How do models “learn” from user actions (spam reports, opens, deletes)?

Every time someone hits the spam button, opens your email, or deletes it without reading, they're casting a vote. Mailbox providers like Gmail and Outlook collect millions of those votes and feed them into the machine learning models that decide where your next email lands.

Here's how each action actually gets interpreted:

• Spam reports are the loudest signal. One report won't sink you, but a pattern of them tells the model your messages aren't wanted. The model updates its picture of you as a sender, and that update ripples outward to other users too.
• Opens, clicks, and replies are positive signals. They tell the model that recipients are genuinely engaging, not just tolerating your emails. Replies carry especially high weight because they're rare and deliberate.
• Deleting without opening is a quiet negative signal. The model notices. It doesn't scream like a spam report, but enough deleted-unopened patterns builds a case against you over time.
• Moving from spam to inbox (or starring an email) is a positive correction. It tells the model the filter made a mistake for that user.

The learning works on two levels. At the individual level, the model personalizes each person's inbox based on their own behavior. At the collective level, it aggregates signals across all users. So if a large chunk of your list is reporting your emails as spam, people who haven't complained yet will start seeing your emails filtered too. That's why sender reputation is a collective score, not just a personal one.

What can you do about it? Quite a bit, actually. Send to people who actually want your emails. Remove unengaged subscribers before they turn into spam reporters. Make sure every email you send has a clear reason to exist for the recipient. The model is essentially a proxy for your audience's opinion of you. Treat your list well, and the model will reflect that.

If you want to see how your current setup looks from the outside, our free Blocklist Checker is a good starting point. And if you're seeing sudden filtering changes and can't figure out why, the SOS hotline is free.