Draw a line in the sand for where metadata pulls from
... and please publish that info.
SmugMug reads metadata from multiple possible locations, but this is currently only known via trial and error, or assumption.
In some cases this behavior is favorable, as a way of displaying data as a string potentially. Usually, being able to pull from multiple fields but display one of them for a common field is a good idea when you (the viewer application, in this case smugmug.com and related apps) can't predict which fields will be populated.
The trick is in prioritizing the fields. And so, assumptions are still being made. This is fine, but let us know exactly what they are so that we may act accordingly. It would be helpful if this information were published by SmugMug.
**For example, this is stated on the site: "Specifically, we look for keywords in the IPTC keywords field."
This is incomplete. For some users, it'll be functionally wrong.**
SmugMug does pull data from IPTC:Keywords but only if (for example) XMP:Subject is blank. If for whatever reason XMP:Subject has text in it, that text will appear on a SmugMug image page as "Keywords."
Now, say you've added additional keywords (or even deleted a few) in your external editor (in IPTC:Keywords) and then replaced those images on SmugMug. None of your changes appear on SmugMug, because XMP:Subject has been given priority over anything that might be living in IPTC:Keywords. You must delete the values present in XMP:Subject in order for IPTC:Keywords to be picked up by SmugMug.
Ask me how I know.
Why not just edit keywords on SmugMug? I don't, because SmugMug isn't an archive (metadata edited there will exist no where else ... the site won't embed metadata, which makes exporting/downloading "originals" only relevant insofar as however you uploaded them.
In all cases, SmugMug reads an incoming file with the open source exiftool (a perl program). What it DOES with that information, is the unknown.