How we measure which openings punish mistakes
Every opening in the trainer carries one number per rating band: of the opponents who reach it, the share who make a mistake on the lines we teach. This page is the whole method, with the exact figures the current build uses. Nothing here is typed by hand; it is read from the same index the app ranks with, and a test fails the build if the two ever disagree.
What a finished claim looks like
- Vienna Gambit: 62% of Lichess players rated 1600–1799 who reach the Vienna Gambit make a mistake or blunder on the lines we teach within the first 10 moves (1.9M games, a lower bound).
- Evans Gambit: 3% of Lichess players rated 1600–1799 who reach the Evans Gambit make a mistake or blunder on the lines we teach within the first 10 moves (1.5M games, a lower bound).
- Danish Gambit: 5% of Lichess players rated 1600–1799 who reach the Danish Gambit make a mistake or blunder on the lines we teach within the first 3 moves (5.1M games, a lower bound).
1. Where the games come from
The Lichess opening explorer, blitz, rapid, classical games, split into seven rating bands (<1200 · 1200s · 1400s · 1600s · 1800s · 2000s · 2200+) plus an all-levels total. We use per-move aggregates only: how many games in a band reached a position and what was played there. A band prints a number only when at least 1.0Kgames in that band reached the opening's root position.
2. How a move is graded
Each popular reply is evaluated with Stockfish at depth 20and compared with the engine's best move in the same position, in win-probability points, from the mover's point of view. The label comes from the classifier the game review uses (classifyMove): a drop above 5.75 points is a Mistake, above 20 a Blunder. Only Mistake (?) and Blunder (??) count. Inaccuracies are deliberately left out: that band is narrower than engine noise between depths.
3. How it adds up
Every opening has a declared root: the position after which it is on the board, with the opponent to move. From there we walk the taught lines as a tree. At each position the share of games that played a counted mistake is added, weighted by the share of games that followed our line to that position. Where the opponent has several taught replies their branches add, because they are different games. Where we have two answers to the same reply, only one branch is followed, the one a learner would study, so the total can never double-count. Games that leave our lines with a good move drop out and never return: the figure is a lower bound, and each row also records how much of the reached traffic it graded. The bars next to each opening show the slip rate move by move among opponents still on the line; the horizon is the mined depth, 11 moves.
4. What we verify at depth 30
Depth-20 evaluations move by a few win-probability points between depths, so a reply graded 6.15points is not a reliable Mistake. Any counted reply that moves an opening's headline by at least 0.5 of a point and sits within 3 points of a bar is re-graded from full depth-30 searches of both positions, and the deeper verdict is final.
276 headline-moving replies across 70 openings · 86 close calls · 86 verified at depth 30 — every close call verified.
5. What we do not claim
- Not a win rate, and not a measure of how good an opening is: it measures how often opponents slip early against these lines.
- Not a share of players: a share of games in a band that reached the position.
- Not exact: a lower bound, because ungraded and rare replies never count.
- Not the whole game: the horizon is the first 11 moves of the taught lines.
Frequently asked questions
Of the Lichess games in that rating band that reached the opening's root position, at least 22% saw the opponent play a move our engine grading labels a Mistake (?) or Blunder (??) somewhere on the lines we teach, within the first 11 moves. It is a share of games that reached the position, never a share of all players, and it is a lower bound.
Only moves we have listed and graded can count. A rare reply that never made the table, or a popular one the engine pass has not graded yet, contributes nothing — even if it is a mistake. We publish how much of the reached traffic each figure has actually seen (the coverage), and hide the number when that falls under 80%.
A win rate blends opening quality with everything that happens afterwards. This measures one thing: how often opponents at a given level go wrong early against a specific set of taught lines, graded move by move against the engine. It tells you where preparation pays, not who wins in the end.
The reply data was last mined on 2026-09-02 and the index rebuilt on 2026-09-09. A freshness check in our build warns when the mine is older than four months.
The numbers live on every opening guide and inside the trainer, ranked for your own rating band.
Browse the openings