Your inbox knows people your address book forgot.
The first version counted how often an address appeared. That produced too much noise. Newsletters, receipts, automated services, and copied distribution lists can generate many messages without representing a personal relationship.
The useful filter was two-way conversation. A thread counted only when I had sent at least one message as well as received one. The experiment also counted conversation threads rather than every reply inside a long chain.
Several bidirectional threads provided enough evidence to surface a candidate. Automated senders, obvious services, and my own aliases were excluded before the remaining addresses were checked against Google Contacts.
The output was not an automatic import. It was a small review list of people who appeared to be real correspondents but were not already saved. A person could add, dismiss, or ignore each suggestion.
That structure could fit Goodword as well. A connected mailbox already contained relationship signals. The same rules could periodically suggest people who belonged in the user's network without treating every sender as a meaningful contact.
The private experiment includes names, email addresses, counts, and the contents of a personal mailbox. None of those details are reproduced here. This article does not publish the candidates, the size of the result, or the thresholds used in the test.
The public-safe event is the filtering method. Look at recent email, require real back-and-forth, count separate conversations, remove non-people, and compare the result with the existing address book.
I sent the experiment to the Goodword product team as an idea for a future suggestion feature.
The test showed the shape of the workflow. The next product decision was how to ask permission, explain why each person appeared, and let the user control whether any suggested contact was saved.