What actually drives sender reputation

Mailbox providers do not publish their filtering models and never will, because a published model is a model that can be gamed. What they do publish, consistently and for years, is the set of behaviours they are measuring. That is enough to work with, and it points somewhere unglamorous.

The signals that genuinely move the number

Complaint rate

The proportion of delivered messages that recipients mark as spam is the most direct negative signal available. Sustained rates above roughly 0.3% of delivered mail attract active filtering at the major consumer providers; a healthy operating target is below 0.1%.

The number is small enough that it is easy to dismiss and large enough in effect to dominate everything else. One badly targeted send to a stale segment can cost weeks of recovery.

Unknown-user rate

Repeatedly delivering to addresses that do not exist tells a receiver that your list was not collected carefully or is not being maintained. It is the fastest self-inflicted damage available, and it is entirely preventable: hard bounces should be suppressed on the first occurrence, permanently, across every system that sends.

The usual failure is architectural rather than conceptual. The invoicing platform suppresses correctly, the CRM has its own list, and nobody ever built the shared suppression store. Three weeks later the same dead address has been mailed eleven times from four systems.

Spam traps

Two kinds, with very different meanings. Recycled traps are addresses that were once real, were abandoned, and have been repurposed by a provider to catch senders who never clean their lists. Hitting one says your hygiene is poor. Pristine traps were never valid and never opted in to anything — they exist only to be scraped. Hitting one says your data was scraped or purchased, and it is treated far more harshly.

Engagement

Opens, replies, deliberate moves out of the spam folder, and deletions without reading all feed provider-side models. Positive engagement is the strongest thing you have going for you. The consequence is uncomfortable but consistent: recipients who never engage actively suppress placement for the recipients who do.

Volume consistency

Steady, predictable volume reads as legitimate. A domain that sends nothing for six weeks and then dispatches two hundred thousand messages in an afternoon looks statistically identical to a compromised account, and gets treated accordingly.

Authentication

Not merely passing, but passing consistently from a stable set of sources. Authentication is table stakes rather than an advantage — it will not get you into the inbox on its own, but its absence or instability will keep you out.

Domain reputation now matters more than IP reputation

This has been shifting for a decade and the direction is settled. It has two practical consequences.

First, changing infrastructure does not reset your history. Organisations still arrive expecting that moving to a new platform will clear a reputation problem. It will not: the domain carries the record.

Second, dedicated IP addresses are worth considerably less than they are sold as. Below roughly a few hundred thousand messages a month, a dedicated IP never accumulates enough volume to establish a stable reputation, and a well-managed shared pool will usually place better. Buying one because it feels more serious is a common and expensive mistake.

Folklore that will not die

• "Spam word" checkers. Tools that score your copy for trigger words are measuring something that stopped being decisive many years ago. Content is a weak signal compared to behaviour, and writing awkwardly to avoid the word "free" costs you engagement, which is a strong one.
• Image-to-text ratios. A specific ratio has never been a rule. A message that is one large image with no text is a weak signal, mostly because it correlates with senders who do not care; the number itself is invented.
• "An unsubscribe link hurts us." The opposite. A recipient who cannot find an unsubscribe presses the spam button instead, and a complaint costs far more than a lost subscriber. One-click unsubscribe is now effectively mandatory for bulk senders at the major consumer providers, and it must work without a login or a confirmation step.
• Seed lists as a health measure. Sending to a panel of test mailboxes tells you where mail to those mailboxes landed. It says nothing about your actual recipients, whose engagement history is what the model is scoring.
• Blocklist chasing. A handful of blocklists genuinely influence delivery. A great many exist mainly to sell delisting. Know which is which before you pay anyone.

Separate your streams

Transactional mail — password resets, invoices, delivery notifications — should not share a reputation with marketing mail. The two have entirely different engagement profiles, and the transactional stream is the one that must never be delayed. Separate subdomains, separate sending pools, separate suppression logic beyond the shared hard-bounce list.

The intervention that works

If you take one thing from this: implement a sunset policy. Recipients who have not engaged within a defined window stop receiving non-transactional mail. It is consistently the single most effective change available, it is uncomfortable to propose because the list gets smaller, and it reliably improves placement for everyone remaining on it.

Enrol in the free postmaster tooling the major providers offer, watch complaint and unknown-user rates weekly, and resist the urge to fix reputation by writing better subject lines. The model is not reading them.