Chess Move Classifications Explained: What Every Label Actually Means

By Yuval Incze, Founder of Chess DNA, competitive chess player & chess improvement researcher · Published Sep 27, 2026 · Updated Sep 27, 2026 · ~8 min read

Every game review hands you a column of labels. Brilliant, excellent, inaccuracy, blunder. Almost nobody is told what they are measuring, so they get read as praise and scolding instead of as numbers. They are numbers. This page gives you the exact one behind each label, where the boundaries sit, and why the same move can be an inaccuracy on one site and a mistake on another.

What a move label actually measures

An engine reports a position in centipawns: hundredths of a pawn, from the side to move's point of view. Plus 300 means you are a piece up, roughly. That number is useful for comparing moves, and misleading for judging them, because centipawns are not evenly spaced in terms of results.

Consider two moves that each drop the evaluation by 200 centipawns. The first takes a position from dead equal to two pawns down. The second takes a position from plus 900 to plus 700, which is winning either way. The same 200 centipawns cost you a great deal in the first case and almost nothing in the second. Any grading system built on raw centipawn loss treats them identically, and gets the second one badly wrong.

So the evaluation is first converted into a winning chance, a percentage, using a curve that flattens out at both extremes. The difference between your winning chance before the move and after it is what the label is based on. Chess DNA uses the conversion published by Lichess, a sigmoid that depends on the evaluation and nothing else.

This is also why the labels do not track material. Hanging a rook when you are already lost will often be graded a mistake rather than a blunder, because there was little winning chance left to lose. A quiet pawn move from an equal position that walks into a mating net is a blunder, even though nothing was captured. In our analysis of 3,540 engine-analyzed rapid games, the average game's single worst move cost 754 centipawns for players under 600 and still 506 centipawns, a whole rook, for players at 1500 and above. Those are the moves the labels exist to find.

Every label, and the exact threshold

These are the boundaries Chess DNA uses, at Stockfish depth 18. They are the same ones the app applies to your imported games and the same ones the free game review tool applies to a pasted PGN.

Two labels you will see elsewhere are deliberately absent here. Chess DNA does not mark book moves, because it runs no opening database: an opening move is graded like any other move, on what it did to the position. And it does not award a great move label, the one some sites give to a single move that is much better than every alternative. It is a real idea, but it is a second measurement layered on the first, and we would rather show you one number that means one thing.

The boundaries themselves are not arbitrary or unique to us. They match the ones Chess.com publishes for its Expected Points model. What sits underneath them is where the systems part company, which is the next section but one.

What makes a move brilliant

Brilliant is the most misread label in chess, mostly because people assume it means the engine was impressed. Engines are not impressed by anything. A brilliant move is a sacrifice that survives inspection, and three conditions all have to hold:

1. You genuinely gave up material. Not an even trade, and not a capture the opponent has to recapture. The whole exchange sequence on that square has to end with you down at least a minor piece.
2. The move is still best or excellent. Under 2 percent of winning chance lost. If the sacrifice costs you real ground, it is not brilliant, it is just a sacrifice.
3. It leaves you at least equal. This is the condition that does the real work.

That third one deserves an explanation, because it is the difference between a label that means something and a label that fires constantly. Winning chance saturates at both ends of the scale. When a position is already completely lost, your winning chance is near zero, and throwing a rook away cannot push it much lower, so the move registers as almost no loss at all and sails past the first two conditions. Without a check on the resulting position, every desperate hang in a lost game would come back marked brilliant. Requiring the position afterwards to be at least equal removes them all.

The practical consequence: brilliancies are rare, and they should be. If a review hands you several per game, the system generating them is not measuring what it claims to.

Why the same move gets a different label on each site

Run one game through three review tools and you will get three slightly different verdicts. This is not a bug in any of them. They are different instruments.

The probability model differs. This is the big one. Chess.com's Expected Points is rating-aware: its documentation says the evaluation needed to count a position as winning, equal or losing varies with the player's rating, on the reasoning that a 500-rated player's practical chances in a given position are not a grandmaster's. Lichess converts the evaluation with a fixed formula and does not look at who is playing. Chess DNA uses the Lichess conversion. So the same engine line, the same evaluation, can sit either side of a threshold depending on which model graded it, and the gap widens at the rating extremes.

The depth differs. Evaluations move as the search gets deeper. A shallow search misses a resource three moves out and marks a sound move as an error; a deeper search does not. Chess DNA analyzes every move at depth 18, the same setting for every game, which is what makes the trend across your own games meaningful. Comparing a label produced at one depth with a label produced at another is comparing two different measurements.

The engine differs. Version, hardware and time budget all move evaluations a little, especially in sharp positions where the tree is wide.

How to actually use the labels

The labels are a filter, not a report card. Read that way, they save a great deal of time; read as a score, they mostly produce a feeling.

1. Go straight to the two or three worst moves. Sort by win chance given away and stop after the top few. In a typical game two or three moves account for the entire result, and studying the rest is studying noise.
2. Name what you missed, not what you should have played. The engine gives you the move. You have to supply the category: a hanging piece, an open back rank, a fork you never looked for. The category is the part that transfers to the next game.
3. Count across twenty or thirty games, not one. A single game cannot tell a bad day from a habit. The theme that keeps reappearing is your actual weakness, and it is usually not the one you would have guessed.
4. Watch the blunder count, not the accuracy. Across our 3,540 game dataset, blunder rate explained 58 percent of the variance in player rating and average centipawn loss 72 percent, while accuracy explained 31 percent and inaccuracy rate just 12 percent. The big errors are where the rating lives.

That last point is worth sitting with, because accuracy is the number every platform puts in the largest font. It is a genuine measurement, and we explain exactly how it is computed in how chess accuracy is calculated. It is simply not the number that best tracks how well you play. The count of moves that lost 20 percent or more of your winning chances is.

Blunders also fall more slowly than most players expect. In the same dataset, blunders per game went from 1.66 under 600 Elo to 0.90 at 1500 and above: a real improvement, and still nowhere near zero. Getting better does not mean you stop collapsing. The full breakdown is in how chess games are actually lost, and if the collapses in your own games feel random rather than explainable, why you keep blundering in chess covers the three causes and what to do about each.

For the wider vocabulary, centipawns, the eval bar, mate scores and search depth, see chess engine analysis explained, and the short definition of a blunder lives in the glossary entry. The standard written annotation symbols, the ones with question marks and exclamation marks that predate engines by a century, are catalogued in the Wikipedia entry on chess annotation symbols.

Frequently Asked Questions

What is the difference between a mistake and a blunder in chess?

The difference is a threshold, not a feeling. A mistake gives away between 10 and 20 percent of your winning chances compared with the engine's best move. A blunder gives away 20 percent or more. Both are measured in win probability rather than material, which is why hanging a rook in an already lost position is often graded as a mistake, while a quiet pawn push that lets in a mating attack from an equal position is graded as a blunder. The label follows the damage to your result, not the size of the piece.

What counts as a brilliant move in chess?

A brilliant move is a sacrifice that the engine still approves of. Three things have to be true at once: you give up real material, the move is still best or near best, and the position it leaves you in is at least equal. That last condition is what stops the label being meaningless. Without it, throwing a piece away in a position that was already hopeless would score as brilliant, because when you are completely lost your winning chances can barely fall any further. Brilliancy is rare by design.

Why do Chess.com and Lichess grade the same move differently?

Because they convert the engine evaluation into winning chances differently. Chess.com's Expected Points model adjusts for the player's rating, so the same evaluation counts as more or less winning depending on who is playing. Lichess uses a published formula that depends on the evaluation alone. Search depth, engine version and hardware differ too. A label is a measurement made with a particular instrument, so comparing labels across sites is only meaningful if you know both instruments agree.

Is an inaccuracy in chess bad?

Barely, on its own. An inaccuracy costs between 5 and 10 percent of your winning chances, which in most positions is recoverable within a few moves. It matters as a count rather than as an event. In our analysis of 3,540 rapid games, inaccuracy rate explained only 12 percent of the variance in player rating, the weakest of every metric we tested. If you are choosing what to study, the blunders are where the rating is, not the yellow marks.

What does a miss mean in a chess game review?

A miss is a specific kind of error: your opponent has just blundered, you were already clearly better, and instead of punishing it you let the advantage go. It is flagged separately from an ordinary mistake because the lesson is different. An ordinary mistake means you did not see a problem. A miss means the problem was your opponent's, the answer was on the board, and you did not look for it. Missed punishments tend to repeat, which is why they are worth tracking on their own.

How many blunders per game is normal?

More than most players expect, and it never reaches zero. Across 3,540 engine-analyzed rapid games, players under 600 averaged 1.66 blunders per game and players at 1500 and above still averaged 0.90. The average game's single worst move cost 754 centipawns under 600 and 506 centipawns at 1500 and above, which is a rook. Improving does not mean you stop collapsing. It means you collapse less often, and recover better when you do.

Can I get move classifications for my games for free?

Yes. Chess DNA grades every move of any game at Stockfish depth 18 with no account and no daily limit: type your Chess.com or Lichess username and pick a game, or paste a link or a PGN. Lichess computer analysis is also free and unlimited. Chess.com gives free members a limited number of Game Reviews per day. All three will label the same game slightly differently, for the reasons set out above.