Quick answer
Quick answer: An accuracy score measures how much your winning chances dropped with each move, judged by an engine. It isn't a measure of how good your moves were in the abstract, and two sites using different formulas will give the same game different numbers.
One of the big sites publishes its equation. The other doesn't.
I read the published formula rather than describing the idea from memory, because the detail changes how you should read your own score.
What the number is actually measuring
Every accuracy system works from an engine evaluation of the position before and after your move.
The drop in your position between those two evaluations is what gets scored. A move that changes nothing scores near perfect. A move that hands over a win scores near zero.
So the score's a measure of damage, not of insight.
A brilliant move and an obvious one both score the same if neither loses anything, which is the first thing people misunderstand.
The published formula
Lichess publishes both steps, which makes it the useful example.
First it converts the engine's centipawn evaluation into a winning percentage, using a curve fitted to real games played by 2300 rated players.
Then it scores each move by how far that winning percentage fell, through a second equation that turns the drop into a number out of a hundred.
The accuracy page has the exact equations if you want to reproduce them.
Why the 2300 detail matters
That curve was fitted to games between strong players, and this is the most interesting thing in the whole topic.
It means a winning percentage is answering the question of what happens when two 2300 players continue from here. Not what happens when two 1200 players do.
At club level a position that a strong player converts easily is often drawn or lost anyway, because converting is hard.
So your accuracy score is measured against a standard of play that neither you nor your opponent was ever going to reach.
The other formula is private
I checked what the other big site publishes, and the answer is a description rather than an equation.
Its score is called CAPS2, short for Computer Accuracy Precision Score, and moves are compared against the engine's top recommendations. The company says the newer version was designed to feel more like being graded on a test, because under the old one almost everyone scored high.
That's a reasonable design goal and it makes the number more readable.
It also means you can't reproduce or audit it, so two numbers from two sites are not comparable.
Why your accuracy is higher in short games
This surprises people and the reason is structural.
A game with fewer moves doesn't give you many chances to lose ground, and quiet positions produce tiny evaluation changes.
So a short quiet draw can score higher than a hard fought win, and a long grinding endgame will drag your number down even when you played well.
Accuracy has a length bias, and nobody puts that on the results screen.
Why a won position lowers your score
Once you're clearly winning, the engine has very high standards and you don't need to meet them.
Giving back half your advantage while still being completely winning is a big evaluation drop, so the move scores badly. In practice it cost you nothing.
This is why players report their accuracy falling in games they won comfortably, and why a safe simplifying move that wins easily can be marked as an inaccuracy.
The score doesn't know that a five pawn advantage and a three pawn advantage both win. It only sees the gap between them.
In a real game, taking the simple win is correct and the engine's preference is irrelevant.
Engine depth changes the answer
The evaluation comes from an engine, and engines change their minds as they look deeper.
A quick analysis pass and a deep one will label different moves as mistakes, and both are honest. The shallow pass isn't wrong, it simply hasn't seen as far.
So the same game analysed twice can produce two different accuracy figures, on the same site, with no bug involved. Engine versions change too, and a game analysed a year apart can be scored by two different programs.
If the difference matters to you, run the deeper analysis. If it doesn't matter, stop reading the number so closely.
What counts as a good accuracy
There isn't a fixed answer, and the ranges people quote are usually invented.
What you can say is relative. Your own accuracy across twenty games, compared to your accuracy across twenty games last month, means something.
A single game's figure means very little, because the opening theory you knew and the length of the game both move it. Play a twelve move miniature you've memorised and the number will flatter you badly.
It also shifts with the time control. Blitz accuracy and classical accuracy aren't the same measurement, and comparing them tells you about the clock rather than about you.
Treat it like a weekly weight measurement rather than a daily one.
The cheating question
High accuracy is the reason people accuse each other, and on its own it proves nothing.
A short game in a known opening line can produce a very high figure from a normal club player, because they played moves they memorised.
Detection systems look at far more than one number, including timing patterns and how the accuracy is distributed across the game.
We went through what actually gets used in is my opponent cheating.
Centipawn loss is the older measure
Before accuracy percentages, everyone used average centipawn loss, and some players still prefer it.
It's simply the average amount of evaluation you gave up per move, in hundredths of a pawn.
Lower is better, and there's no upper bound, so it never compresses the way a percentage does.
It's harder to read and more honest, which is a familiar trade in every kind of measurement.
How to use the number properly
There's a version of this that helps you:
- Ignore the game level figure and look at where the drops happened.
- Check whether your worst moves cluster in the opening, the middlegame or the endgame.
- See whether they cluster in time trouble.
That tells you what to work on. The headline percentage doesn't.
What it cannot tell you
Accuracy has no idea what you were trying to do. It doesn't know your plan and it never did.
It can't tell a deliberate practical choice from a blunder, and at club level practical choices are often correct even when the engine disagrees.
It also can't see that a move was hard to find. A difficult only move and an obvious recapture score identically.
So it measures outcome, never process, and improvement is a process problem.
Should you chase a higher score?
No, and chasing it will make you worse.
Playing for accuracy means avoiding complications, because complications are where evaluation swings live. Avoiding complications is how players stop improving.
The players who climb take on positions they might get wrong, and their accuracy scores look worse for a while.
If your rating is stuck, that pattern is worth reading about in breaking out of a rating plateau.
Two sites will not agree
Worth knowing before you compare your number with somebody else's.
Each site runs a different engine, at a different depth, through a different formula. The same game analysed in two places can come back several points apart.
Neither one is wrong. They're measuring slightly different things and presenting both as a percentage, which invites a comparison the number can't support.
So compare your accuracy only against your own, on the same site, at the same time control. Across sites the figure means very little.
What to do now
Read your next analysis by looking at where you lost ground, not at the percentage.
If you want the underlying idea, engines and their evaluations are described plainly in the chess engine reference.
Then go and play a game where you deliberately allow complications, and expect the score to drop. You can start one on the play page.
The score falling while you improve is normal, and it's the part nobody warns you about.