The objective of this study was to evaluate the effects of reducing crude protein (CP) and CP substitution strategy on performance when primiparous dairy cows were fed diets continuously or according to a change-over experimental design. Fifty-four primiparous, Holstein cows were randomly assigned to either a randomized complete block design (CONT; n = 36, initial mean ± SD: 129 ± 36 DIM, 580 ± 40 kg BW, 44.1 ± 3.2 kg milk) or to a replicated, 3x3 Latin square design balanced for the effects of previous treatment (CHANGE; n = 18, initial mean ± SD: 129 ± 35 DIM, 583 ± 39 kg BW, 44.4 ± 3.7 kg milk). Experimental designs were run concurrently with three 28-d periods and sampling on d 22–28 of each period. Cows were milked 3× daily and were individually fed once daily a diet that was predicted to be either adequate (ADMP; 16.7%CP, 28.3%NDF, 25.5% starch) or deficient (LOMP) in metabolizable protein (MP). ADMP contained expellers soybean meal, whereas this was removed in LOMP diets and replaced with either dry ground corn (STARCH; 14.9%CP, 28.1%NDF, 28.6% starch) or soyhulls (FIBER; 15.1%CP, 30.7%NDF, 25.4% starch) for a total of 3 diets. Contrasts for the effects of MP (ADMP vs LOMP) and carbohydrate source (CHO: STARCH vs FIBER) were evaluated for both experimental designs, with P < 0.05 significant and P < 0.10 trends. DMI was greater for ADMP than LOMP in CONT, whereas DMI only tended to be greater in CHANGE. Yield of milk and protein were greater for ADMP in both designs whereas yield of fat was greater for ADMP in CHANGE, but only tended to be greater for ADMP in CONT. Within CONT, CHO did not affect DMI, milk, or component yield; however within CHANGE, cows fed STARCH produced more milk and tended to produce more protein. This discrepancy was not due to statistical power of the designs because the results were directionally opposite from CONT where cows fed FIBER had yields that were numerically, though not significantly, greater. Inferences about MP status were generally similar for production measures in both designs whereas inferences about CHO were affected by experimental design.