This depends largely upon the intent of the test. The flexibility of AMP facilitates many different types of testing. Large scale spidering can be a useful way to test large volumes of pages in a situation where an organization wants to create a quick snapshot of the site’s level of accessibility. Determining how many pages to test depends on whether you’re after qualitative or quantitative data and also depends on the size of the site.
SSB BART Group has performed limited work in determining what a statistically significant sample size would be for quantitative data, but as a general rule if you have a site with few pages, you'll want to test a higher percentage of the site’s pages than if you had a very large site. For instance, for a 10,000 page site, you’ll want to spider about 600 pages (6%) whereas for a small site (say, 100 pages), you may need to test 80 pages (80%).
For qualitative data, it may be best to pick and choose the documents you want to test, saving the source of several documents with representative content. Doing so will increase the relevance of your results, because while the spider tests every document it can find, those documents might be repetitious in nature – such as pages which are primarily text, rather than those which are more diverse and may contain content typically known to create accessibility problems.