How to read Stockfish analysis: evaluations, depth, lines and blunders

To read Stockfish analysis, look at three things: the evaluation (a number such as +1.5 or −3.2, where plus means White is better and the size is roughly the advantage in pawns, or "M3" for a forced mate), the engine line (the moves Stockfish expects next) and the change in evaluation caused by your move. A move gets called a blunder when it throws away a big slice of your winning chances. In one of my games, a single bishop move took the evaluation from +0.31 to −4.22.

This guide shows you how to read each part of an engine's output, then walks through three real positions from my own Chess.com games (I'm nacsol3, an adult improver who went from 631 to 1385 in rapid). You'll see a classic blunder, a "blunder" that still left me better, and a huge evaluation drop that got no label at all. By the end you should be able to answer the question most beginners actually have: why did Stockfish call my move a blunder?

How do I read Stockfish evaluation numbers without guessing?

Start by checking the display's point of view. The examples on this page are converted to White's perspective: positive favours White and negative favours Black. Raw engine output can instead be relative to the side to move; a user interface may convert it. Do not assume that turning the board around changes the sign convention. Check the interface or a position with an obvious material imbalance before comparing two tools.

Next identify the unit. A centipawn is one hundredth of the traditional pawn unit, so a score of 150 centipawns is displayed as 1.50 on a pawn scale. Modern Stockfish normalizes its evaluation; it is not an inventory of captured pieces. Its official FAQ, checked October 31, 2026, explains both the scale and its calibration using engine self-play. Those self-play probabilities are not a forecast of your next game against another beginner.

For the concrete examples below, read in this order: which side the score favours, whether a forced mate is shown, then the suggested reply. Only after that compare the before-and-after numbers. If the second decimal changes when you let the engine search longer, keep the explanation provisional. A verified capture or mating line is a more useful lesson than trying to memorize an unsettled score.

Step 1: read the evaluation number

Stockfish gives every position a score. Its own FAQ explains that the score is measured in pawns or centipawns, where one pawn equals 100 centipawns (Stockfish FAQ, checked October 26, 2026). Analysis boards usually show pawns with two decimals, so you'll see +0.31 rather than 31 centipawns.

Three things beginners often miss:

• Check the sign convention, rather than reading it as your side. These examples and Chess Coach use White's perspective: + is good for White and − is good for Black. If you played Black, a falling number is good news.
• It isn't a material count. The number includes things like king safety and active pieces. Being "+1.5" doesn't mean you're a pawn and a half up on the board; it means your position is worth about that much.
• It's tied to winning chances. Stockfish's FAQ says its normalized scale links an advantage of 1.0 to about a 50% chance of winning. That's why +0.30 is "fine" and +3.00 is usually decisive at club level.

A mate score is different from a big number. "M4" means the engine has found a forced checkmate, not a very large advantage. If you see M-something after your opponent's move and a normal number after yours, you missed a forced mate.

Step 2: understand depth

Depth is how far the engine searched. The Stockfish FAQ defines it as the number of half-moves (plies) the search goes through, so depth 18 is roughly nine moves for each side. It also warns that depth is a guide, not a guarantee: some lines get looked at much deeper and others much shallower.

What this means for you:

• Numbers change a little with depth and time. In my game analyses, Stockfish searches up to depth 18. When I ran the first position below a second time for this guide, the best move was the same but the numbers moved: +0.23 instead of +0.31, −4.56 instead of −4.22. The verdict didn't change.
• Trust the size of a swing more than the second decimal. +0.31 versus +0.23 is noise. +0.31 versus −4.22 is a lost piece.
• Sharp positions need more depth. If a verdict surprises you, let the engine run longer. If the number keeps jumping around as the depth climbs, the position is complicated, and that's useful information too.

Step 3: read the engine lines

Next to the number, Stockfish shows a line (also called the principal variation): the sequence of moves it expects if both sides play well. Many boards show the top two or three lines, each with its own evaluation.

• The first move of the top line is the engine's best move. The rest of the line is the reason it's best.
• Read it one move at a time. For each move ask: what does it attack, what does it defend, what does it leave unprotected? The idea usually shows up within the first three or four moves.
• Compare the lines' numbers. If the top three moves are all within a few tenths of a pawn, you had several good choices. If the second line is already much worse, the position demanded one precise move.
• Don't memorize long lines. After four or five moves an engine line is just one possible future. Take the idea, not the sequence.

Step 4: read the drop, not just the label

Here's the key to "why did Stockfish call my move a blunder?". Stockfish itself only produces evaluations. The words inaccuracy, mistake and blunder come from the site that runs it, and they're based on how much your move changed things.

To read a drop:

1. Take the evaluation before your move. That's what the position was worth with the best move.
2. Take the evaluation after your move.
3. Turn both into winning chances and look at the difference.

Step 3 matters because a pawn isn't worth the same everywhere. Lichess publishes the formula it uses to turn centipawns into a win percentage (Lichess accuracy page, checked October 26, 2026). With that formula, +0.31 is about 53% and −4.22 is about 17.5%. But +10.42 and +17.66 are both above 97%: once you're completely winning, losing a few more pawns of advantage barely changes the outcome.

Lichess's open-source code then labels a move by how much winning chance it loses: 5 percentage points or more is an inaccuracy, 10 or more a mistake, 15 or more a blunder (lila source, Advice.scala, checked October 26, 2026; the code expresses these as 0.1, 0.2 and 0.3 on a −1 to +1 scale). Chess Coach uses the same thresholds.

Chess.com uses a different model. Its Help Center says moves are classified by Expected Points, which estimate a player's winning chances from their rating and the engine evaluation; a move that loses 0.20 expected points or more is a Blunder, 0.10–0.20 a Mistake and 0.05–0.10 an Inaccuracy (How are moves classified?, Chess.com Help Center, updated February 9, 2026). Because ratings are part of the model, the same move can get a different label on the two sites. If you want the categories themselves explained, see inaccuracy vs mistake vs blunder. This guide is about reading the numbers behind them.

Example 1: the bishop sacrifice that wasn't (a classic blunder)

March 7, 2024. I'm White, rated 1045, against soytueternokarma (1080) in a 10-minute game (game on Chess.com). Black has just pushed a pawn to b5, attacking my queen on a4.

r . . q k b . r
. . p . . p p .
p . n . p n . p
. p . p . b . .
Q . . P . B . .
. . P . P N . .
P P . N B P P P
R . . . K . . R

What I played (red arrow): my bishop took the pawn on b5 (Bxb5). It looks active. The bishop attacks the knight on c6 and pins it: the knight can't move without exposing its king.

What Stockfish said: before my move the evaluation was +0.31, about equal. After it, −4.22. In winning chances that's roughly 53% down to 17.5%, a drop of about 35 points, more than twice the 15-point blunder line.

How to read the line: the engine expects the pawn on a6 to take my bishop, my queen to take back on b5, and then Black's queen to step to d7, defending the pinned knight. Count the material: I gave a bishop (worth about three pawns) for two pawns, and the pin goes nowhere because the knight is now defended. That's the whole story behind −4.22.

The better move (green arrow): the queen simply goes back to d1 (Qd1). Stockfish's line continues with Black developing a bishop and both sides castling, about equal.

The lesson in plain words: a pin only wins something if you can attack the pinned piece more times than it's defended. Before you give up a piece to create one, count the defenders.

What happened next? On the very next move my opponent returned the favour: instead of defending with the queen, Black moved the other knight, and the evaluation swung from −4.52 back to +2.07. I won the game. Club chess is like that, and it's one reason the week 10 diary of my 2024 games is full of these swings.

Example 2: why did Stockfish call my move a blunder when I was still winning?

June 21, 2025. I'm White, rated 1270, against Peoples_champ_NYC (1272) (game on Chess.com). The evaluation before my move is M1: I have checkmate in one.

. n . . . . k .
. . q . . . r .
. . . . p . Q .
p . . . . . . .
P p . P . . . .
. P r . P N b .
. . . . . P P .
R . . . K . . R

What I played (red arrow): my queen took the pawn on e6 with check (Qxe6+). After it the evaluation was +2.43, still clearly better for me.

What I missed (green arrow): the queen to e8 is checkmate. The black king can't go to f7 or f8 because the queen covers them, and my rook on h1 guards the open h-file, so h7 and h8 are covered too.

Why it's still a "blunder": a forced mate counts as a certain win, at the top of the scale (97.5% in my analysis). +2.43 is about 71%. The drop is about 26 points, so it's labelled a blunder even though I was still better afterwards. The label doesn't say "you're losing now". It says "you gave away a lot of what you had".

And here the label was right to shout. The game went on for another 33 moves and ended in a draw. You can see the rest of that week in my week 25 diary of 2025.

How to spot this in your own reviews: whenever the evaluation before your move is an M-number and the evaluation after it isn't, stop and look for the mate. Checks, captures and threats, in that order.

Example 3: a seven-pawn drop with no label at all

February 27, 2025. I'm White, rated 1265, against busao (1253) (game on Chess.com). I'm completely winning on move 36. My queen takes a pawn on a5, and my game analysis has the evaluation going from +17.66 to +10.42. Stockfish preferred a rook check on c7 (when I re-ran the position for this guide, it gave that check +16.13).

More than seven pawns of advantage gone, and no inaccuracy, mistake or blunder mark. Why?

Because both numbers mean "White wins". On Lichess's formula, +17.66 is about 99.9% and +10.42 about 97.9%, a drop of about two points, far below the five-point line for an inaccuracy. Chess Coach caps the scale at +10, so both count as 97.5% and the drop is zero. I won the game a few moves later.

The lesson: a big number swing in a won position usually just means "there was a faster way". It's worth a glance, not a panic. The same swing at +0.3 would have decided the game.

Common mistakes when reading Stockfish analysis

• Reading the sign from your own side. If you're Black, −2 is good for you.
• Treating the evaluation as material. +1.5 can come from a strong attack with equal material.
• Studying the label instead of the line. "Blunder" tells you that something went wrong; the first two or three moves of the engine line tell you what.
• Chasing decimals. +0.4 versus +0.2 rarely teaches anything at club level. Look for drops that cross the blunder line, missed mates and moments where the evaluation changed sign.
• Ignoring your opponent's mistakes. The move after a blunder is often a test you also fail. In Example 1 my opponent missed a simple defence one move later.
• Assuming the engine's move is the human move. Lichess's accuracy page itself notes that the engine judges whether a move is sound, not how hard it was to find.
• Asking a chatbot to evaluate without an engine. Language models guess numbers. We compare the two in ChatGPT vs Stockfish.

How Chess Coach helps you read the analysis

Chess Coach is a connector for Claude and ChatGPT that reviews your finished Chess.com and Lichess games with Stockfish running on our server. It picks the two or three key moments, shows them on a board with a red arrow for your move and a green arrow for the better one, and draws the evaluation graph. Every number comes from the engine; your assistant turns the line into plain words, and you can press "Why?" or ask a follow-up such as "why is −4.2 so bad here?" in the same chat. The free plan includes one in-depth review a day. See the AI chess analysis features and how to connect it.

If Chess.com's review gave you a label without a reason, why Chess.com Game Review doesn't explain moves covers what each plan shows. And if you'd rather use ChatGPT with your own engine output, the ChatGPT chess coach prompts include one for explaining an engine line.

Chess Coach is independent and not affiliated with Chess.com, Lichess, Anthropic or OpenAI.