For a shopper, not being able to find what you want (or trust what you find) is a high cost.
The cost of switching to another site, or shopping mode, is low.
More to the point: it's vastly lower than the shopper being able to effectively fix Amazon's search.
This is a general principle of networks and positive- vs. negative-value members or additions.
The naive Metcalf's Law notion, that all members of a network are a net positive, is false, and even the far more useful Tilly-Odlyzko formulation (V = nlog(n)) fails to account for nodes contributing a negative cost. Since all* information imposes an attention cost, you can approximate the actual network value as:
V = n*log(n) - k*n
Where k is some cost constant.
In fact the size of the network is determined by the cost constant. The lower the constant, the larger the sustainable network size.
At some point, adding more members reduces total network value. Worse, since you have high-value and low-value contributors, and quite possibly a higher value-sensitivity of high-value members, as the network value approaches and passes the zero point, high-value members tend to defect. That's what happens as a social network tends to low-quality posts, content, and discussion.
Or a shopping market tends to counterfeit goods, mislabeled content, fraud and the like.
Just as MySpace found itself walking dead, and numerous earlier retail establishments, Amazon could find itself on the wrong side of this line and quickly.
Also: though I represent k as a constant, it's better to think of it at any given point in time as being mostly undifferentiated amongst nodes. But over time that constant might increase or decrease, whether due to the behaviour of nodes, additions or deletions in nodes, or in environmental factors.
Is this backward?