Game review features compared: how platforms define a blunder

"Blunder" sounds like a universal verdict, but every platform computes it differently. This guide sets out how Lichess and Chess.com actually define their move labels, using Lichess's open-source code and Chess.com's help center, so you can read a review without over- or under-reacting. ChessHelper (chesshelper.ai), our own tool, takes a different approach to feedback, with an evaluation bar, alternative lines and engine strength matched to your level, and we cover it at the end.

Why the labels differ

An engine gives an evaluation in pawns. A label such as "mistake" needs a rule that turns a change in evaluation into a verdict. A raw pawn drop is a poor rule: losing 1.00 when the position is level is serious, while losing 1.00 when you are already up a rook changes little. Both Lichess and Chess.com therefore convert the evaluation into a measure of winning chances first, and label the size of the drop. They use different curves and different cut-offs, so the same move can get different labels on the two sites.

How Lichess defines inaccuracy, mistake and blunder

Lichess converts centipawns into winning chances on a scale from −1 to +1 with a logistic curve whose constant, −0.00368208, is set in its code. It then measures how much the move reduced the mover's winning chances:

Forced mates have their own rules. A move that allows a forced mate is a blunder, unless the mover was already clearly lost: if the evaluation was already worse than −7 pawns it is graded a mistake, and worse than −10 pawns an inaccuracy. Missing a forced mate you had is graded the same way in reverse: if you are still more than 10 pawns ahead after the slip it is an inaccuracy, more than 7 pawns a mistake, and otherwise a blunder. Lichess shows the label with the move the engine preferred, for example "Mistake. Nf3 was best."

Lichess's accuracy percentage is built from the same win percentages. Each move gets a score from a fitted curve: a move that keeps your win percentage gets 100, and the score falls off exponentially as the drop grows. The per-move scores are then averaged into the game figure.

How Chess.com classifies moves

Chess.com's help center describes an "expected points" model: the share of a point you would expect to score from the position, from 0 to 1. Each move is classified by how many expected points it lost:

On top of that scale sit rule-based labels without published numbers. A Great move is one that was critical to the outcome, and Chess.com says newer players are labeled more generously. A Brilliant move is a good piece sacrifice that does not leave you worse and that you did not need because you were already completely winning. A Miss is a failure to capitalize on your opponent's mistake, with thresholds that vary by rating.

Chess.com's accuracy score uses a separate algorithm, CAPS2. The company describes it as working like a grade on a school test, with most scores falling roughly between 50 and 95, and it has not published the formula.

The two scales side by side

Lichess's thresholds are in win percentage (0 to 100), and Chess.com's expected points run from 0 to 1, so 0.05 expected points corresponds to 5 percentage points if both curves agreed. They do not: the curves converting engine scores into chances are different, and Chess.com's is not published. Treat the comparison below as a reading guide, not an exact conversion.

The visible difference is at the top: Chess.com's blunder line sits higher on its own scale, and Chess.com adds praise labels (Best, Excellent, Good, Great, Brilliant) that Lichess does not use.

Worked example: a blunder by any definition

87654321

abcdefgh

After 1.e4 e5 2.Bc4 Nc6 3.Qh5, Stockfish at depth 20 rates 3...g6 best at −0.38, slightly better for Black. The natural developing move 3...Nf6 allows 4.Qxf7#. Under Lichess's rules this is a move that allows a forced mate from a position where Black was not lost, so it is a blunder. On any expected-points scale it throws away the whole game, which puts it at the top of Chess.com's blunder band.

Worked example: where labels can disagree

Closer calls are where platforms split. In the Two Knights Defense after 1.e4 e5 2.Nf3 Nc6 3.Bc4 Nf6 4.Ng5 d5 5.exd5, the recapture 5...Nxd5 moves the evaluation from −0.09 to +2.11 at depth 20 in our checks. Plugged into Lichess's formula, Black's win percentage drops from about 51 to about 31, around 19 points, which is a blunder on Lichess. A shallower search, or a different curve, could put the same move in the mistake band. The full walk-through of this position is in our guide on how to analyze your chess games.

Other tools: explanations instead of labels

Not every tool leads with labels. DecodeChess translates the engine's view into threats, plans and concepts in plain language, with one decoded game a day on its free plan. En Croissant's reports show an evaluation graph, a board heatmap and the best moves. ChessHelper shows the evaluation bar, arrows and alternative lines on your Chess.com and Lichess games, with engine strength you can set from 800 to 3200 Elo, and stores your games with accuracy and rating trends on its dashboard. The feature matrix lines them up.

How to use labels well

Use labels to find where to look, not as a grade. Start with the largest drop, which is usually the one that decided the game. Ignore a string of inaccuracies in a position that was already lost. And compare like with like: an 85% accuracy on one site is not the same as 85% on another, because the formulas differ. The controls that sit behind these labels are explained in our analysis board guide.