[pacificenigma]

You start with SPF, DKIM, DMARC and IP reputation.

But then I'd use the moat GMail enjoys as a primary mail destination by statistically looking at the frequency an incoming email's metadata and content similarity has recently arrived at GMail as a whole, then moving suspicious arrivals into a progressive exposure pathway that tentatively delivers a small percent to known currently-active users (eg mobile GMail app open on screen and unlocked) and see what percentage are flagged as spam, archived or deleted. Then use that real time feedback to vary the delivery flow of remaining messages to inbox vs spam.