The Chess Analysis Board: How to Read What It Gives Back

A chess analysis board is a board with an engine attached. You paste in a game, load a position, or play moves directly, and it tells you continuously who is better and by how much, along with the moves it would prefer. Every major site has one and most of them are free, so the practical question is not where to find one. It is what to do with the number it gives you.

Where to get one, free

Lichess. Server-side Stockfish on any game you upload or play there, with no daily cap, plus accuracy percentages and average centipawn loss. There is no paid tier on lichess at all, for this or anything else. For a large number of players this is the correct final answer.

Chess.com. Its analysis board is separate from its Game Review product and, per a July 2026 chess.com forum thread, pulls newer Stockfish releases such as Stockfish 18 run client-side, while Game Review runs Stockfish 16 on chess.com's own servers. Free accounts get one full Game Review a day per chess.com's own Help Center, while the analysis board itself is more open.

The independents. WintrChess is a free, open-source browser analysis tool with move classifications. Chessigma advertises Stockfish 17 with no paywall and no daily game limit. Several of these run the engine in your browser, which they pitch as a privacy feature, and it is.

Free analysis is not a differentiator anymore, and any page telling you otherwise is selling something. What still differs between tools is what they do with the output.

How to read the evaluation

The number is in pawn units. Chess.com's own Help Center puts it plainly: +1.00 is roughly a one-pawn advantage for White, -2.50 roughly a two-and-a-half-pawn advantage for Black. A centipawn is one hundredth of that unit, so a 35-centipawn swing is about a third of a pawn.

Stockfish calibrates this further through its WDL model: an advantage of 100 centipawns corresponds to roughly a 50 percent probability of winning from that position in engine self-play. Two things follow from that sentence and both matter. The scale is anchored to engine play, not human play. And the same centipawn count means completely different things in different positions, which is not a subtlety, it is the main thing to understand.

Lichess makes this argument on its own accuracy page: it converts the Stockfish evaluation into a win percentage first and grades moves on the change in that, rather than on a fixed centipawn cutoff, and it criticises raw centipawns directly on the grounds that losing 300 centipawns means something entirely different depending on where you started. Nate Solon's chess newsletter makes the harder version of the same point: centipawn scores are not tied to any discernible chess quantity and drift across engine versions.

What depth actually promises

Depth looks like the setting that decides whether you can trust the verdict. It is softer than that. Stockfish's own FAQ states that a reported search depth is a useful guide but not a strict guarantee that every move to that ply was examined, because pruning and reductions mean the engine searches some branches far deeper than others.

Chess.com users ask what depth means on chess.com's own forums, in threads titled exactly that, which tells you how well the interfaces explain it. For reviewing your own games, the honest answer is that depth stopped being your limiting factor a long way below where the slider goes. We took this apart in detail in our piece on what chess analysis cannot tell you.

The column no analysis board prints

Here is the gap, and it is the reason a correct analysis board can still leave you no better off. Every one of these tools answers one question: what is the best move? None of them answer the question you actually had: was that move findable by someone like me?

We measured that directly. Running Maia3, a model trained to predict human moves at a chosen rating rather than to find the best one, against the finish of the Opera Game with Stockfish 17 at depth 20 as the reference, we get what we call findability: the probability a player at a given rating plays the exact move the engine rates best. The full dataset was generated on 28 August 2026.

The first column behaves the way you would expect: stronger players find the engine's move far more often. The second column runs backwards, and that is the interesting one. The higher the rating, the less likely the player is to play the engine's recommended knight recapture, because 13...Rxd7 is the natural human recapture and stronger players are more consistent about playing natural moves. At 2000, Maia3 gives the human rook recapture about 96.71 percent.

Average findability across the seven non-forced positions in that finish is nearly flat: 16.83 percent at 1000, 16.38 percent at 1400, 17.65 percent at 2200. That flatness is not a claim that rating does not help. It means rating helps enormously in some positions and works against the engine's preference in others, and an average quietly cancels the two out. The fuller breakdown is in our piece on Maia chess.

Practically: an analysis board that flags 13...Rxd7 as your mistake is correct and close to useless. Almost nobody at any human rating plays the alternative. The moves worth your twenty minutes are the ones where a player at your level does find the better move, often enough that learning the pattern changes your next game.

How to actually use an analysis board

Whatever tool you are on, the routine is the same and it is short.

Replay the game once without the engine. Stop where you remember feeling something: confusion, excitement, the moment you thought you were winning. Note what you were thinking. Your emotional turning point and the engine's evaluation turning point are usually different moves, and the gap between them is the lesson.

Then turn the engine on, and look for three things only. The last moment the game was balanced. The single biggest swing. And one good move you found, because a review that only hunts errors trains you to play scared.

For each error, back up three moves. The blunder is a symptom. What put your pieces in a relationship where it was available? A free pawn grab, a rook lift that abandoned a defender, a check played because it was a check. That is where the rule you can actually reuse lives.

Ignore small swings and deep lines. A move that takes the evaluation from +0.30 to -0.10 is not a mistake at club level. And the engine's eight-move refutation is true and unusable: if you cannot see why the first move works, the remaining seven teach you nothing. Name the mechanism, a fork, a weak square, a trapped piece, then move on.

Side by side

The number was never the hard part

Any analysis board on this page will tell you that you were winning on move 19 and losing on move 23. Producing that verdict is a solved problem and it is free everywhere. What still separates one tool from another is whether anything after the number tells you why the move was wrong, whether you could have found the alternative, and what to practise so it stops happening. That gap is what an AI chess coach is for, and it is also, honestly, most of what a human coach does with your games.

Findability figures come from our own run on 28 August 2026: maia3-js running the maia3-5m checkpoint against the finish of the Opera Game, Morphy versus the Duke of Brunswick and Count Isouard, Paris 1858, with Stockfish 17 at depth 20 and MultiPV set to every legal move. Third-party engine versions, free-tier limits and pricing change without notice; confirm current details on each product's own pages.