Allegedly the 'clear' answer is much easier to manipulate than gaming PageRank ever was:
Currently it is probably just mostly quantity that does the trick w.r.t. training data. So e.g. spam the Internet with "product comparisons" featuring your product as the winner.
It’s been a long time since agents couldn’t even conduct web search and could only riff off their model. But the examples in this thread are things an agent would search for immediately, and agents are leaning harder on tool calls and external info over time, not less.