How My Chess Review works

This page explains what happens during a chess game review on My Chess Review: what a PGN is, which engine checks your moves, how the accuracy score is worked out, and how each move gets its grade.

What is a PGN?

PGN stands for Portable Game Notation. It is a plain text format for recording chess games, and nearly every chess site, app and database can read and write it. A PGN has two parts: tags in square brackets that describe the game, and the moves themselves.

[Event "Casual game"]
[Site "Paris FRA"]
[Date "1858.??.??"]
[White "Paul Morphy"]
[Black "Duke Karl of Brunswick and Count Isouard"]
[Result "1-0"]

1. e4 e5 2. Nf3 d6 3. d4 Bg4 4. dxe5 Bxf3 5. Qxf3 dxe5
6. Bc4 Nf6 7. Qb3 Qe7 8. Nc3 c6 9. Bg5 b5 10. Nxb5 cxb5
11. Bxb5+ Nbd7 12. O-O-O Rd8 13. Rxd7 Rxd7 14. Rd1 Qe6
15. Bxd7+ Nxd7 16. Qb8+ Nxb8 17. Rd8# 1-0

The moves use standard algebraic notation. Each square has a name from a1 to h8. Pieces are written K (king), Q (queen), R (rook), B (bishop) and N (knight), and pawn moves have no letter. An x means a capture, + means check, # means checkmate, and O-O and O-O-O are castling on the king side and the queen side.

You do not need the tags to start a review. A list of numbered moves is enough.

The engine: Stockfish in your browser

Stockfish is a free, open-source chess engine and one of the strongest chess programs ever made. My Chess Review runs a version of Stockfish that works inside a web page, so the analysis happens on your own computer or phone rather than on a server.

For each position, the engine searches ahead through possible moves and replies. How far it looks is called depth. More depth gives a more reliable verdict but takes longer. The review searches every position to depth 18, or for at most 0.4 seconds, whichever comes first. A game of 40 moves has 81 positions, so the main pass takes at most about half a minute, however slow the device. When the engine's top move seems to lose a lot in the next position, the review checks those two positions again with a deeper search. Because the work happens on your device, a review costs us nothing to run, which is why there is no limit on how many games you can review.

How accuracy is calculated

Accuracy turns a whole game into one number from 0 to 100 for each player. It is built in four steps.

  1. The engine evaluates the position before and after every move.
  2. Each evaluation is turned into winning chances from 0% to 100%. A pawn matters far more in a level position than in one that is already won, and winning chances capture that difference. The review uses the same curve as Lichess: an advantage of one pawn is about 59%, three pawns about 75%, and a forced mate counts as about 98%.
  3. For each move, the review measures how many percentage points the move lowered your winning chances, compared with the engine's verdict on the position before it. A move that keeps them where they were scores 100. A move that loses 5 points scores about 81, one that loses 10 points about 65, and one that loses 20 points about 41.
  4. The scores for all of your moves are combined into your accuracy for the game. As on Lichess, this is the average of two averages: one that gives more weight to the sharp moments of the game, and one that is pulled down hard by any single bad move.

A few things are worth knowing. Accuracy is not a rating and does not predict your rating. A short game with one big blunder can still show a high number if every other move was sound. And because every tool uses its own method, accuracy from different sites is not directly comparable. Compare your own games with each other instead.

How moves are graded

Every move gets one of ten grades. Six of them are decided only by how many percentage points the move lowered your winning chances. The cut-offs are fixed: the same for every game and every player, so the same loss always earns the same grade. They follow the ranges Chess.com publishes for its own grades.

Best
The engine's top move. Only the move the engine would have played itself gets this grade.
Excellent
Loses 2 points or less. In practice it is as good as the best move.
Good
Loses more than 2 and up to 5 points. The position stays healthy, and this is how most solid moves look.
Inaccuracy
Loses more than 5 and up to 10 points. Often a passive move or a slow plan.
Mistake
Loses more than 10 and up to 20 points. Typical causes are a missed tactic, a weakened king or a pawn given away.
Blunder
Loses more than 20 points, often turning a win into a loss. Hanging a piece and allowing mate are the classic examples.

The other four grades have rules of their own. Chess.com describes these grades in words but does not publish its numbers, and it adjusts them to a player's rating, which this review does not know. The rules below are ours. They aim at the same idea, but the same move can get a different grade here than it does there.

Brilliant
A piece sacrifice that is the right move. The move is the engine's first choice or within 2 points of it, and it leaves a knight, bishop, rook or queen where the opponent can come out at least 2 points of material ahead by taking it, counting a pawn as 1, a knight or bishop as 3, a rook as 5 and a queen as 9. You are at 40% or better after it, and it is at least 5 points better than any other move. It does not count when the next-best move would already leave you at 85% or more, or when the same piece was already on offer a move earlier.
Great
The only good move. It is the engine's first choice, and every other move would have lost 20 points or more, so anything else would have been a blunder. You are at 40% or better after it. Simple recaptures, captures that just pick up 2 points of material or more, and moves that give checkmate stay Best.
Miss
A missed chance. Your opponent's last move lost more than 10 points and gave you a winning position, 75% or more, that you did not have before. Your reply then lost more than 5 points and left you below 75%.
Book
A standard opening move. The review carries a list of about 280 lines from well-known openings, from one move to about a dozen moves deep. A move is Book while the game stays on that list, also when the moves come in a different order. The list is short, so a normal opening move that is not on it simply gets one of the other grades. Games that start from a set-up position have no Book moves.

A few moves are treated differently. Moves where you had only one legal choice are not graded. When a game is already decided, with one side's winning chances at 97% or more before and after the move, or at 3% or less before it, the grade stops at Inaccuracy. A move that gives checkmate is always graded Best. To find Brilliant and Great moves, the review needs to know how good the next-best move was, so it looks a second time at up to 24 positions with a shorter search: depth 14, or at most 0.2 seconds each. All of these moves still count toward accuracy.

Key moments

A long game has a lot of moves and only a few that decided it. When a review finishes, the panel first shows up to 3 key moments, and the full move list stays one click away. The key moments are the mistakes that cost the most winning chances, largest first. At most 1 of them is a Brilliant or Great move, unless the game has too few mistakes to fill the list. Mistakes made after the game was already decided are left out.

When the game has a winner, the review also looks for the move that the loser never recovered from: the last moment the losing side still had winning chances of 40% or more, followed by a move of their own that lost more than 10 points. That move is always among the key moments, with a note that the position never came back.

What to take from a game

Above the key moments, the review sums up one player's mistakes in a single sentence. It sorts the mistakes that have a concrete reason into five kinds: a piece or pawn left to be taken, a tactic walked into, a chance missed, a mate allowed, and a win let slip. It names the kind that cost the most winning chances, lists up to 4 of those moves, and adds one habit to practise in the next game. When none of the mistakes has a concrete reason, it says so. You can switch between White and Black. When you load a game by username, the review knows which side you played: it shows your side first, turns the board your way, and picks the key moments from your own moves.

The same block checks a few opening principles, looking only at a player's first 10 moves. It reports at most 2 of these findings: the queen came out in the first 6 moves while 2 or more knights and bishops had not moved; a piece that had already moved, and was not under attack, was moved again 2 or more times; the king moved and gave up castling, or had not castled after 10 moves; 2 or more knights and bishops had not moved after 10 moves; 6 or more of the 10 moves were pawn moves. These are facts about the moves, not engine verdicts. Games shorter than 6 moves and games that start from a set-up position are not checked.

When a note uses a term such as fork, pin, skewer, discovered attack or back-rank mate, a line under it says what the term means.

How the explanations are written

When a move is graded as an inaccuracy, a mistake, a miss or a blunder, the review adds a short note in plain English. The notes come from fixed rules. Each rule looks for one familiar pattern, and it only speaks when two things agree: the pattern is on the board, and the engine's own line shows that it costs material. The rules are checked in this order, and the first one that fits writes the note:

The notes are not written by a chatbot. They only state what the rules can confirm, so they never invent a reason. Before a note says that material is lost or won, the review follows the engine's line until the captures and checks stop, and compares the result with the line after the engine's preferred move. A sacrifice that wins the material back a few moves later is not called a loss. When no rule fits, the note stays with the facts: how the winning chances changed, the engine's preferred move and the opponent's strongest reply. That happens often, because the rules only know the common tactical patterns. They do not explain positional mistakes, such as a weak pawn structure or a badly placed piece.

Brilliant and Great moves get a line that says what was offered and where the next-best move would have left you. Book moves name the opening. Moves graded Best, Excellent or Good get a short line with the change in winning chances and the engine's preferred move when it was different.

Limits worth knowing

An engine is a guide, not a judge. In a sharp position it can change its mind as it searches deeper, and a deeper review can grade a few moves differently from a quick one. Some moves that a human would rightly call good, like a practical choice when short on time, can still grade as inaccuracies because the review only looks at the board. Use the grades to find the moments worth studying, then look at them yourself on the analysis board.

Privacy in one paragraph

The engine runs in your browser, your games are never uploaded, and there is no account. When you load games by username, your browser fetches them from Chess.com or Lichess directly. My Chess Review is not affiliated with Chess.com, Lichess or the Stockfish project. See the privacy page for everything else.

Review a chess game