Why your puzzle rating is 600 points above your game rating

The short answer

A puzzle tells you a tactic exists. Your game does not. That single difference accounts for most of the gap, and it is why a 2100 puzzle rating next to a 1200 blitz rating is ordinary rather than alarming.

The two ratings also come from different pools. Your puzzle rating is calibrated against other solvers on that site, your game rating against other players on that site, and nobody has tied the two scales together. Comparing them is like comparing your 5k time to your bench press. The number that matters is not the gap itself but which skill the gap points at.

What each number is actually measuring

A puzzle hands you three things for free: it is your move, something decisive is available, and the position has already been filtered so that one line works and the rest fail. You are doing search, not detection.

In a game you get none of that. Most positions have no tactic. You have to decide, forty times, whether this is one of the rare positions where you should stop and calculate — and you have to do it while the clock runs and while you are also tracking your own hanging pieces.

So a high puzzle rating tells you your calculation engine works when it is switched on. It says nothing about whether you switch it on. Those are separate skills and they improve from separate training.

Your errors are not evenly spread, and that is the useful part

Across 243,265 engine-classified moves from 7,709 games analysed on Mated, 3.8% of moves are outright blunders and another 4.2% are mistakes. That is roughly one real error every 13 moves, before you count the 18.4% classified as inaccuracies. Most players assume they are losing games to a slow accumulation of slightly worse moves. They are not.

The same set shows where the errors live. In the middlegame, 9.0% of moves are a mistake or worse. In the endgame, 2.3%. Phase, not opening knowledge, is where the damage happens.

And the damage is lopsided. The average move in that set gives away 93 centipawns against the engine's best, but the median is far lower, because a small number of very bad moves carry almost all of it. Your typical move is fine. You lose games to two or three moves, and those moves cluster in the middlegame.

This is what a large puzzle-over-game gap usually means in practice. You can solve the tactic. You are not noticing the two or three middlegame positions per game where one exists — for you or against you.

Gap in the usual direction: puzzles far above games

If your puzzle rating is hundreds of points above your game rating, stop grinding puzzle streaks. You have already demonstrated the skill they train. Train detection instead.

Concrete version. You are in a quiet position, you have a plan, you play Bd3 to complete development. Three moves later your opponent's knight lands on f4 and you are losing a pawn and the bishop pair. The engine says Bd3 was a blunder. You would have found the refutation in eight seconds if someone had shown you the position with a caption saying White to play and win. Nobody did.

Things that build detection, in rough order of how much they cost you:

• Before every move, name every undefended piece on the board — yours first. Out loud if you are alone.
• When a piece lands on a new square, ask what it now attacks. Not what it defends. What it attacks.
• Play slower time controls. Detection failures at 3+0 tell you nothing you can fix.
• Replay your own losses with the engine off first. Find the losing move yourself, then check. If you cannot find it unaided, that is the skill that is missing, not calculation.
• Take one lost game a week and find the last position where you were still equal. Look at what changed in the move after.

Gap in the other direction: games above puzzles

This is rarer and it usually means one of two things. Either you never do puzzles, in which case the number is stale and meaningless, or you play solid, low-risk chess and win on your opponent's errors rather than your own tactics.

The second version has a ceiling. Somewhere around 1600 to 1800, depending on your pool, opponents stop handing you material and you have to generate something yourself. If that describes you, puzzle training is genuinely the highest-value thing you can do, and calculation drills without moving the pieces matter more than pattern flashcards.

One caveat, and it is an opinion rather than a measurement: a puzzle rating built entirely on rapid-fire puzzle modes with a clock is not the same skill as calculating a six-move line. If you want the second, solve slowly and write the line down before you play the first move.

A puzzle set that you always solve is not training you

Across 2,174 rated puzzle attempts on Mated, 94.4% were solved first time. That is a pleasant number and a bad one. If you are getting nineteen out of twenty, the set is pitched as a warm-up, not as training, and your puzzle rating will drift up without your play improving.

Nobody has measured the ideal failure rate for chess puzzles well, so treat any specific percentage you read with suspicion, including one we might give you. But the direction is clear enough: if you almost never fail, the material is below you, and the rating you are collecting is mostly measuring how fast you recognise things you already know.

The fix is not complicated. Push the difficulty until you are failing often enough to be slightly annoyed. Slow down on the ones you fail and work out whether you missed the idea or miscalculated a line you had already found. Those two failures need different responses.

What the gap cannot tell you

Neither number distinguishes between the mistakes you make with twenty minutes on the clock and the ones you make with twenty seconds. That distinction is usually bigger than the puzzle-versus-game gap, and both ratings hide it completely.

A rating is one number summarising thousands of decisions, so it throws away everything about which decisions went wrong. If you want to know whether your problem is detection, calculation, time management or endgame technique, the ratings will not tell you — you need the positions. Running an engine over your own games and sorting the errors by phase and by clock time answers it in an afternoon; that split by phase is exactly what Mated builds its daily session from, but you can do a rough version yourself with any engine and a spreadsheet.

One more thing the gap hides: your opponents' blunders. If one move in 13 is a real error in that 243,265-move sample, the same is true across the board from you. Half of tactical training is noticing the mistakes that have already been made.