Open any post-game report on Chess.com or Lichess and two numbers stare back at you: an accuracy percentage and an average centipawn loss (ACPL). Most improving players treat them like a school grade — 92% accuracy feels like an A, an ACPL of 45 feels like a C-minus, and the instinct is to chase the higher number game after game. That instinct quietly stalls more adult improvers than almost any opening they could pick badly.
After auditing hundreds of game reports from club-level players, the pattern is consistent: the metric isn’t lying, but it isn’t measuring what you think it is. This guide breaks down what each number actually represents, why two players with identical accuracy can be on completely different improvement trajectories, and how to read these metrics the way a coach does instead of the way a leaderboard does.
What Average Centipawn Loss Actually Measures
A centipawn is one hundredth of a pawn — the engine’s base unit for measuring an advantage. If a position is evaluated at +1.00, the engine thinks one side is up the equivalent of a clean pawn. Average centipawn loss is the average evaluation drop, in centipawns, across all of your moves in a game, measured against the move the engine considered best.
So an ACPL of 30 means that, on average, each of your moves gave back about a third of a pawn compared to the engine’s top choice. Crucially, this is an average of regret, not a count of mistakes. That distinction is the entire reason the metric gets misread.
Why averages hide the story
Imagine two games, both finishing at an ACPL of 35:
- Game A: 39 near-perfect moves and one catastrophic blunder that hung a piece (a 700-centipawn swing on move 24).
- Game B: 40 slightly imprecise moves, each leaking 30–40 centipawns, with no single disaster.
The number is identical. The diagnosis is opposite. Game A is a calculation-and-blunder-check problem — your play is fundamentally sound but one lapse cost the game. Game B is a positional-understanding problem — you don’t actually know where the pieces belong, and the engine quietly disagrees with you a dozen times a game. Training advice that fixes one will do nothing for the other. This is exactly why an aggregate score, taken alone, can send you to the wrong drills for months.
Accuracy Is a Repackaged Version of the Same Number
The accuracy percentage you see is not an independent measurement. Both Chess.com and Lichess derive accuracy from the same per-move evaluation losses that feed ACPL, run through a conversion formula that maps centipawn loss onto a 0–100 scale. Lichess publishes its formula openly; it converts the “win percentage” lost on each move into an accuracy figure and averages it. Chess.com uses its own proprietary curve.
Two consequences follow, and both matter for how much you should trust the number:
1. Accuracy is non-linear, so it flatters won games
Because the conversion is based on win probability rather than raw evaluation, a mistake made in a position that’s already winning barely dents your accuracy — the win probability hardly moves whether you’re up +6 or +5. The same size error in a balanced position craters it. This is deliberate and reasonable, but it means a 95% accuracy game where you were winning from move 10 tells you far less than an 80% accuracy game fought in a sharp, equal middlegame. High accuracy in easy positions is not evidence of strong play.
2. The numbers are not comparable across platforms
An 85% on Lichess and an 85% on Chess.com are not the same achievement, because the curves differ and the engines, depths, and even the definition of an “inaccuracy” differ. Comparing your accuracy to a friend’s on a different site, or to a streamer’s, is meaningless. The only valid comparison is you against yourself, on one platform, over time.
How a Coach Reads These Numbers (The Three-Layer Method)
Instead of asking “was my number good?”, strong players interrogate the report in three layers. This is the framework we build the personalized plans around, and you can run it manually on any analyzed game.
Layer 1: The shape of the loss, not the size
Open the move list and look at where the centipawns leaked. Is it one cliff (a blunder) or a gentle slope (chronic imprecision)? Sort your errors by magnitude. If 80% of your total centipawn loss comes from one or two moves, you have a vigilance problem, not a knowledge problem — and no amount of positional study will fix a hung queen. If the loss is spread evenly across fifteen quiet moves, the reverse is true.
Layer 2: The phase where the loss happens
Tag each significant error by phase: opening, middlegame, or endgame. A player with an ACPL of 25 whose entire loss is concentrated in the first ten moves has an opening-preparation gap that’s trivially fixable. The same ACPL spread across rook endgames points somewhere completely different — toward technical endgame study like the Lucena and Philidor positions every tournament player needs. The aggregate number is identical; the curriculum is not.
Layer 3: The position type, not the move
This is where the engine’s own blind spots matter. Modern engines like Stockfish and Leela evaluate positions differently from each other, and both can flag a perfectly sound human move as an “inaccuracy” simply because it isn’t the sharpest computer line. We covered this in depth in our look at the turning points engine reviews routinely miss. When the report dings you for a “mistake” that kept the position simple and within your skill, that’s often a better practical decision than the engine’s preferred razor’s edge. Reading these metrics well means knowing when to overrule the machine — something we explore further when comparing Stockfish and Leela as training partners.
The Trap: Why Chasing Accuracy Stalls Improvement
Here’s the uncomfortable mechanism. Accuracy and ACPL reward not losing centipawns, and the easiest way to not lose centipawns is to avoid complexity. A player optimizing for the number learns, often unconsciously, to trade pieces early, steer toward dry positions, and decline sharp lines — because simple positions have fewer ways to drop evaluation. Their reports get prettier and their real strength flatlines, because strength is built precisely in the messy positions they’re now avoiding.
This is the metric equivalent of studying for the test instead of learning the subject. The improvers who break through plateaus are usually the ones whose accuracy gets worse for a stretch — they’re deliberately playing sharper, less familiar structures, accepting more errors now to build patterns that pay off later. If your accuracy has been climbing while your rating sits still, that mismatch is the single most useful signal in your entire report. It almost always means you’ve optimized your style around the scoreboard.
A Practical Reading Routine for Your Next Game
Run this every time, and the numbers become a coach instead of a grade:
- Ignore the headline number for the first 60 seconds. Go straight to your two largest single-move losses and understand them before you ever look at the average.
- Ask “blunder or drift?” One cliff means train vigilance and calculation; a slope means train positional understanding and plans.
- Tag the phase. Opening losses are the cheapest to fix; endgame losses are the highest-leverage to fix.
- Check the difficulty discount. Was your high accuracy earned in a sharp equal game, or handed to you in a position that was already winning?
- Watch the trend, never the single game. Ten games on one platform tell a story; one game tells you almost nothing.
Done consistently, this turns a vanity metric into a diagnostic. The number stops being something to chase and becomes a map of exactly which skill to train next — which, conveniently, is the entire job of a training plan.
Frequently Asked Questions
What is a good average centipawn loss for a club player?
As a rough guide, players around 1200–1500 often post an ACPL in the 50–80 range, 1500–1800 lands roughly 35–55, and 1800–2100 tends toward 20–40 — but these are descriptive, not targets. ACPL drifts with the sharpness of your games and the engine depth used, so a “good” number in a wild tactical game differs from one in a quiet positional grind. Track your own trend rather than aiming at a universal threshold.
Is accuracy the same on Chess.com and Lichess?
No. Both derive accuracy from per-move evaluation loss, but they use different conversion formulas, different engines, and different analysis depths. An 85% on one platform does not equal 85% on the other. Only compare your accuracy against your own past games on the same site.
Can a high-accuracy game still be a bad game?
Yes, frequently. Because accuracy is weighted by win probability, errors made while already winning barely register. A 95% accuracy score earned in a position you were dominating from the opening reveals far less about your skill than a hard-fought 80% in a balanced, complex middlegame.
Should I always trust the engine when it flags a “mistake”?
Not blindly. Engines reward the objectively sharpest move, which is sometimes a line no human should enter over the board. A move that keeps the position simple and within your skill can be the stronger practical choice even when the engine marks it down. Reading these reports well means knowing when your decision was better for a human than the computer’s.
Stop Grading Your Games — Start Diagnosing Them
Accuracy and centipawn loss are genuinely useful, but only once you stop reading them as a report card and start reading them as a symptom list. The number tells you something happened; the three-layer method tells you what and what to do about it.
That diagnostic step is exactly what MyChessPlan automates. Get your free chess archetype report to see which type of player you are and where your centipawns are really leaking — or go straight to the $14.99 personalized improvement plan, which reads your actual games, separates your blunders from your drift, and hands you the specific drills for your weakest phase instead of a number to chase.
Leave a Reply