particular player. Games are affected by time control, position type and repeated habits, so averaging ten games does not establish a promised error bound. Treat a single game’s figure as a rough description of that performance.
Calculate the accuracy of any game
Three ways in: paste the PGN from any site, paste a Chess.com or Lichess game link, or look a Chess.com username up and pick from recent games. Stockfish 18 runs inside your browser tab, so analysis stays local unless you opt into contribution, and nothing is metered.
