How Many Chess Games Should You Actually Analyse?

The short answer

One game deeply if you want to understand a position type. Twenty to fifty games shallowly if you want to find out what you keep getting wrong.

These are not two intensities of the same activity. They are different tasks. Deep analysis of a single game tells you about that game. It tells you almost nothing about whether the thing you found is a pattern or a one-off. Shallow analysis across many games tells you what repeats, but it will never explain why you misplayed a rook endgame — it will only tell you that you did, eleven times.

Most improving players default to deep analysis on one game because that is what the advice says, and it is the wrong default. You have a limited number of problems that cost you most of your rating, and you cannot see them from one game.

Why the wide pass comes first

Across 197,870 engine-classified moves in 6,298 analysed games on Mated, 3.8% of moves played are outright blunders and another 4.2% are mistakes. That's roughly one move in twelve that is a real error, before you even count inaccuracies, which run at 18.5%.

Do the arithmetic on a single game. A 40-move game gives you about 40 of your own moves. At those rates you'd expect around three real errors in it. Three data points. You cannot tell from three errors whether you have a problem with hanging pieces to knight forks or whether you had a bad Tuesday.

Twenty games gives you roughly sixty errors. Now the same shape shows up four or five times and you can name it. That is the entire argument for going wide first: you need enough errors for the repeats to surface above the noise.

This is where a tool helps, because running Stockfish over 20 games by hand is dull and you will stop after four. Mated pulls your recent games from the Chess.com or Lichess API and runs the engine over every position in the browser, so the wide pass is something you look at rather than something you do.

The average move is misleading, and that's the point

In the same dataset, the average move gives away 94 centipawns against the engine's best. The median is far lower. A small number of very bad moves carry most of the damage.

That gap between mean and median is the most useful thing in the numbers. Your play is not uniformly 94 centipawns worse than Stockfish's. It is mostly fine, punctuated by disasters. Which means your accuracy percentage — the number you glance at after a game — is telling you about the wrong moves. It is dominated by the three moves that fell off a cliff.

So when you do go deep on one game, go deep on the cliffs. Not the opening novelty on move 9. Not the inaccuracy that cost 40 centipawns. The two or three moves where the evaluation moved by a queen. Those are where the thinking broke, and the thinking is the thing you are trying to fix.

Analyse the middlegame, skip most of the rest

Errors are not spread evenly through a game. In that same set of 197,870 moves, 9.2% of middlegame moves are a mistake or worse, against 2.3% in the endgame.

That is a factor of four. If you are budgeting attention, the middlegame is where it goes. This is not an argument that endgames don't matter — they matter enormously when they occur, and part of the low endgame error rate is simply that many games never reach one, and that the positions reaching an endgame are often already decided. But if you are deciding which twenty moves of a game to spend forty minutes on, the data points at the middle.

A practical consequence: when you do the wide pass, you can skip the opening entirely on the first sweep. Look at where the eval graph drops after both sides have castled and pieces are still on. That's your material.

A concrete schedule that works

Here is what I'd actually do, as opinion rather than fact — nobody has measured the optimal ratio and I'd distrust anyone who claims a number.

The wide pass is cheap and you should repeat it often. The deep pass is expensive and you should do it rarely and properly.

When you do the deep game, the rule is: cover the engine line. Look at the position, write down what you'd play now and why, and only then compare. If you scroll the engine line first you are reading, not thinking, and you will remember none of it by Thursday.

• Every two weeks: wide pass over your last 20–30 games. You are looking for one repeating error shape, not a list.
• Every week: one game deep. Pick a loss where you did not understand what happened, not a loss where you hung a rook.
• On the deep game: 30–45 minutes, engine off for the first pass, engine on only to check your conclusions.
• Never: deep analysis of a game you won comfortably. The information density is terrible.
• Always: a 60-second look at any game you lost on time or in under 20 moves. Those have one cause and you already know it.

When one game deeply is the right call

There is a case where the wide pass is the wrong tool, and it's worth naming. If you have just started playing a new opening, or you've reached a structure you don't understand — a Carlsbad, an isolated queen's pawn, a closed Ruy — then the frequency data can't help you. You have played the position four times. There is nothing to aggregate.

Then you go deep on one game, and you go deeper than the engine. You look up two master games in the same structure. You work out what the plans are for both sides. The engine will tell you move 23 was a mistake; it will not tell you that the minority attack was your plan and you abandoned it on move 18 for no reason.

That kind of work is slow and produces understanding rather than a fix. Both are worth having. They just answer different questions, and confusing them is why people spend an hour on a game and come away with nothing they can use.

What the number of games can't tell you

There's a limit to all of this that the advice usually skips. Engine analysis shows you where the evaluation moved. It cannot show you why you played the move, and the why is what changes.

Two players blunder the same knight. One didn't see the fork. One saw it, calculated wrong, and convinced himself it didn't work. One miscounted defenders. Same move, same centipawn loss, three completely different problems, and no amount of analysing more games separates them.

The only thing that does is writing down what you were thinking before you look at the engine. One line, in plain language: "I thought Rd1 was forced." Do that on ten blunders and you will find that most of them share a sentence, and that sentence is worth more than the ten games.