How spam filters decide where your email lands

Spam filtering is a layered decision: acceptance at the SMTP door, machine-learned scoring, then per-user placement. What actually weighs, in order.

Last updated 19 July 2026

There is no single moment when "the spam filter" reads your email and votes. Filtering is a pipeline of decisions made at different times by different systems: whether to accept the message at all, how to classify it, and where to place it for each individual reader. The same campaign can be refused at the door for one provider, inboxed for one Gmail user and foldered for another, all correctly, by the same rules. Once you see the layers, most deliverability confusion, including nearly all of the "spam words" folklore, falls apart on its own.

Layer one: the SMTP door

Before any content exists to judge, there's a connection. A sending server contacts the receiving server over SMTP, the delivery protocol from the first lesson of this course, and announces who it is and what it wants to deliver. The receiver's first decision is made right here, on connection data alone: the sending IP and its reputation, blocklist results, whether reverse DNS (the lookup that maps an IP address back to a hostname) resolves sanely, how the sender behaves against the protocol.

Three outcomes are possible at the door. Accept the message for further processing. Reject it outright with a permanent error, which is what a blocklisted IP typically earns. Or defer it with a temporary error that says try again later, which is how providers throttle senders they don't yet trust, and how greylisting (deliberately deferring first attempts, since spamware often never retries) works.

The door matters because mail refused here generates a bounce your ESP can show you. Mail that gets through the door and then goes to spam generates nothing. That asymmetry explains a pattern I hear constantly: "our delivery rate is 99%, but nobody's seeing our emails". Delivered means accepted at the door. It says nothing about the folder.

Layer two: classification

Once accepted, the message is scored. Modern filters at the major providers are machine-learned systems weighing thousands of signals per message, and "thousands" is the providers' own characterization, no exaggeration. Reputation of the sending domain and IP. Authentication results. The recipient's history with this sender. URL and attachment analysis. Content features. Similarity to mail already being complained about across the network. Nobody outside those teams knows the weights, and the weights shift continuously as models retrain, so any vendor selling you "the algorithm" is guessing.

The signals do have a knowable hierarchy, though, and two facts about it are worth carrying for life.

Authentication is an entry ticket, not a boost. Passing SPF, DKIM and DMARC doesn't earn you inbox placement; it makes you eligible for it, by letting the filter attach this message to your domain's identity and history with confidence. Fail authentication and you look worse than unidentified: you match the signature of forged mail. Perfect authentication with a bad reputation still folders. It just folders as verifiably you.

Reputation dominates content. The accumulated behavioral record from the sender reputation lesson outweighs anything inside the message body, which is why the same words deliver fine from a trusted sender and folder from a distrusted one. The word "free" in a subject line does not send email to spam; twenty years ago naive keyword filters made that partially true, and the advice has outlived the mechanism by about two decades. Content still participates: a message that is one enormous image, link shorteners pointing through disreputable domains, or a missing unsubscribe link all give a filter easy reasons. But content is the tiebreaker, never the headline act. Senders convinced they have a spam-word problem almost always have a reputation problem.

Layer three: your reader, specifically

The layer that surprises people most: at large providers, placement is personalized. Gmail's filtering openly incorporates each user's own behavior. A sender one user always opens and another always deletes unread can inbox for the first and folder for the second, same campaign, same minute. "Not spam" rescues and "report spam" clicks teach the filter about you globally and about that reader's relationship with you specifically.

Two practical consequences. First, "test it to my own inbox" is nearly worthless as a placement test, because your inbox has a personal history with your mail (you can check what a test can legitimately tell you: authentication, headers and content are measurable per message; placement across your list is not). Second, there is no binary state of "landing in spam". You inbox for some fraction of readers and folder for the rest, and the fraction moves with your reputation.

This is also where I should say plainly what no tool can do, ours included. Nobody outside a mailbox provider can observe per-user placement across your actual list. Services that claim to measure "inbox placement rate" with seed lists are sampling fake mailboxes with no human behavior behind them, which is precisely the signal modern filters weigh most. We deliberately don't do seed-list theater. Inbox Decay measures the verifiable layer, the records and headers and blocklists filters actually consume, and says so, because a measurable input honestly reported beats a fabricated placement percentage.

The big three, qualitatively

The layers exist everywhere, but the providers have personalities, and anyone who mails at volume learns them.

Gmail is the engagement machine. Its filtering leans hardest on behavioral signals and per-user history, it's aggressive about the Promotions tab as a category distinct from spam, and it gives senders Postmaster Tools as a reputation window. Mail that people demonstrably read gets treated remarkably well at Gmail; mail that people ignore slides, whatever its technical polish.

Microsoft runs its own reputation systems for the consumer domains (outlook.com, hotmail.com, live.com) and has a long-standing reputation among practitioners for weighting IP behavior in its own particular way; what's fine at Gmail is sometimes deferred at Outlook, and vice versa. On the business side, Microsoft 365 adds a layer nobody else has at scale: tenant-level rules. The receiving company's administrator can set policies of their own, allow-lists and block-lists included, that override or supplement Microsoft's verdicts, which is why B2B senders sometimes find their mail quarantined by one customer's IT department while every other recipient sees it fine. No global fix exists for a local rule.

Yahoo pioneered much of modern authentication (DKIM has roots in Yahoo's DomainKeys), runs engagement-informed filtering of its own, and moved jointly with Google on the 2024 bulk sender requirements, which the bulk sender rules lesson covers in full. Senders see less tooling from Yahoo than from Google, and broadly similar expectations.

Different personalities, one shared direction: authenticated identity as the entry ticket, behavioral reputation as the decision, content as the tiebreak, and the individual reader as the final judge. Build for that hierarchy, in that order, and the personalities mostly take care of themselves. One receiving world plays by different rules entirely, and it is where B2B revenue lives: when the receiver is a business, not Gmail.

Terms from this lesson

SMTP-time rejection - refusing a message during the delivery conversation, before content processing, typically on IP reputation or blocklist grounds. Produces a bounce the sender can see.
deferral - a temporary "try again later" refusal, used to throttle untrusted senders.
greylisting - deliberately deferring a first delivery attempt, since legitimate servers retry and much spamware doesn't.
classification - the machine-learned scoring stage that weighs reputation, authentication, content and history to place an accepted message.
per-user personalization - placement adjusted per recipient from their own behavior, so one campaign lands differently across readers at the same provider.
entry ticket - what authentication is: passing SPF, DKIM and DMARC makes you eligible for the inbox by tying mail to your identity, without earning placement by itself.
seed list - a set of test mailboxes used to estimate placement. Weak evidence, because fake mailboxes have no human behavior for engagement-driven filters to weigh.
tenant-level rules - policies a Microsoft 365 administrator sets for their own organization, able to override Microsoft's global filtering for that company's mail.

Check yourself

1. A message is refused during the SMTP conversation before any content is examined. Which layer made that call?

2. What does passing SPF, DKIM and DMARC do for placement?

3. The same campaign inboxes for one Gmail user and folders for another. Why?

4. Which matters more to modern filters: sender reputation or the words in the message?