In early 2007, the California State University (CSU) system initiated a program to make all information technology accessible to persons with disabilities; the authors were appointed to lead the Web portion of this effort. Campuses initially hoped to rely on automated testing of Web sites to evaluate their level of accessibility and determine what repairs would be needed. However, we found that automated testing tools were severely limited both in their ability to identify relevant problems and, surprisingly, in their scalability to sites of CSU magnitude. The first manual (non-automated, human intelligence based) testing protocols that we developed proved to be awkward in practice; this paper reports our substantial recent refinements after extensive use and incorporation of the newest accessibility guidelines. We also suggest how this protocol can become part of the Web development workflow, rather than being used solely to check legacy sites.