Accuracy is the worst number in your game report

The short answer

Your accuracy percentage is a summary statistic dominated by two or three moves you already knew were bad. It tells you the size of the damage, not where it came from, and it swings for reasons that have nothing to do with how well you played — game length, how forcing the position was, how long you stayed in a book line, how deep the engine ran.

Read the classified move list instead, in this order: which moves were mistakes or blunders, what phase of the game they happened in, whether your opponent punished them, and what the position looked like immediately before them. That is four pieces of information the percentage throws away.

Why the number is unstable

Accuracy is derived from how much evaluation you gave away per move, converted through a win-probability curve and averaged. Neither Chess.com nor Lichess fully publishes its formula, and the two do not produce the same number for the same game. So a cross-site comparison is meaningless, and a comparison of your Tuesday number to your Wednesday number is close to meaningless too.

Three things push accuracy around independently of your play. Forcing sequences inflate it: if there is one legal recapture, you played the best move, and the report credits you. Quiet positions inflate it too, because in a symmetrical structure most reasonable moves lose almost nothing. Sharp positions deflate it, because when three moves are within 40 centipawns of each other and the fourth loses on the spot, you get punished hard for a defensible choice.

The result is that a grinding 70-move endgame win where you never had to think can score higher than a game where you calculated a sound sacrifice and got one move of the follow-up slightly wrong. This is opinion, but I would rather play the second game.

Roughly one move in 13 is a real error

Across 107,460 engine-classified moves from 3,437 analysed games on Mated, 3.8% of moves played are outright blunders and another 4.2% are mistakes. That is one real error roughly every 13 moves, before you count the 18.3% classified as inaccuracies.

Put that against a typical game. If you play 40 moves, expect three or four moves that are a mistake or worse. Those are the moves that decided the game. Everything else in the report is texture.

This is why the percentage feels uninformative even when it is technically correct. It is computed mostly from the 92% of moves where nothing happened, then shifted by the handful where everything did. You cannot reverse the arithmetic to find out which was which. You have to open the move list.

The average centipawn loss lies in the same way

In the same set of 107,460 moves, the average move gives away 95 centipawns against the engine's best. The median is far lower, because a small number of very bad moves carry most of the damage.

That gap is the whole problem with single-number reporting. A mean pulled up by outliers describes a player who does not exist: nobody plays 40 moves each about a pawn short of best. You play a long run of moves that cost almost nothing and then you drop a rook.

So when your average centipawn loss goes from 45 to 30 across ten games, you have not become 33% more accurate. You have most likely blundered less often, or blundered smaller. That is worth knowing, but it is a different claim, and the number does not distinguish them.

Where the errors actually are

Errors are not spread evenly through a game. In that same set of 107,460 moves, 9.0% of middlegame moves are a mistake or worse, against 2.2% in the endgame.

That ratio should change what you look at. If your report shows a big accuracy drop, the useful question is not "how much" but "which moves" — and the prior says to start with the moves right after the last book move and before material simplifies. That is where the position stops having a plan you already know.

One caveat on the endgame figure: fewer errors per move does not mean endgame errors matter less. A 2.2% error rate on a move that decides a queen-versus-rook race costs you the game outright, where a middlegame inaccuracy often costs you a tempo. Frequency and cost are separate axes and no report I have seen combines them well.

This is the part of a report that is tedious to assemble by hand across twenty games, which is what Mated does — it runs Stockfish over every position from your last games and groups the errors by phase and pattern rather than handing you one figure per game.

A worked example

Invented numbers, but the shape is one you will recognise. You play 41 moves. Thirty-nine of them are within about 15 centipawns of best, which is 585 centipawns of drift in total. On move 19 you allow a fork and lose 350. On move 33 you hang the exchange and lose 620.

Total loss: 1,555 centipawns. Average per move: about 38. That will report as a respectable accuracy, and the two red marks in the move list account for 62% of everything you gave away.

Now the questions worth asking, none of which the percentage answers. Was move 19 a calculation error or did you not consider the opponent's knight at all? Was move 33 played with 40 seconds left? Had you seen the same fork pattern in a previous game? Did your opponent actually take the exchange, or did the engine notice and neither of you did?

The last one matters more than people expect. If your opponent missed it, your rating did not move, your win-loss record did not move, and the habit is still there waiting. Unpunished blunders are the ones you never learn from, and they are invisible in any single-number summary.

What to open first, and in what order

A practical order for a post-game report, whether you are using a site's built-in engine review or something else:

• Filter to your own mistakes and blunders. Ignore your opponent's; they are not your habits.
• Note the move number of each and which phase it fell in. Three middlegame blunders in a row across different games is a pattern; one endgame slip is noise.
• For each one, look at the position two moves earlier, not the blunder itself. The blunder is usually the consequence of a plan that was already wrong.
• Check the clock at the moment of the error. Time-pressure blunders and calculation blunders need different fixes and reports do not separate them.
• Check whether it was punished. Unpunished errors go at the top of the list, not the bottom.
• Only then look at the accuracy figure, and only to sanity-check that the game was roughly as clean as it felt.

What engine analysis still cannot tell you

An engine knows the evaluation of every position. It does not know what you were looking at. It cannot distinguish "I saw the knight and miscalculated" from "the knight never entered my head", and those two failures need opposite training. Nothing in current AI game analysis solves this; you have to supply the introspection yourself, ideally within a few hours while you still remember.

Nor is there good public evidence on transfer — whether drilling the specific position type where you keep collapsing actually reduces the error rate in your next fifty games, or by how much. Everyone in chess training assumes it does, including me. Nobody has measured it well on ordinary club players, and you should treat confident numbers about it with suspicion.

One thing you can measure yourself, cheaply: count blunders per game over a rolling twenty games and write the number down. It is a coarser statistic than accuracy but it is stable, comparable to itself, and it tracks the thing that decides your games.