How do AI and ML algorithms measure engagement patterns?

You can't see Gmail's spam filter code. Nobody can. But from the outside, it behaves like a system that's learned what "good" and "bad" look like from billions of interactions across billions of mailboxes, and then applies that pattern recognition to every new email that comes in.

That's machine learning in the context of email filtering. And the signals it's learning from are almost entirely behavioral.

What signals feed the models

Mailbox providers track what users actually do with emails: opens, clicks, replies, deletions, spam reports, moves to spam, moves out of spam, and (at Gmail) implicit signals like how long a thread stays in the inbox before being archived or deleted. These are the training labels for the model.

Over time, the model learns that emails from domain X tend to get replied to, while emails from domain Y tend to get deleted immediately. It learns that a certain subject line pattern correlates with high spam complaints. It learns that senders with certain IP reputation histories tend to produce complaints at higher rates.

How it applies to your domain

The model doesn't just learn general patterns. It learns per-sender patterns and per-recipient patterns. Your domain builds a history based on how its emails have been treated across all Gmail users who've received them. A new subscriber's placement depends on both your domain history and that subscriber's personal behavior patterns (what they typically do with promotional email, for instance).

So this is why engagement degradation is cumulative. If you send to a large inactive cohort for six months, the model has accumulated a lot of evidence that those recipients don't want your emails. That's not easy to reverse quickly.

What you can actually control

You can't directly influence the model. What you can do is create the behavioral signals that feed into it positively: segment to engaged subscribers, send content people actually open and click, remove inactive addresses before they dilute your signal, and make unsubscribing frictionless so disinterested people leave cleanly rather than ignoring you.

For more on how those engagement signals translate to sender reputation, and what negative engagement looks like to these systems, those two answers cover it well together.