AI Filters & User-Level Trust
Explore how AI filters and per-user trust models are reshaping inbox placement.
Big mailbox providers like Gmail don’t only judge whether you are a trustworthy sender. They also learn whether each reader wants your mail, so one email can reach one person’s inbox and another person’s spam folder. This article covers what is known about that shift, what is still guesswork, and what you should change in how you send.
How filtering has changed
A spam filter is software at the mailbox provider that decides whether an email goes to the inbox, to the spam folder, or nowhere. Filters have gone through three broad stages. Today’s filters still use ideas from all three.
- Rules (1990s–2000s). Filters looked at the message itself. Certain words, odd headers or known bad servers added points, and too many points meant spam. SpamAssassin is the best-known example. Spammers soon learned to dodge the rules.
- Reputation (2000s–2010s). Filters started asking who sent the email. Providers tracked each sending IP address and domain: how often people complained, how much mail went to dead addresses, how steady the volume was. Authentication made that identity reliable. SPF, DKIM and DMARC are DNS records that prove an email really comes from your domain.
- Machine learning (2010s–today). Machine learning is software that learns patterns from huge numbers of examples, instead of following hand-written rules. Providers train it on billions of messages and on what their users do with them. It can make a separate decision for each message and each reader.
What the big providers look at
Providers don’t publish how their filters work. Below is what they have said in public, plus what senders widely observe.
Gmail
Google has said that Gmail’s filters use machine learning and learn from what users do, such as clicking Report spam or Not spam. Signals Gmail is known or widely believed to use include:
- your domain and IP reputation
- whether SPF, DKIM and DMARC pass
- how this reader treated your past emails
- how many readers report your mail as spam
- the links in the message, checked against lists of harmful sites
- attachments and the technical details in the headers
- how steady your sending volume is
Gmail also sorts mail into tabs such as Primary and Promotions. Tabs are about the type of email, not about trust. A newsletter in Promotions has still reached the inbox.
Microsoft
Microsoft filters mail for Outlook.com and Microsoft 365. It uses IP and domain reputation, authentication, the content of the message, and reports from users who click Junk. It offers senders two feedback tools:
- SNDS (Smart Network Data Services) shows how your IP addresses look to Microsoft.
- JMRP (Junk Mail Reporting Program) sends you a copy of each complaint.
Focused Inbox splits each user’s mail into Focused and Other, based on what that user reads.
Yahoo
Yahoo weighs spam complaints heavily. Its complaint feedback loop tells you when Yahoo users mark your mail as spam. Like Gmail, it also uses reputation, authentication and how its users engage.
Sender reputation vs. user-level trust
Sender reputation is one view of you, shared across all readers: “Does this sender usually send wanted mail?” User-level trust is a separate view for each reader: “Does this person want mail from this sender?”
Picture a shop, shop.example, sending one sale email to three Gmail users. The shop’s reputation is the same each time. But Ana opens and replies to its emails, Ben hasn’t opened one in a year, and Chloe once reported it as spam. The filter may well put the email in Ana’s inbox and Chloe’s spam folder.
This changes how you should read your numbers:
- Averages hide groups. A 25% open rate overall might be 70% among fans and 2% among people who stopped reading. Those two groups get very different inbox placement.
- Sending to everyone gives mixed results. Engaged readers get your email in the inbox. Readers who ignore you are more likely to get it in spam.
- Each reader counts. A good reputation gets you in the door. Whether each reader wants your mail decides where it lands for them.
What builds or breaks trust with one reader
No provider publishes its formula. The table reflects what providers have said and what senders consistently see. Treat the “likely effect” as a rough guide, not a measurement.
| What the reader does | Likely effect |
|---|---|
| Moves your email out of spam (“Not spam”) | Strong positive |
| Replies to you | Strong positive |
| Adds you to their contacts | Positive |
| Opens and reads your emails regularly | Positive |
| Clicks a link | Positive |
| Deletes your emails without opening them | Negative |
| Ignores your emails for months | Slowly negative |
| Reports your email as spam | Strong negative |
Two more things matter:
- Recent behaviour counts more. A reader who opened last week is a better sign than one who last opened six months ago.
- How often you send matters. Someone who opens half of your weekly emails looks keener than someone who opens a few of your daily ones.
The provider sees what readers really do in its own app. Your open tracking does not. Apple Mail Privacy Protection loads images automatically, so many “opens” in your reports aren’t real people reading. Look at clicks, replies and purchases too.
What this means for how you send
Send to the people who want it
Trust builds on itself, in both directions. Mail to engaged readers lands in the inbox, they respond, and trust grows. Mail to silent readers lands in spam, nobody responds, and trust falls.
Silent addresses carry a second risk. Some are abandoned and later turned into recycled spam traps: old addresses a provider reuses to catch senders who never clean their lists.
So segment your list, which means splitting it into groups. Group readers by when they last engaged, for example within 30, 90 or 180 days. Send your regular mail to recent readers. Send less often, or a different message, to the rest.
Ask quiet readers if they still want your mail
A re-permission campaign asks inactive subscribers to confirm they want to stay. Remove the ones who don’t answer. Your list gets smaller, but everyone left wants your mail, and that is what filters reward.
Make every email worth opening
Useful personalisation is based on what a reader bought or chose, not just “Hi Ana”. Triggered emails also help: they go out when a reader does something, such as leaving items in a cart. Relevant mail gets more replies and clicks, which builds trust with each reader.
Make leaving easy
Gmail and Yahoo require bulk senders (5,000+ messages a day to their users) to offer one-click unsubscribe and to keep spam complaints under 0.3%. A reader who can leave in one click is less likely to press Report spam, which hurts you far more.
What may come next
The points below are reasonable expectations, not confirmed facts. Providers rarely announce filter changes.
- More context. Filters may weigh when and where a reader usually reads, or which kinds of email they tend to open.
- Predicting before delivery. A filter could guess whether a reader will want a message and place it accordingly. Some of this may already happen; providers haven’t said how much.
- Signals from outside email. Some people expect providers to use browsing or app activity too. There is no public evidence that they do this for spam filtering.
- More data for senders. Google Postmaster Tools already shows your spam complaint rate and whether you meet Gmail’s sender requirements. Providers may share more over time.
- AI inside the inbox. Gmail, Apple Mail, Outlook and Yahoo Mail now offer AI features that summarise emails or highlight important ones. They shape what readers notice after delivery. See AI & the new deliverability landscape.
Checklist
- Set up SPF, DKIM and DMARC, and confirm they pass with the test inbox.
- Track results by engagement group (last 30, 90, 180 days), not only as one average.
- Send most of your mail to readers who engaged recently.
- Run a re-permission campaign for long-inactive readers, then remove those who don’t answer.
- Personalise from real interests and behaviour, not just first names.
- Offer one-click unsubscribe and keep spam complaints well under 0.3%.
- Check your spam rate in Google Postmaster Tools every week.