When you click the yellow flag, or when Gmail decides on its own to apply one, you are participating in a feature Google calls importance prediction. According to Google's own documentation, the system evaluates messages using signals including:
When you hover and Gmail says "Teach Gmail this conversation is important," you are giving the model a labeled training example. That signal is then used to predict importance on future messages — from the same sender, on the same thread, with similar content, or in similar patterns.
The output is not just visual (the yellow flag). It feeds the Priority Inbox, default sorting, and notification behavior on mobile.
Both. The "who" and "when" portions are pure metadata — sender, recipient, timestamps, frequency, in-reply-to relationships. The "keywords you frequently read" portion requires Google to look at message content. The model lives on Google's servers and is trained on your account's data.
Importance prediction is one of several systems that read your mail. The major ones, all running on Google's servers:
| System | What it does | Data used |
|---|---|---|
| Spam & phishing detection | Filters obvious junk and malicious mail before it reaches your inbox | Full content of all messages |
| Smart Reply | Suggests one-tap replies ("Sounds good!" / "I'll check") | Content of the incoming message |
| Smart Compose | Predicts the next phrase as you type | Content you're writing + history |
| Auto-categorization | Sorts mail into Primary, Promotions, Social, Updates | Content + sender reputation |
| Nudges | "You may have forgotten to reply" reminders | Sent-message history + read state |
| Calendar / Travel integration | Auto-extracts flights, hotels, packages into Google Calendar and Maps | Full content of confirmation emails |
| Tabbed inbox | Identifies marketing mail | Sender, unsubscribe headers, content patterns |
All of this requires server-side processing of your message content. Gmail is end-to-end encrypted in transit (TLS), but at rest on Google's servers it is readable by Google's systems — by design, because the features above could not exist otherwise.
In June 2017, Google announced it would stop scanning the contents of consumer Gmail messages for the purpose of personalizing ads. This was a real change, and it remains in effect — content of your Gmail is no longer used to target advertising shown in other Google products.
What did not change:
The 2017 change is often misremembered as "Google stopped reading my email." It did not. It stopped one specific commercial use of email content.
Even setting content aside, the metadata Google builds from Gmail is extraordinarily revealing:
Phone metadata showed, in the post-Snowden era, that who you talk to and when is often more sensitive than what you say. The same is true of email — arguably more so, because email metadata includes subject lines, which sit ambiguously between metadata and content.
If you use Gmail, the relevant settings are:
The only way to remove Google from the inferential loop is to move email off Gmail. Privacy-focused providers — Proton Mail, Tutanota, Mailbox.org, Fastmail — operate under different business models and, in the case of Proton and Tutanota, encrypt messages at rest in a way the provider itself cannot read. This trades some convenience (no Smart Reply, no calendar auto-extract) for a structurally different privacy posture.
The "Teach Gmail this conversation is important" tooltip is a small, honest disclosure of a much larger reality: Gmail is a learning system that profiles your communications continuously, drawing on both content and metadata, and most of the inferences happen invisibly. The 2017 end of ad-scanning narrowed one commercial use of that data but did not pause the analysis itself.
For most users the practical action items are: