Every chess player who clicks “Computer Analysis” on Lichess or Chess.com sees the same thing: a row of green, yellow, and red dots, an accuracy percentage, and a centipawn loss number. Most players glance at it, feel either smug or defeated, and close the tab. They miss the actual point of the analysis entirely.
Engine output is not feedback. It is raw data. A coach turns that data into a diagnosis. The difference between players who improve from engine review and players who don’t is not the engine they use — it is the framework they apply to what the engine spits out. This post is that framework.
Why Raw Engine Numbers Mislead Most Players
Stockfish 16 evaluates positions with near-perfect accuracy at depth 20+. That is precisely the problem. It judges your moves against a standard no human will ever match, then condenses the verdict into a single number — centipawn loss — that hides almost everything useful about why the move was bad.
A player who loses 80 centipawns by missing a 14-move tactical sequence has made a categorically different mistake than a player who loses 80 centipawns by playing the wrong pawn break in a closed position. The engine prints the same number. The first mistake is unfixable for a 1400. The second is the single most important thing that player needs to learn this month.
This is why we have an entire post on how engine analysis differs from coaching — and why simply running games through Stockfish does not produce improvement on its own.
The Three Layers of Engine Output
Every modern chess engine report contains three layers of data. Players who improve learn to read them in a specific order, weighted by what is actionable.
Layer 1: Move Classifications (Blunders, Mistakes, Inaccuracies)
These are the colored dots. Chess.com and Lichess use slightly different thresholds, but the standard is roughly:
- Inaccuracy: 50–100 centipawns lost (a noticeable error, but the position is still playable)
- Mistake: 100–300 centipawns lost (a real positional or tactical concession)
- Blunder: 300+ centipawns lost (a game-changing error)
This is the most overrated layer of the report. Players obsess over their blunder count and ignore that where in the game the blunders happened matters far more than how many there were. Five inaccuracies in the opening phase from the same player almost always indicate a single recurring repertoire gap — not five separate problems.
Layer 2: Centipawn Loss and Accuracy Percentage
The “accuracy” score most platforms display (e.g. 87.3%) is derived from average centipawn loss per move. It is a useful comparison metric across your own games at the same time control. It is nearly worthless as a comparison against other players.
Here is the rule that actually matters: your accuracy should be roughly stable across game phases. A player whose accuracy is 92% in the opening, 76% in the middlegame, and 81% in the endgame has just diagnosed themselves. The middlegame is where their skill drops off. That is the training target — not “play fewer blunders.”
Layer 3: Evaluation Swings (The Layer Almost Nobody Reads)
This is the most important layer and the one no platform highlights well. It is the graph of how the evaluation changed throughout the game. The pattern of swings — not the individual values — tells you what kind of player you are.
Three common patterns:
- Sawtooth: Evaluation oscillates wildly between +2 and −2. Indicates poor risk assessment and impatient play. Common in attackers who push positions before they are ready.
- Cliff: Evaluation holds steady for 20+ moves, then drops sharply once. Indicates a knowledge gap (usually endgame or transition into a specific structure). Common in well-prepared defenders.
- Slow leak: Evaluation declines by 30–50 centipawns every few moves with no single bad move. Indicates strategic drift — the player does not have a plan. Most common pattern at 1200–1600.
This is the diagnostic information a coach extracts in five seconds and most players never see.
A Coach’s Three-Question Framework
When a strong coach reviews an engine report, they ask three questions in order. You should ask the same three.
Question 1: Where Does My Accuracy Drop?
Open the move-by-move centipawn loss graph. Identify the phase (opening, early middlegame, late middlegame, endgame) where your accuracy is consistently lowest across your last 10 games. That is your training target for the next month. Not the blunder in move 34 of last night’s game.
Question 2: Are My Mistakes Tactical or Strategic?
Look at the engine’s recommended move in each flagged position. If the engine’s suggestion is a forcing sequence (a capture, check, or threat that wins material), your error was tactical — you missed calculation. If the engine’s suggestion is a quiet positional move (a pawn break, piece reroute, or prophylactic move), your error was strategic — you misread the position.
This single distinction determines your entire study plan. Tactical mistakes are fixed by puzzle work. Strategic mistakes are fixed by studying annotated master games in similar structures. Our framework on calculation training covers the first case in depth.
Question 3: Is This Move a Pattern or a One-Off?
A single blunder is noise. The same type of mistake across three games is signal. Before you “fix” anything, check whether the same kind of position has tripped you up before. The engine cannot do this for you. You do it manually by scanning your last 5–10 game reports for the same diagnostic flag in Question 2.
Most rating plateaus are caused by a single recurring weakness that the player never identified as a pattern because they reviewed each game in isolation. Our diagnostic method post walks through how to maintain this pattern log.
Three Common Misreads That Waste Your Study Time
Even with the framework above, players consistently misuse engine output in three ways.
Misread 1: Treating “Best Move” as the Lesson
The engine’s top move is often a computer move — a line that requires 8 moves of perfect calculation that you will never reproduce. Don’t memorize it. Instead, look at the engine’s second and third choices. Those are usually the moves a human coach would have recommended, and they teach the underlying idea without requiring engine-level calculation.
Misread 2: Trusting the Opening Evaluation
Engines evaluate opening positions based on a long-horizon search that does not reflect practical playability. A line evaluated at −0.3 may be the most testing line for your opponent. A line evaluated at +0.2 may be a dry equality you cannot win. Use a database (Lichess opening explorer) for opening decisions, not raw engine evaluations.
Misread 3: Reviewing Won Games Less Carefully Than Lost Ones
This is the single most common mistake at 1500–1800. Players review their losses obsessively and skim their wins. But the engine often reveals that a “won” game was actually lost on move 18 — the opponent simply blundered later. Reviewing wins is how you find your real weaknesses before your rating starts to reflect them.
How This Connects to Your Playing Style
The patterns above are not random — they correlate strongly with playing style. Attackers consistently show sawtooth evaluation graphs. Defenders show cliffs. Strategists show slow leaks. Tacticians show clean accuracy with occasional huge swings on missed combinations.
This is why a generic “review your games with Stockfish” recommendation produces such inconsistent results. The same data means different things depending on what kind of player is generating it. If you have not yet identified your archetype, our chess archetypes guide is the place to start — it determines which engine patterns are diagnostic for you and which are just noise.
From Diagnosis to Plan
Reading engine analysis correctly gets you a diagnosis. Turning that diagnosis into a training plan is a separate skill. A diagnosis says “your middlegame accuracy drops 16% versus your opening.” A plan says “spend 20 minutes per day for 3 weeks on prophylactic thinking drills in IQP positions, then re-measure.”
If you want this done for you — a full diagnostic on your last 50 games, an archetype assessment, and a 30-day training plan calibrated to your specific weaknesses — that is exactly what the $14.99 MyChessPlan personalized improvement plan produces. It is the same workflow a $150-per-hour coach uses, automated against your real game data. You can also get a free archetype report first if you want to see the framework before committing.
Frequently Asked Questions
What is a good average centipawn loss for my rating?
Roughly: 1000 rated ≈ 60–80 ACPL, 1500 rated ≈ 35–50 ACPL, 2000 rated ≈ 20–30 ACPL. But comparing your ACPL across different opponents and time controls is misleading. Use your own historical ACPL as the benchmark, not other players’.
Should I use Stockfish, Leela, or Chess.com’s engine?
At depth 18+, all three give equivalent verdicts on practical mistakes below master level. Use whichever is convenient. The engine is not the limiting factor in your improvement — the interpretation is.
How many games per week should I analyze?
Two to four, deeply, beats ten games skimmed. Pattern recognition across your last 10 games matters more than depth on any single game. Block 30 minutes per analysis session and stop when you have identified one actionable pattern.
Does engine analysis still work for opening preparation?
Only when combined with a master games database. Pure engine prep produces theoretically sound lines that are practically unfamiliar. Use the engine to validate the candidate moves a strong player would consider, not to generate them.
Leave a Reply