Later on in the article they explain that the initial approach was to include the profitability of an item directly in its metadata so that the algorithm could

Later on in the article they explain that the initial approach was to include the profitability of an item directly in its metadata so that the algorithm could use that in its ranking criteria. They changed their approach to instead measure how much profit Amazon makes on a given search. This allows them to analyze whether various changes to the search algorithm increase or decrease profit, and optimize for those that do. It's a subtle distinction but it means that instead of explicitly promoting their products in search, they are instead making changes based on other product attributes that naturally boost the ranking of their products (and thus increase their profits). It's kind of like a blind study: they're telling the engineers, "we won't tell you which products are ours, but we'll see if you can figure it out based on these 100 other traits."