How do mailbox providers use machine learning for filtering?

You send a perfectly crafted email to 10,000 subscribers. Most land in the inbox. A handful end up in spam. A few disappear entirely. No error messages, no explanation. What just happened?

Mailbox providers like Gmail, Outlook, and Yahoo Mail aren't running your email through a simple checklist. They're running it through machine learning models trained on billions of messages. These models have seen more spam than any human ever could, and they've learned to spot patterns that no hand-written rule would catch.

Here's the basic idea. The models are trained on labeled examples: this is spam, this isn't. They learn which combinations of signals predict the right outcome. Over time, as spammers adapt, the models retrain on new data and adapt right back. It's not static. It never stops learning.

What are the signals? That's actually the more useful question for senders, and it's covered in detail in the next entry in this series. But broadly, the models are looking at three layers.

• Technical signals. Your IP reputation, domain age, whether SPF, DKIM, and DMARC are set up correctly. These are the first things checked before a human ever reads a word of your email.
• Content signals. Subject lines, body text, link patterns, image-to-text ratios, HTML structure. The model isn't just scanning for banned words. It's looking at the overall profile of the message.
• Behavioral signals. What recipients have done with your previous emails. Opens, clicks, deletes without reading, and spam reports all feed back into the model. Gmail in particular is known for personalizing filtering at the individual recipient level, not just the sender level.

That last point matters a lot. Even if your domain has a good reputation overall, if a specific segment of your list consistently ignores or deletes your emails, the model learns that those people don't want your mail. And it acts accordingly (for them, specifically).

Is it a black box? Partly, yes. The exact weights and architecture are proprietary and change constantly. But it's not random. Good senders who generate positive engagement signals consistently see better placement. The model is, at its core, trying to answer one question: does this recipient want this email? Everything you do as a sender either helps or hurts that answer.

If you want to know where your domain stands technically before the ML layer even gets involved, our free Blocklist Checker is a good starting point.