Chess.com vs Lichess: Why the Same Game Gets Different Accuracy Scores

Analyze one game on Chess.com, on Lichess and in another tool, and you will get three different accuracy percentages. None of them is wrong. They use different formulas, different engines and different rules about which moves count. This article explains the differences and shows how far apart the numbers actually are.

By the RookSac teamSeptember 29, 20264 min read

Same game, three numbers

Accuracy looks like a fact about a game, like the number of moves. It is not. It is the output of a calculation with several choices in it, and every site makes those choices differently. Three things account for almost all of the differences:

1. The formula. How engine scores become win chances, how a move's loss becomes an accuracy, and how the moves are combined into one game score.
2. The engine and its settings. Different Stockfish versions and different search depths disagree about positions, especially sharp ones, and a disagreement about one move shifts that move's score.
3. What is counted. Whether book moves score, whether forced moves count, and how moves played with a few seconds left are treated.

What each tool does

Lichess publishes its method. Engine scores are converted to win chances with a logistic curve fitted to games between strong players, each move's accuracy is an exponential function of the win chances it gave away, and the game score combines an average weighted by how volatile the position was with a harmonic mean, so a few big errors pull the score down more than a plain average would.

Chess.com uses its own scoring system for Game Review, which it calls CAPS, together with its own move-classification rules (this is where Brilliant and Great moves come from). Chess.com describes its approach in general terms, but the exact computation is not published, so it can't be reproduced independently.

RookSac reads games from either site, runs Stockfish 17.1 in your browser, and uses Lichess's published win-chance and per-move curves. For the game-level score it applies its own aggregation: the average of the move accuracies, with the distance from 100 scaled by 2.375. That factor was chosen by fitting to 25 games that Chess.com had already reviewed (50 player scores), where it brought RookSac's numbers to within about three points on average. The details are on the how the analysis works page.

How far apart are the numbers?

The three bundled sample reports include Chess.com's own accuracy for every game, so we can compare directly. Across 300 real games (600 player scores; the games were not used to set the calibration), the difference between RookSac's accuracy and Chess.com's was:

So two tools that are deliberately aligned still disagree by about 3 points on a typical score, and by more than 5 points about 12 times in 100. Two tools that make no attempt to align can be expected to differ at least as much. A gap of a few points between two sites is not an error in either one.

Why individual games differ so much

• Sharp positions. When a position is tactically complicated, two engines (or one engine at two depths) may rate a move differently, and a single move can swing the average.
• Long won or lost endings. Once a game is decided, small differences in how a curve flattens add up over many moves.
• Time trouble. Sites treat moves made with almost no time left differently. RookSac leaves out flag-race moves (started with under two seconds and no increment); other tools may count them.
• Book moves. Tools disagree on what counts as opening theory, and a book move is usually credited as perfect.

Move labels differ too

The same holds for labels. Lichess marks inaccuracies, mistakes and blunders and has no Brilliant or Great labels. Chess.com adds those, and each site chooses its own thresholds for what counts as a mistake. RookSac's thresholds are listed on the move classifications page. If a move is a "blunder" in one review and a "mistake" in another, look at the win-percentage drop behind it: the two tools may agree about the position and only disagree about where to draw the line.

How to compare sensibly

1. Don't mix sources. Track your progress with one tool. A rise from 78 to 82 means something if both numbers come from the same method, and nothing if the first came from Lichess and the second from Chess.com.
2. Compare inside the game. Your accuracy against your opponent's, from the same tool, is far more informative than your accuracy against a number from somewhere else.
3. Treat small gaps as noise. Under about three points, two numbers for the same game are the same number.
4. Look behind the percentage. Errors per 100 moves and the list of mistakes tell you more than any single accuracy figure. See what accuracy really means.
5. Use many games. One game's accuracy is noisy on any platform; see why reviewing many games beats reviewing one.