Content rules decide far less than reputation, and they still reject mail on their own. What rule-based scoring measures, what Bayesian filters learn, and what to fix.
Content filtering is the part of deliverability everyone can see and almost everyone overweights. Rewriting a subject line is cheap and changing infrastructure is not, so teams reach for the word list first and spend weeks there while the actual problem sits in their stream separation or their list quality. The honest ordering is that reputation and authentication decide most placement outcomes and content decides the remainder. That remainder is still real. A message can be rejected on content alone, and knowing which rules exist stops you from optimising against superstitions.
How a score becomes a verdict
SpamAssassin remains the reference implementation for rule-based scoring, and although Gmail, Outlook, and Yahoo run proprietary machine-learned systems instead, a large share of business mail still passes through a gateway running it or something derived from it. The model is simple: dozens of independent rules each contribute a positive or negative number, the numbers are summed, and a threshold decides the verdict. No single rule condemns a message. A message reaches the threshold by accumulating small penalties, which is why the useful question is never whether one phrase is banned, but how many marginal penalties a message carries at once.
X-Spam-Status: No, score=3.9 required=5.0 tests=BAYES_50,
HTML_IMAGE_ONLY_28,HTML_MESSAGE,MIME_HTML_ONLY,
RDNS_NONE,SPF_HELO_NONE,T_REMOTE_IMAGE
X-Spam-Level: ***Read that report and the content rules turn out to be the least interesting entries in it. HTML_IMAGE_ONLY_28 and MIME_HTML_ONLY are structural: the message is essentially one image with no meaningful text and no plain-text alternative, which is the shape of both a badly built campaign and a filter-evading spam. RDNS_NONE and SPF_HELO_NONE are infrastructure faults that cost more points than any phrasing choice ever will. BAYES_50 means the statistical classifier has no opinion, which for an established sender is itself a signal that the mail does not resemble the recipient's normal correspondence. The remediation list here contains one copywriting item and four engineering items.
What the Bayesian layer learns
Rule scores are static and the Bayesian classifier is not. It trains on what recipients in a given deployment mark as spam and as legitimate, tokenising headers and body and building per-token probabilities from that history. In practice this means an identical message scores differently at two organisations, and that your own sending history matters more than the industry-wide reputation of any phrase. It carries a practical consequence worth internalising: consistency helps. A sender whose messages look structurally alike from month to month accumulates a stable classification, while a sender alternating between plain-text updates and image-heavy promotions from the same address is training the classifier on two different senders.
The content faults worth fixing
Template review that earns its time
- A real text/plain part carrying the same substance as the HTML, not a stub telling readers to enable images
- An image-to-text ratio that leaves the message readable with images blocked
- Link domains that match the sending domain, with no URL shorteners and no redirect chains through third parties
- No hidden text, white-on-white styling, or character substitution to evade matching, all of which score heavily
- Valid, closed HTML: unbalanced tags and pasted word-processor markup trigger structural rules
- A subject and preheader that describe the message honestly, with no fake Re: or Fwd: prefixes
- No attachments on bulk mail; link to the file instead
The order of operations matters more than any individual item on that list. Content optimisation applied to a domain with a reputation problem produces nothing measurable, and teams who run the sequence backwards conclude that deliverability work is futile. Fix authentication, fix the sending patterns, then treat content as the last few percent. When mail from a healthy sender lands in spam, the template is a reasonable suspect. When mail from an unhealthy sender lands in spam, the template is a distraction.
Frequently Asked Questions
Do certain words automatically send mail to spam?
Does adding an unsubscribe link improve the score?
How much does the plain-text part matter?
Are scoring services accurate for Gmail?
Key Takeaways
- Reputation and authentication decide most placement, content decides the remainder, and the remainder is real
- Rule-based filters accumulate small penalties toward a threshold, so no single word condemns a message
- Read rule names rather than scores: most points on a typical report come from infrastructure faults
- Structural faults such as image-only mail, a missing text part, and mismatched link domains outweigh word choice
- Fix content last, because template work on a damaged sending reputation produces no measurable change
Related articles
Half-Year Review: Email in the AI Inbox Era
Six months that rearranged the reading layer: Gemini in Gmail, Microsoft rejecting outright, DMARC finally a standard. What the first half of 2026 means for senders.
Deliverability Postmortems: Learning From Incidents
The incident is over, placement recovered, and the pressure to move on is enormous. The blameless postmortem practice that converts each incident into prevention.
Writing for AI Readers: Structure When Gemini Summarizes You
Five months of AI Overviews data shows opens up and clicks down: recipients read summaries. How to structure email so the model's version still does your job.