Sender reputation

Sender reputation is the score mailbox providers keep on you — per domain, per sending infrastructure, per behaviour pattern. It decides whether authenticated, well-formed mail actually reaches the inbox. Reputation builds slowly and burns fast: a single bad campaign can cost weeks of inbox placement.

What moves it

In rough order of impact:

1. Complaint rate — recipients clicking "mark as spam". Providers treat this as the strongest negative signal; sustained rates above ~0.1% are dangerous, 0.3% is a cliff.
2. Hard bounce rate — sending to addresses that don't exist reads as list buying or negligence.
3. Spam-trap hits — addresses that exist only to catch senders with bad list hygiene.
4. Engagement — opens, replies, moves-to-inbox versus instant deletes and ignores.
5. Volume consistency — steady, predictable volume looks legitimate; sudden spikes look compromised.

MiniMailer's two-layer model

Most email platforms pool customers on shared infrastructure, so one customer's bad day becomes everyone's deliverability problem. MiniMailer isolates reputation at two levels:

• Per-customer isolation. Your sending runs on infrastructure whose reputation is isolated from every other customer's — a noisy neighbour can never drag down your deliverability, and your good record is yours alone.
• Per-agent scoring within your account. Every send can carry an agent_id, and reputation is scored per agent on top of the per-domain view. When you run a fleet of AI agents, one misbehaving agent shows up in its own metrics instead of silently poisoning the shared domain.

The layers stack — neither replaces the other. Domain-level reputation is what mailbox providers see; agent-level reputation is how you find the culprit before they do.

Agent reputation tiers

Every agent sits in one of four tiers, scored continuously from its own hard bounce rate, complaint rate, and unsubscribe rate:

The tiers are not just labels. A Building or Restricted agent is rate-capped on its own, independently of your account's limits — restricted agents most tightly. That containment is the point: one agent's bad run can neither burn through your account's whole allowance nor keep damaging your domain while you work out what went wrong. Healthy and Dedicated agents send at your account's full tier limit, and sends with no agent_id are never capped this way.

Why unsubscribes count, and why they can't restrict you

Bounces and complaints both need a real person on the other end reacting badly. A list of addresses nobody consented to can produce neither for a long time — and an agent working through one would otherwise score a flawless 100 right up until the damage showed elsewhere. How fast people leave your list is the earliest honest read on whether they wanted the mail, and it is the one number you cannot improve by gaming: the only way to push it down is to send mail people want to keep receiving.

Two deliberate limits on how it is used. A normal unsubscribe rate costs nothing — every healthy marketing programme has people leaving, and the score only reacts to a rate well above that band. And unsubscribes on their own can never push an agent into Restricted, no matter how high the rate goes. Unsubscribing is a right your recipients have and you are obliged to make easy; a system that throttled you for honouring it would be pushing you to bury the link, which is precisely how senders get themselves blocked. A high unsubscribe rate keeps an agent out of Healthy and tells you about it — it does not cap you.

How you're kept informed

• Dashboard — delivery, bounce, and complaint rates per domain, and per-agent reputation metrics.
• Webhooks — domain.reputation_warning and domain.reputation_recovered for domain-level movement; agent.reputation.damaged, agent.pool.promoted, agent.pool.demoted, and agent.pool.recovered for agent-level events. Wire these into your alerting — reputation problems compound while nobody is looking.
• Automatic suppression — hard bounces, complaints, and unsubscribes are suppressed immediately, so the addresses that just hurt you can't hurt you again.

Protecting it

• Double opt-in for marketing lists. The confirmation click is cheap insurance against typos, bots, and spam traps.
• Give every agent its own agent_id. Attribution is free when things are fine and priceless when they aren't.
• React to warnings. A reputation_warning webhook is the moment to pause the offending traffic and inspect it — not after the campaign finishes.
• Retire the unengaged. Recipients who haven't opened anything in months are pure downside; suppress or win them back deliberately.
• Keep transactional and marketing on separate domains so a marketing misstep never delays a password reset.