Cancer stem cells (CSCs) have been studied intensively in recent years due to their potential importance for understanding neoplastic disease and the design of antitumor therapies. A number of properties attributed to CSCs have been used to define the CSC population, the most important of which is the ability to initiate reproducibly the growth of tumors in vivo. Other assays such as spheroid formation, expression of particular markers and label retention are also used for defining CSCs, although the degree to which these assays invariably reflect the ability to form tumors in vivo remains to be carefully evaluated. Given the importance of correctly defining and isolating CSCs if valid conclusions about their characteristics are to be made, we used syngeneic animal models to compare these different assays. In standard spheroid assays, cell aggregation rather than spheroid growth from single cell suspensions ensued, but aggregation was circumvented by the inclusion of methylcellulose in the medium. Label-retaining subpopulations did not reliably exhibit an enhanced ability to form spheroids and were enriched for senescent cells. Spheroid-forming ability was found to correspond to expression of established CSC markers, although not invariably. Furthermore, spheroid-forming ability was not always reflected in tumor-initiating properties in vivo. Long-term culture of primary mammary tumor cells as adherent monolayers increased their tumor-initiating ability in vivo. This increase was attenuated when the cells were subsequently cultivated as spheroids. Together these data indicate that assays that are widely used to define CSC subpopulations do not invariably reflect tumor-initiating properties in vivo.