How do Bayesian and machine-learning filters influence placement?

Spam filters don't just scan your email for suspicious words anymore. Today's placement decisions happen inside systems that have been trained on billions of messages, and they're making judgments that go far deeper than whether you used the word "free" too many times.

Here's how the two main approaches work, and why together they shape where your email lands.

Bayesian filters: the word-pattern layer

Bayesian filters calculate the statistical probability that a message is spam based on the words and phrases it contains. They learn from examples: here are 10,000 confirmed spam emails, here are 10,000 legitimate ones. Over time the filter builds a model of which words, combinations, and patterns skew toward spam.

These filters are fast and still useful. But they're easy to game if you know what you're doing, and they can't see anything beyond the message itself. They have no idea who you are, what your history looks like, or whether your subscribers actually want to hear from you.

Machine learning filters: the full picture

Modern ML-based filters (the kind running inside Gmail, Outlook, and Yahoo Mail) evaluate hundreds of signals at once. Content analysis is just one input. The bigger factors are behavioral and reputational.

What those signals include:

• Your sending history. How long you've been sending from this domain and IP, and whether your sending volume is consistent or erratic
• Authentication status. Whether your SPF, DKIM, and DMARC records are properly configured and passing
• Engagement patterns. How recipients have responded to your previous emails (opens, clicks, moves to spam, deletes without reading)
• Recipient-level behavior. Whether this specific inbox has opened your emails before, which personalizes the placement decision to that individual
• URL and domain reputation. Whether the links inside your email point to domains that have appeared in spam campaigns
• Sending infrastructure signals. Shared vs dedicated IP, ramp-up patterns, bounce rates, complaint rates

The key thing to understand is that Bayesian analysis typically feeds into the larger ML model as one signal among many. It's not either/or. The word-pattern check happens, then that output gets weighted alongside everything else.

What this means for your placement

Because ML filters learn and adapt continuously, your placement outcome isn't fixed. It changes based on how your list responds to you. A campaign with strong opens and clicks teaches the filter that your emails belong in the inbox. A campaign with low engagement, high deletes, and spam reports teaches it the opposite.

And this is why engagement metrics matter so much more than they used to. Cleaning up your subject line helps at the Bayesian layer. But cleaning up your list, removing people who never open, and sending to people who actually want your emails, those actions move the needle at the ML layer.

Still it also means that no single factor determines inbox vs spam. An email from captain@deepcurrent.io with perfect authentication but a disengaged list can still land in spam. An email with a few imperfect content signals but a strong sender history and engaged subscribers often makes it through fine.

The practical takeaway is that you can't word-craft your way out of a deliverability problem. You have to earn it, over time, by sending things people actually want to receive.

If you want to check how your authentication is sitting (since that's one of the first signals ML filters evaluate), our free Email Header Analyzer shows you what filters are actually seeing when your message arrives.