What Chess Accuracy Really Means (and What a Good Score Is)

You finish a game and the review says Accuracy: 84.2. Is that good? It sounds like a school grade, and it is easy to read it as "84% of my moves were right". It doesn't mean that. This article explains where the number comes from, so you can decide what to do with it.

By the RookSac teamSeptember 29, 20265 min read

From engine score to winning chances

An engine such as Stockfish scores a position in centipawns: hundredths of a pawn, from the side to move's point of view. A score of +150 means "about a pawn and a half better". But a pawn is worth different amounts in different positions, and a raw score is a poor guide to how likely you are to win.

So RookSac, like Lichess, converts the score into a win percentage using a smooth S-shaped curve: win% = 50 + 50 × (2 / (1 + e−0.00368208 × cp) − 1). The curve is steep near equality and flat when a side is already winning: the tenth pawn hardly changes anything, while the first fifty centipawns matter a lot.

This is the same idea behind the evaluation bar on most sites, and it is discussed further in how to read an engine evaluation.

From winning chances to a move's accuracy

Each move is judged by what it did to your winning chances. The win percentage before your move (with the best play available) is compared with the win percentage after it. The difference is the drop. A move that keeps your chances where they were has a drop of zero; a move that turns a 70% position into a 40% one has a drop of 30 points.

The drop is turned into a score from 0 to 100 with an exponential curve: accuracy = 103.1668 × e−0.04354 × drop − 3.1669. The curve is unforgiving, as this table shows:

A move that gives away only 5 win% points, a small slip, already scores about 80 rather than 91. That harshness is deliberate: it makes the difference between a game of small slips and a clean game visible. See move classifications for how the same drops map to labels.

From moves to a game score

The game's accuracy is built from the average of its move accuracies. RookSac then adjusts it: the distance of the average from 100 is multiplied by 2.375, because Stockfish 17's evaluations sit closer together than the older engine's, and on a set of 25 games reviewed by Chess.com this scale made the numbers match theirs closely (a mean error of about three points). In short, game accuracy = 100 − 2.375 × (100 − average move accuracy). The details are on the how the analysis works page.

Three details make the number behave in ways that surprise people:

• Book moves count as fully accurate. Opening theory is not scored against the engine, so a game with 15 book moves starts with 15 perfect scores.
• Forced moves are perfect. If there is only one legal move, you get full credit.
• Flag-race moves are left out. With about two seconds left and no increment, a premove decides the game, not judgement, so those moves keep their label but don't count towards accuracy.

What 90% does and doesn't mean

Suppose a player makes 40 scored moves. The three invented games below show what the average rewards:

The one-blunder game scores higher than the steady game even though the blunder probably lost it. Accuracy measures the average quality of your play. It doesn't measure how decisive your mistakes were, and it says nothing on its own about who won. That is why a chess review shows the error count and the critical moments next to the percentage. A single accuracy number can hide the one move that mattered.

Two more consequences follow:

• A lopsided game inflates accuracy. When a position is completely won, several moves keep the win% near 100 and all of them score full marks, even if one is much cleaner than another. Long, easy conversions push accuracy up.
• Accuracy is not comparable across game types. A calm 60-move game and a sharp 25-move game have different natural ranges. Compare your accuracy over many games in the same time control.

What is a good score?

There is no universal answer, and anyone who gives you a single number ("above 90 is great") is oversimplifying. What the data do show:

• Even elite players are nowhere near 100. In the three RookSac sample reports (100 real Chess.com games each, players rated about 2,900–3,400 in blitz), average accuracy was 84.9 to 90.4, and their opponents averaged 82.3 to 86.7. See the sample report.
• Accuracy depends on time control. Bullet and blitz produce lower numbers than classical because moves are made quickly.
• It depends on the opponent. A strong opponent makes every game sharper, which lowers both players' accuracy.

So the useful comparisons are your accuracy against your opponents' in the same games, and your own average over the last 50 games against the previous 50. Both remove most of the noise that makes a single number meaningless.

Accuracy versus average centipawn loss

You may also see ACPL, average centipawn loss: how many hundredths of a pawn each move gave away, on average. ACPL is easy to understand but has two weaknesses. A large centipawn loss in an already-winning position (going from +9 to +6) is huge in centipawns and irrelevant in practice, and a mating score can produce an enormous single-move value. Accuracy avoids both because it works in win percentage, which flattens when a side is already winning. In the RookSac samples, ACPL ran from 17.5 to 27.3; the two measures agree on order but not on size.

How to use the number

1. Treat one game's accuracy as noise. See why one game means little.
2. Look at accuracy by phase to find where you lose the most.
3. Compare with your opponent's accuracy in the same game.
4. Watch the trend over dozens of games, and pair it with the count of mistakes and blunders.
5. Don't optimise the percentage. Improve the moves behind it: the mistakes and blunders that a review lists.