Three Openings Where the Right Reply Depends on Rating

Case 1: The King's Indian's Bishop Test

1. d4 Nf6 2. c4 g6 3. Nc3 Bg7 4. e4 d6 5. Nf3 O-O
White has two natural squares for the king's bishop here: d3 or e2. The popularity ranking between them flips with rating — Bd3 is narrowly ahead at u1400, but Be2 pulls steadily away at higher tiers.
White's 6th move, by level
u1400: 6.Bd3 (24.6%) narrowly ahead of Be2 (21.5%)
u1600: 6.Be2 (27.1%)
u1800: 6.Be2 (33.8%)
u2000: 6.Be2 (44.1%)
That part isn't surprising on its own. What's striking is what happens after White plays Be2 specifically. At u1600 and u1800, the data's recommended reply is the same: the immediate central break 6...e5 (eval -0.49). At u2000 — facing the exact same White move, from the exact same position — the recommended reply changes to 6...Na6 instead, delaying e5 by a full move (eval -0.54).
The point
Same position. Same opponent move. Different correct reply, purely as a function of which rating band the line is calibrated for.

Case 2: The London's Silent Pivot

1. d4 d5 2. Bf4
White's first three moves never change across any rating band in the London — d4, Bf4, and e3 are fixed regardless of who's on the other side of the board. The interesting movement happens entirely on Black's side.
Black's most common reply to 2.Bf4
u1400: 2...Nc6 (38.8%)
u1600: 2...Nc6 (33.3%) — share already shrinking
u1800: 2...Nc6 (25.5%) — shrinking further
u2000: 2...Nf6 (29.2%) — the lead changes hands
Because White's own setup is constant, this looks at first like it shouldn't matter much. It does — the resulting middlegame structures are genuinely different at each tier: a fairly symmetrical fight at u1400, a sharp Bg4-pin structure at u1600, Nimzo-style piece play at u1800, and a patient Queen's-Gambit-Declined-style buildup at u2000 — all stemming from one shift in what the opponent tends to choose.

Case 3: The Scandinavian's Outlier in the Middle

1. e4 d5 2. exd5 Qxd5 3. Nc3
White's 3.Nc3 attacks the queen, and textbook theory doesn't actually agree on a single reply here — Qa5, Qd6, and Qd8 are all established main systems, each with its own body of theory and its own proponents. RBOT's data picks a clear winner among them at three of the four tiers.
Black's reply to 3.Nc3, by level
u1400: 3...Qd8 (eval -0.54)
u1600: 3...Qa5 (eval -0.54) — the outlier
u1800: 3...Qd8 (eval -0.57)
u2000: 3...Qd8 (eval -0.59)
This is the case that breaks the simplest possible story. If opening theory simply got "more precise" as rating climbed, you'd expect a clean progression — maybe Qa5 holding up at lower levels and Qd8 taking over once opponents play accurately enough to punish it. That's not what the data shows. The middle tier is the outlier, sandwiched between two tiers that agree with each other on either side.
The point
A model built on intuition about "how theory should refine with rating" would not have predicted this. Only looking directly at what real games at each level actually produce does.

Why This Happens

None of this means the underlying chess is unstable or that engine evaluation is unreliable. It means "best move" in practical opening preparation was never really being computed in a vacuum — it's computed against the realistic distribution of what happens next. A reply that scores best against the follow-up play typical of 1600-rated opponents isn't automatically the reply that scores best against 2000-rated follow-up play, even from an identical starting position, because the moves after it interact with what the opponent is statistically likely to do.
Most opening books can't capture this, because they're written for one assumed level of opposition — usually, implicitly, a strong one — and applied uniformly to everyone reading them. These three cases are about as clear a demonstration as exists of why that's a real gap, not a theoretical nitpick. It's also the entire premise behind calibrating a repertoire to rating in the first place, rather than handing every player the same "main line" regardless of who they actually face across the board.

More statistics of this kind is presented at rbotchess.com.

Stefan Rebner is the developer of RBOTChess (Rating Based Opening Training), a platform for opening preparation built around Lichess game data at specific rating levels, using top engine analysis applied to opponent move frequencies at each rating level rather than to theoretical best play. Previous articles in this series: What Your Opening Book Doesn't Know About Your Opponents and Where Do the Pieces Go? and also A Different Approach to Chess Improvement— all available on Lichess.