An accuracy score is not the percentage of moves you got right. Here is what it is really counting, and why two engines can rate the same game ten points apart.
Almost every game review hands you a percentage. It feels like a mark out of a hundred, so most players read it that way: 85% means I played 85% of my moves correctly. That is not what it measures, and the gap between what people assume and what the number does matters, because it changes which games you should worry about.
Behind every accuracy score is one question asked on every move: how much did this move cost you compared with the best available? The engine evaluates the position before your move and after it. The difference is your loss on that move. Play the top move and you lose nothing. Hang a rook and you lose a great deal. Your accuracy for the game is a summary of those per-move losses, not a count of how many times you matched the engine.
This is why two players can both score in the high eighties while playing completely different chess. One may have found six only-moves in a sharp game. The other may have shuffled pieces in a quiet position where nothing was ever at stake. Neither gave much away, so both score well.
Raw centipawn loss punishes the wrong things. If you are winning by nine pawns and play a move that drops the evaluation to six, you have lost three pawns of material on paper. In practice you have given up nothing that matters: the game is just as won. Meanwhile, an error that takes a position from level to a pawn down can be the mistake that loses the game.
So modern accuracy scores convert the evaluation into a win probability first, then measure how much of that probability your move handed over. Going from +9.0 to +6.0 barely moves your chances of winning. Going from 0.0 to -1.0 moves them a lot. This is closer to how a human judges a mistake, and it is why the same blunder is punished differently depending on when it happened.
The same logic has an awkward consequence. Once a position is completely won or completely lost, win probability stops moving, so mistakes there cost almost nothing. A player who wins a piece early, then plays the rest of the game sloppily, can still score in the eighties. Long scrappy games where both sides are lost several times over are the worst case: the numbers can look far kinder than the chess deserved.
We ran into this directly. Comparing our own scores against Chess.com's on a few hundred analysed games, ordinary games agreed closely, but games where a quarter or more of the moves came from already-decided positions read several points too generously. The fix was to keep charging for a large evaluation drop even when the win probability has flattened out. If your own review ever tells you a shambolic game was 80% accurate, saturation is usually the reason.
There is no official accuracy. Every site invents its own formula, and the differences are real:
None of this makes any of them wrong. It does mean comparing your accuracy from one site with a friend's from another is meaningless, and that chasing small differences is a waste of time.
Accuracy is a screening tool, not a verdict. It is good at three jobs:
It is bad at deciding whether a game was well played. A 72% game full of sharp defence can be better chess than an 88% game where nothing happened. Look at what the mistakes were and when they came, not just how many points the summary knocked off.
Statmate calculates accuracy for every game it analyses with Stockfish 18, and the Accuracy tab puts your whole history in one place: your average, the trend over time, your accuracy by opening, and every reviewed game sorted so the worst are easy to find. Games reviewed on Chess.com also show that site's own figure, so you can see where the two methods disagree instead of wondering which is broken.
The board itself shows the same numbers per game, with the classification of every move and the evaluation graph beside them, so a low score always comes with the moves that caused it.
Accuracy measures how much winning chance you gave away, weighted so that mistakes in positions that were still alive count for more. It is comparable within one system and meaningless across two. Use it to find the games worth reviewing, then let the moves tell you what actually went wrong.