What actually happens inside a spam filter before your email arrives
A spam filter isn't one gate reading your words and voting. It's three systems — identity, content and collective memory — scoring you in real time.
By Norbelys Chirinos, Co-founder
Founder-reviewed ·How we research and correct articles
Most explanations of spam filtering describe it as a checklist: authenticate your domain, keep your complaint rate low, avoid trigger words. All true — we’ve written that checklist ourselves. But that’s the diagnostic view, the one useful when something’s already broken. It doesn’t answer a different, more interesting question: what is the filter actually doing, mechanically, in the half-second between your message arriving and a folder being chosen? Here’s that machine, opened up.
It isn’t one gate. It’s three systems running at once.
A modern spam filter isn’t a single program that reads an email top to bottom and outputs a verdict. It’s closer to three separate systems voting, each looking at a different kind of evidence, each finishing its part of the judgment on a different timescale.
Identity runs first, and it’s deterministic. Before a single word of your email is evaluated, the receiving server already knows who’s connecting, whether your authentication checks out, and what your sending domain’s history looks like. This part isn’t a model making a probabilistic guess — it’s closer to a lookup: pass or fail, known-good or known-bad, computed in milliseconds from records that already exist.
Content runs second, and it’s a trained model, not a keyword scan. This is the part most people picture when they think “spam filter” — and it’s the part that’s changed the most. It no longer works the way a spam filter did in 2005.
The content model reads text the way a person does — that’s the actual innovation
For years, content filtering worked roughly like a very literal-minded search: does this message contain “free,” “guarantee,” excessive punctuation, known spam phrases? That approach was always beatable — swap a letter, add a space, use a homoglyph, and the literal match fails even though a human reader instantly recognizes the trick.
Google’s public engineering writeup on RETVec, the text-encoding system behind Gmail’s current spam classifier, describes the fix directly: instead of matching known words character-for-character, the model encodes text more like a human perceives it — resilient to typos, emoji substitution, and the kind of deliberate obfuscation spammers have used for two decades. When Google swapped this into Gmail’s production classifier, they reported a 38% improvement in spam-detection rate and a 19.4% reduction in false positives — catching more actual spam while wrongly flagging fewer legitimate emails — while cutting the model’s compute cost by 83%. That last number matters more than it looks: a cheaper model is one a provider can afford to run on every single message, not just a sampled subset.
The third system is memory — and it updates itself
Here’s the part that’s genuinely different from how filtering worked a decade ago: the model isn’t static between updates. Every time a recipient reports a message as spam, or rescues one from the junk folder, that action becomes a labeled example the underlying model can learn from. Multiply that by a user base in the billions, and the system isn’t just applying rules someone wrote — it’s continuously re-learning what its own users currently consider spam, at a scale where patterns that would take a human analyst weeks to notice can shift the model’s behavior within a much shorter window.
Google states that Gmail’s defenses now stop more than 99.9% of spam, phishing and malware before it reaches an inbox, blocking on the order of 15 billion unwanted messages a day. That figure isn’t really a statement about any single email — it’s a statement about a system operating at a scale where “a human read this and decided” stopped being physically possible a long time ago. Nobody is looking at your specific message. A system trained on the reactions of everyone else’s mail is.
Why the same email can pass for one sender and fail for another
Put the three systems together and the mechanic that actually matters comes into focus: identity, content and memory aren’t three separate checks run in sequence toward a single pass/fail gate. They’re three inputs blended into one continuously updated score attached to you — your domain, your sending pattern, your history of what real people have done with your mail before. [We’ve written about the practical version of this] (/blog/why-emails-go-to-spam/): reputation outweighs content, and it decays back toward neutral slowly rather than resetting the moment you fix something. The mechanical reason that’s true is everything above — reputation isn’t a separate box on a checklist, it’s the accumulated output of the memory system, feeding directly into the same score the content model contributes to.
That’s also why the same subject line and body can land in the primary inbox for a sender with a clean two-year history and get quietly folder-ed for a sender nobody’s mail system has any relationship with yet. The words are identical. The score isn’t, because the score was never really about the words alone.
What this means if you’re the one sending
None of this is a reason to write worse copy — content still contributes to the score, and an obviously spammy message won’t survive a good identity history either. But it reframes where the leverage actually is: a filter built to learn continuously from collective behavior rewards senders whose mail keeps getting treated well by real recipients over time far more than it rewards any single clever subject line. [Real engagement — opens that survive privacy-proxy noise and actual replies] (/blog/half-your-opens-are-robots/) — is quite literally training data the system uses against your identity going forward, for better or worse.
Frequently asked questions
Does this mean the words in my email genuinely don't matter?
They matter less than most people assume, but not zero — the content model is real and does score obviously spammy patterns down. What's changed is that it's no longer a literal keyword match you can dodge with a synonym; it's a learned pattern that's harder to game and easier to satisfy just by writing a normal, relevant email.
If the model updates itself constantly, can my reputation recover fast after a bad send?
Recovery is gradual, not instant, because the system is weighing a rolling history rather than resetting on your most recent action. Consistent good behavior over weeks, not a single clean campaign, is what shifts the accumulated score back.
Is this the same for every mailbox provider, not just Gmail?
The general architecture — identity checks, a learned content model, and behavior-based reputation — is common across major providers, though the specific models, thresholds and how much weight each factor gets differ by provider and aren't fully public.
The part you can actually control
You can’t see the model. You can control what it learns about you: whether your authentication is clean on every send, whether the people you email actually want to hear from you, and whether your sending pattern looks like a real correspondent’s or a fresh burst of unfamiliar volume. That’s the whole job of a deliverability-first sending platform — Norbelys keeps DMARC, SPF and DKIM monitored continuously, runs new mailboxes through a real warmup ramp instead of throwing them at full volume on day one, and reports opens and clicks with the bot and scanner noise filtered out, on every plan, not a paid tier — because the system judging you is watching those exact signals, every time you hit send. Start sending with Norbelys and let the machine’s memory of you start working in your favor instead of against it.