Accurately estimating stream channel characteristics is essential for managing and restoring populations and aquatic ecosystems. Reach-based sampling designs have been used extensively to collect fisheries related data; however, few studies have examined accuracy and precision of scaling up reach-based sampling designs to stream habitat assessments. Here, we used continuous habitat surveys across multiple reaches to census stream attributes in tributaries in the upper Lewis River, Washington and better understand the potential bias and precision of reach-based designs. We used these continuous data to create simulated outcomes from three different random sampling designs. We found precision of estimates of stream-level habitat attributes (large woody debris, residual pool depth, and grain size) increased with the number of reaches sampled (i.e., sampling intensity); however, effort needed to achieve reasonable precision (coefficient of variation = 0.20) varied across streams, attributes, and designs. Bias (i.e., estimate—the truth) was relatively low, but also varied across streams and attributes. Our findings illustrate the challenges of using reach-based designs for stream-level habitat assessments and the need for novel approaches for broader data collection.