diff --git a/pytket/extensions/cutensornet/structured_state/mps_gate.py b/pytket/extensions/cutensornet/structured_state/mps_gate.py
index c21dc958..af66f432 100644
--- a/pytket/extensions/cutensornet/structured_state/mps_gate.py
+++ b/pytket/extensions/cutensornet/structured_state/mps_gate.py
@@ -96,27 +96,16 @@ def _apply_2q_gate(self, positions: tuple[int, int], gate: Op) -> MPSxGate:
         l_pos = min(positions)
         r_pos = max(positions)
 
+        # Always canonicalise. Even in the case of exact simulation (no truncation)
+        # canonicalisation may reduce the bond dimension (thanks to reduced QR).
+        self.canonicalise(l_pos, r_pos)
+
         # Figure out the new dimension of the shared virtual bond
         new_dim = 2 * min(
             self.get_virtual_dimensions(l_pos)[0],
             self.get_virtual_dimensions(r_pos)[1],
         )
 
-        # Canonicalisation may be required if `new_dim` is larger than `chi`
-        # or if set by `truncation_fidelity`
-        if new_dim > self._cfg.chi or self._cfg.truncation_fidelity < 1:
-            # If truncation required, convert to canonical form before
-            # contracting. Avoids the need to apply gauge transformations
-            # to the larger tensor resulting from the contraction.
-            self.canonicalise(l_pos, r_pos)
-
-            # Since canonicalisation may change the dimension of the bonds,
-            # we need to recalculate the value of `new_dim`
-            new_dim = 2 * min(
-                self.get_virtual_dimensions(l_pos)[0],
-                self.get_virtual_dimensions(r_pos)[1],
-            )
-
         # Load the gate's unitary to the GPU memory
         gate_unitary = gate.get_unitary().astype(dtype=self._cfg._complex_t, copy=False)
         gate_tensor = cp.asarray(gate_unitary, dtype=self._cfg._complex_t)