A program placed participants with host organizations, and at each host organization a specific person supervised them. Asked how well the program retains its partners, you can answer that question two ways from the same records.
Count host organizations returning year over year and you get one retention series. Count individual supervisors returning year over year and you get another. Both are correct. They are not the same number, and they do not stay close to each other.
The consequence is not statistical, it is operational. If the gap between the two is large, the program is retaining institutions while losing the people inside them, and the useful intervention is about relationships with individuals. If the two track together, the relationship is institutional and the intervention is about the partnership itself. The choice of what to count decides which of those two conversations the program has.
None of this is visible if the unit of analysis is chosen after the chart is built, which is the usual order. It is worth settling first, in writing, before anyone opens the visualization tool. In a data model where one entity reaches another through a junction, both counts are equally easy to produce, and equally easy to produce by accident.