ChessGrader / Guides / Is Game Review accurate?

Is Chess.com's Game Review accurate?

For practical purposes, yes — it runs real Stockfish and the grades are broadly right. But the search settings and fitted scoring models remain private, and a few numbers mean less than they look like they mean.

By the ChessGrader team · Updated July 10, 2026

Short answer: Game Review is accurate enough to trust for the thing it's for — finding where your game went wrong. The engine underneath is Stockfish, the strongest chess engine in the world, and a move it calls a blunder almost always is one. The caveats are real but they live one layer up: in the labels, the accuracy score, and the commentary, all of which depend on models chess.com describes publicly without publishing in full.

What Game Review actually runs

Game Review evaluates your moves with Stockfish running server-side, then layers three things on top of the raw evaluations:

• Move classifications — Brilliant through Blunder, assigned by its Classification V2 Expected Points Model. Chess.com publishes the inputs, the role of player rating, and the 2/5/10/20 boundaries for ordinary labels, while keeping the fitted evaluation-to-expected-score model private.
• An accuracy score — 0 to 100, from its CAPS2 system. The broad idea is public; the exact formula is private.
• Coach commentary — generated natural-language explanations of what the engine found.

What's notably absent from all their documentation: the search depth, node count, or time the engine spends per move. That matters, because a shallow Stockfish and a deep Stockfish can genuinely disagree about a sharp position. Chess.com even has a help article titled "Why did my move classification change in Game Review?" — the honest answer is that re-analysis at different settings can flip a borderline label. That's not a scandal; it's how engines work. But it means a single Game Review grade is a good estimate, not a measurement.

Where it can mislead you

Three specific places deserve skepticism:

• Brilliant moves. The !! label is the most debated feature in the product. Because the detection formula rewards sacrifices, players have found reliable recipes for farming Brilliancies with objectively unnecessary queen sacs in already-won positions. If Game Review calls your move brilliant, it was probably good — but the label measures flashiness plus soundness, not depth of idea, and different sites will disagree about the same move.
• Coach commentary. The explanations are simplifications by design. "This loses a pawn" might be technically true while the real problem is that your king never castles. Treat the coach as a first draft of the story, not the story.
• The accuracy score. 85% in a 3-minute blitz game and 85% in a rapid game are not the same achievement, and accuracy runs higher in one-sided games where the moves pick themselves. Comparing your accuracy across time controls, or against a friend who plays different openings, mostly measures noise. Here's what accuracy actually tells you.

Why the same game gets different numbers on different sites

Run one game through chess.com, Lichess, and ChessGrader and you'll get three accuracy numbers and some disagreeing labels. Nothing is broken. Each site chooses its own engine depth and its own formulas for converting evaluations into labels and scores. A move that lost 28 centipawns is a fact; whether that's an "Excellent" or a "Good" is a naming convention. Lichess publishes its accuracy method; chess.com describes CAPS2 but keeps the exact formula private; the two were never going to match. We've written up the full Lichess vs chess.com comparison if you want the details.

What's public vs proprietary

How to sanity-check a suspicious grade

If a label feels wrong — a "blunder" you're sure was fine, a Brilliant that felt routine — don't argue with a black box. Run the game through a second tool whose methodology is published, so you can see exactly why it graded what it did. ChessGrader's grading pipeline is documented: local Stockfish 17.1, fixed 100,000-node searches per move, 600,000-node verification of critical moments, and rating-aware expected-points boundaries with centipawn and mate safeguards. Agreement is useful evidence; disagreement can come from search depth, calibration, or a genuinely sharp position worth studying.

To be clear about our bias: ChessGrader is a free unlimited game review tool, so of course we'd suggest it as the second opinion. The reason it's a fair one is that our numbers are checkable and chess.com's aren't. Not better — checkable.

The verdict

Game Review is genuinely good. The engine is world-class, the blunder detection is reliable, and the presentation has taught more people to review their games than any product in chess history. Trust it for direction: where the game turned, which moves to study. Hold it loosely on precision: exact accuracy points, borderline labels, and anything called Brilliant. And never compare its numbers across sites, time controls, or friends and expect them to mean the same thing.