Email deliverability: why good mail still lands in spam

What deliverability actually means

SPF, DKIM, and DMARC usually get treated as a checklist worked through by hand, one DNS record at a time, with warmup as a slow manual grind on top of it. The AI-native fix is having your agent do the setup and the sending together: verify a domain and add its authentication records through Nitrosend’s MCP server from inside Claude, ChatGPT, Cursor, or Codex, then run every send on the same platform. Correct configuration stops being a project you have to remember to finish; it becomes the default state of a domain your agent already manages. Here’s what actually decides where mail lands once that’s configured.

Email deliverability is whether your message reaches the inbox rather than the spam folder, the promotions tab, or nowhere at all. It is not the same thing as delivery, and conflating the two is the most common mistake in email.

Delivery is a transport fact. The receiving server accepted the message and returned 250 OK. Your logs record a success and your dashboard turns green. Deliverability is a placement fact, and it is decided after that accept, by a filter you cannot see, using signals you mostly do not control directly.

That gap is where the whole discipline lives. A campaign can be 99.8% delivered and 40% inboxed, and nothing in a standard sending log will tell you.

What actually decides placement

I have been doing email for a decade, and my companies have sent over six billion messages. SmartrMail alone had 12,000 customers. Across all of it the same four levers decide placement, in roughly this order of weight.

Engagement is the one people underrate. It is a virtuous cycle: decent engagement improves inboxing, which puts more mail in front of people, which improves engagement. It runs in reverse just as reliably. The more unopened mail you leave sitting in inboxes, the worse your placement gets, which is why sending harder to a quiet list is the exact wrong move.

The 0.30% cliff

Since 2026, Gmail, Yahoo and Microsoft enforce a hard spam-complaint threshold of 0.30% for bulk senders, alongside mandatory authentication and one-click unsubscribe. Three complaints per thousand messages is the line.

Kam, who built our deliverability layer, grades every sending domain's spam rate against that threshold continuously rather than at report time, because the useful moment is before you cross it, not after. A domain drifting toward the line gets throttled automatically. Google Postmaster will also show you the number, but it shows you yesterday.

The reason the cliff matters more than any other metric is that it is the only one with a cliff. Bounce rate degrades you gradually. Complaint rate has a step function in it.

Why open rate no longer measures this

Open rate is noise now. Not as a figure of speech, as a technical fact. Apple's Mail Privacy Protection pre-loads every image, and Gmail proxies them, so an open no longer means a human opened anything. We do not track it as a KPI at all.

We replaced it with reply, activation and retained revenue, and we judge deliverability on inbox-and-primary-tab placement rather than on a contaminated number that makes a dying list look alive. If you are optimising open rate, you are optimising an artefact of image loading.

There is a related trap in content. If you send images only, Gmail, Yahoo and Outlook cannot read an image, so they see a message with no readable content and treat it accordingly.

Fixing it in order

Deliverability problems have a diagnostic order, and doing it out of order wastes weeks. Start at authentication, because nothing downstream matters if the mail is not provably yours. Then check whether you are listed, since a live listing blocks at the MTA before any content check runs. Then look at bounce rate and list hygiene, because that is what moves reputation. Then warm the domain if it is new. Only then does content tuning pay.

Connecting a custom sending domain instead of a Nitrosend brand-subdomain sender is the single highest-leverage move for a new sender. Chong makes the same point from the support side: using a verified customer domain to build portable sender reputation is what actually helps, not tweaking subject lines.

If you want to see where you stand before changing anything, run a deliverability test and measure real inbox placement rather than trusting your delivery log.

Most senders can work that order themselves in an afternoon, and the tooling page covers what to monitor while you do. If placement is still bad after all four are clean, that is the narrow case where deliverability consulting is worth paying for.

All of this applies to application mail too, and arguably matters more there: a password reset that lands in spam is a support ticket, which is the case made on transactional email. If you send through an API rather than SMTP, the email API is where the send path and these controls meet.

Why the standard advice stopped working

I built SmartrMail, took it to 12,000 customers, and sold it in 2022. In those years I watched every single customer spend hours in a dashboard doing work an AI can now do in seconds. Deliverability was the worst of it, because deliverability advice is not really advice. It is a list of chores.

Read the other pages ranking for this term. One hands you about twenty separate tasks. Another gives you seven steps and a closing checklist. Twilio's SendGrid guide runs to six chapters and then says the quiet part twice on the same page: you will either need dedicated staff who understand the ins and outs of email, or you should hand the whole thing to a provider.

That is the honest end of the road for the old model. The chores are real, they work, and almost nobody does them past week one. Lists rot. Nobody re-checks authentication after a DNS change. Nobody watches complaint rate per send. The guides were never wrong. They were addressed to a person who does not exist.

There is a second problem nobody has caught up with. Every content rule in those guides, avoid the all-caps subject line, keep a balanced text-to-HTML ratio, minimise the links, was written for someone eyeballing one email at a time. If a model wrote your copy, nobody eyeballed anything, and that advice has no route to the thing that produced the text.

What changes when an agent sends the mail

Two things, pulling in opposite directions.

The risk goes up. An agent told to email everyone about the launch has no concept of a ramp, no memory of your complaint rate last Tuesday, and no instinct that the subject line it just wrote reads like a payday loan. Speed is the point of an agent, and speed is exactly what wrecks a sending reputation.

The fix also gets better, for the first time in fifteen years. A checklist a human runs once at setup can instead be a check that runs on every send, because the thing sending is software and software can be gated. That is the only structural change this problem has had since SPF. So the question worth asking is no longer what you should do about deliverability. It is what will refuse to send when you get it wrong.

On Nitrosend the answer is eight layers on a single send path, covering marketing and transactional email alike.

The ramp is a property of the account rather than a schedule you are supposed to remember for six weeks. Sending capacity is set by a standing tier, Probation, Standard or Trusted, expressed as a rolling 24-hour recipient limit. On the free plan those tiers are 100, 500 and 5,000. On Pro they are 1,000, 10,000 and 100,000. Standing is earned through attributable delivery history, verified list quality and domain strength, and our own documentation is blunt about the rest: payment alone does not earn trust, and buying prepaid credits does not raise it.

I want to be straight about one thing, because this category is full of people claiming to be first. We are not the only platform an agent can reach. Resend and Brevo both ship MCP servers of their own. What is different here is that the agent is the primary interface rather than a feature bolted onto a dashboard, and that every rule above sits on the send path rather than in a help centre article you were meant to find.

Who this is actually for

Deliverability gets you into the inbox. It does not make the message worth opening, and that is where AI-sent email usually falls over. Machine-written mail tends to look machine-written, and a recipient who can tell is one click from marking you as spam, which feeds straight back into the four levers above.

Our Brand Kit holds your real colours, fonts, logo and voice, down to the hex values and the typeface, and every campaign an agent drafts renders against it. One prompt produces an email that is written, designed on brand, and sent, rather than a template you then spend an afternoon fixing. One honest limitation, because you will hit it: the most common objection we get in demos is design quality, and the usual cause is that an agent will not pull your images in unless you tell it to. Say so in the prompt, or put the assets in the Brand Kit, and it goes away.

This page is not the right answer for everyone, and pretending otherwise would be the same overreach the guides make. If you have an email team, a deliverability consultant on retainer and someone whose actual job is list hygiene, the chores model works. Keep it, and pick the deliverability tools that fit your stack.

Nitrosend is for the other case. You are busy, you already work inside an AI assistant, you are building automations rather than opening dashboards, and you want email to be something you approve rather than a practice you maintain. If that is you, the twenty-task checklist was never going to get done, and the fix is not a better checklist. It is a send path that will not let you break the rules in the first place.

Go deeper

Common questions