Cold email deliverability: what actually decides placement

Cold email deliverability is decided in three gates. Authentication, meaning SPF, DKIM and DMARC, gets you considered at all; since 2024 at Google and Yahoo and since May 2025 at Microsoft, mail that fails it is rejected, not filtered. Reputation, built from bounce and complaint rates, decides whether the mailbox provider trusts the domain. Engagement, meaning replies and opens against deletions and spam reports, decides where an individual message lands. Generated content makes the third gate harder, because the constraint that used to keep lists small was the effort of writing, and removing it removes the thing that kept them relevant.

If your reply rate is falling and your bounce rate is climbing, the first question is not what to change in the sending setup. It is whether you are sending to people who have any reason to hear from you.

Key takeaways

• Google, Yahoo and Microsoft all now require SPF, DKIM and DMARC for anyone sending 5,000 or more messages a day. Microsoft rejects non-compliant mail outright with a 550 5.7.515 error.
• The spam complaint rate that matters is 0.3%. Google says to stay under 0.1% and never reach 0.3%. At 0.3% a domain gets blocked regardless of authentication.
• A mailbox’s technical limit and its safe limit are far apart. A Google Workspace user may send 2,000 messages a day; a cold mailbox that sends more than about 50 is asking to be throttled.
• Generated volume removes the natural brake on list size. That, not the writing, is the risk.
• Past a certain point the fix is sending less, not sending differently. This article says where that point is.

Who this applies to

You are sending outreach to people who did not ask for it, at some scale, and probably using generation to write or personalise it. Applies to sales sequences, partnership outreach and recruitment. The same mechanics govern lifecycle and marketing mail, with the difference that those recipients opted in, which changes the third gate entirely.

What this does not cover: getting unwanted mail past filters. That is not a deliverability problem, it is a consent problem, and the techniques for it stop working within weeks and damage the sending domain permanently.

This is not legal advice. The legal frame is stated below because it constrains the design; take advice on your own situation.

The three gates

Gate one: authentication, which is now mandatory

Three DNS records. Get them right once.

SPF lists which servers may send for your domain, as a TXT record naming your sending services. Watch the ten-DNS-lookup limit: stacking several providers silently breaks it, and it fails without an obvious symptom.

DKIM signs each message so the receiver can verify it was not altered and came from your domain. Your sending platform provides the keys; publish them. Use a distinct selector per service so you can rotate one without disturbing the others.

DMARC tells receivers what to do when SPF and DKIM fail, and asks them to report. Start at p=none with a reporting address, read the reports for a few weeks to find legitimate senders you forgot about, then move to quarantine and eventually reject.

The rules as they stand, from the providers’ own published requirements:

Two things follow for cold senders specifically. First, cold outreach at typical volumes sits under the 5,000-a-day threshold, so the bulk rules do not technically apply. The baseline rules still do, the reputation system still does, and a cold sender with no DMARC record is signalling that they have not read the requirements. Set all three anyway. Second, the one-click unsubscribe rule is written for subscribed mail. A cold email should still carry a working opt-out, because the alternative is the recipient using the spam button, which costs vastly more.

What authentication does not do is make anyone want your mail. It proves you are who you claim. A verified sender of unwanted mail is a verified problem.

Gate two: reputation, and the thresholds that end it

Mailbox providers score the sending domain and IP over time. The signals are few and the thresholds are published.

Bounce rate is the one cold senders control least and ignore most, because a list assembled from a data vendor or scraped from profiles carries dead addresses at a rate that would be unthinkable on an opt-in list. Verify every address before the first send. A bounce rate above 5% tells the provider you are mailing people you do not know, which is exactly the inference you do not want it to draw.

Complaint rate is the one that ends domains. One report per thousand messages is where Google asks you to stay. Three per thousand and the domain is blocked whatever the DNS says. On a cold list, the people who report are the ones who did not recognise you and could not see why you wrote. Both of those are targeting failures, not copy failures.

Gate three: engagement, which is per message

For each message, the provider weighs recent recipient behaviour toward your domain.

Positive: replies, the strongest signal by far; opens followed by other engagement; moving mail out of spam; adding you to contacts; forwarding.

Negative: marking as spam; deleting unread repeatedly; ignoring over time; bouncing.

Content is not on that list. Subject-line wording and spam-word checkers are a small factor compared with whether the last thousand recipients wanted the mail. A message about a genuine problem the recipient has, sent by a person they can identify, with a reason for writing they can see in the first line, gets replies. Replies are what the third gate rewards. Everything else is decoration.

Safe volume per mailbox

The technical limits and the safe limits are far apart, and confusing them is how new senders burn a domain in a week.

Two rules. Cold outreach goes from a subdomain or a separate domain, so the primary domain’s reputation, which carries your invoices, proposals and support replies, is insulated. And a mailbox that sends cold mail sends it at a rate that looks like a person, because a person is the thing the provider is modelling.

Note the difference between that and what some tooling sells: a separate sending domain is standard practice. Twenty lookalike domains rotated so that reputation damage never catches up with the sender is the design described at the end of this article, and it is not the same thing.

Why generated content makes this harder

This is the reason the article exists.

The brake is gone. Writing genuinely personalised outreach was slow, and that slowness limited list size and forced qualification. Generation removes the cost of producing volume without removing the cost of sending it, which is borne by your domain reputation.

Superficial personalisation is legible. Inserting a company name and a scraped fact produces mail that reads as templated to a recipient who has seen a hundred like it. It looks personalised in a dashboard and generic in an inbox, and the behavioural signals follow the inbox.

Generated variation does not fool reputation systems. Rewording each message defeats naive content matching. Reputation is computed on sending domain, IP and recipient behaviour. Different words to the same unwilling audience produce the same complaint rate.

Volume arrives faster than reputation builds. A new domain sending thousands in week one is the classic pattern providers throttle. Generation makes it possible to reach that pattern on day one.

What actually works

Send to people you can name a reason for. Everything else is mitigation. A list of 200 people each of whom has a visible, specific reason to hear from you outperforms a list of 5,000 matched on a filter, on every metric including absolute reply count.

Warm the domain gradually. A new sending domain starts at tens per day and increases over four to eight weeks, prioritising the recipients most likely to reply. This is not a trick to disguise volume. It is how reputation is established, and it works because early engagement is genuinely good.

Verify addresses before sending, particularly for anything not collected first-hand. Bounce rate above 2% is a problem; above 5% you are being throttled already.

Clean the list continuously. Remove hard bounces immediately. Remove anyone who has not engaged in six months. Painful, and the single most effective intervention available.

Make opting out trivial. One line, one click, honoured immediately. Every opt-out you make difficult becomes a spam complaint instead, and a complaint costs vastly more than a departure.

Watch the numbers where the providers publish them. Google Postmaster Tools shows your spam rate and domain reputation for Gmail recipients. Microsoft’s SNDS shows the same for its network. Register for both before the first send, so the baseline exists.

Measure with a holdout. Attribute replies and revenue against a portion of the list deliberately not mailed. See how to measure whether an AI system works for why every channel overstates itself without one.

Consent, and the legal frame

Jurisdiction dependent, and the practical shape is:

United States, CAN-SPAM. Unsolicited commercial email to business addresses is permitted subject to conditions: accurate headers, non-deceptive subject lines, a physical postal address in the message, a working opt-out honoured within ten business days.

EU and UK, GDPR and PECR. Stricter. B2B outreach to corporate addresses is often argued under legitimate interest, which requires an actual balancing assessment rather than an assertion, plus a clear opt-out and transparency about where the data came from. Outreach to individuals generally needs consent.

Canada, CASL, and Australia, the Spam Act. Consent-based regimes with narrow exceptions and real penalties. Most US-style cold outreach is not lawful there.

We are not lawyers and this is not legal advice. The practical point is that “it is legal” and “it will be delivered” are different questions with different answers, and satisfying the first does nothing for the second.

Where the line is

Stated plainly, because the surrounding industry is vague about it.

Legitimate outreach: you have a defensible reason to believe this specific person has this specific problem, the message says who you are and why you are writing, opting out is trivial and permanent, and you stop when asked.

Not legitimate, regardless of what any tool enables: mailing purchased lists of people who never expressed interest, disguising who is sending, using misleading subject lines, making opt-out difficult, or rotating domains and inboxes so that reputation damage never catches up with the sending.

That last one deserves naming because it is sold as a feature. Infrastructure designed so that consequences land on disposable domains is a system for externalising the cost of unwanted mail onto recipients and mailbox providers. It is not a deliverability strategy, and we will not build it.

Where we land

We build lifecycle email automation and targeted outbound, and we decline volume cold-outreach infrastructure. The distinction is not “cold email is wrong”.

Targeted outbound to a researched list, where each recipient is someone you can articulate a specific reason for contacting, is a legitimate and effective channel. We will help build the qualification, the sequencing and the measurement for it, in the United States only, because it is the one major market whose law is opt-out rather than consent-based.

What we decline is scale that only works by staying ahead of reputation: hundreds of domains, thousands of inboxes, volume calibrated to the point just before consequences arrive. Not primarily on ethics, though that too. It produces a system whose performance decays continuously and requires permanent reinvestment to stand still, which is a poor thing to hand a client.

The recommendation that most often disappoints: when someone arrives with falling reply rates, the answer is usually that the list is too big rather than that the setup is wrong. Cutting a list by 70% to the people you can name a reason for typically improves both placement and absolute reply volume. It is a smaller engagement than the one they came for, and it is what works.

When to send less instead

Complaint rate above 0.1%. Stop and fix the list. No sending change compensates, and at 0.3% the domain is gone.

Bounce rate above 3%. Verify before sending again, and find out where the addresses came from.

Reply rate under 1% on a list you researched yourself. The targeting is wrong, not the copy.

When you cannot name why this person should hear from you. If the answer is that they matched a filter, the outreach will perform accordingly.

Frequently asked questions

Do we need a separate sending domain for cold email?

Yes, a subdomain or a separate domain, so that a reputation problem on outreach never touches the domain your invoices and support replies come from. Twenty lookalike domains rotated to spread damage is a different thing, and it is the pattern described above.

How long does domain warming take?

Four to eight weeks to reach meaningful volume, starting at tens per day per mailbox and prioritising the recipients most likely to reply. Faster ramps get throttled.

Does AI-generated copy get filtered for being AI-generated?

Detection of generated text is unreliable and is not the mechanism. What gets filtered is mail people do not engage with. Generated copy is a problem when it enables volume and superficial personalisation, not because of how it was written.

What spam complaint rate is acceptable?

Under 0.1%. Google’s published guidance is to stay below 0.1% and never reach 0.3%, at which point the domain is blocked regardless of authentication. Register for Postmaster Tools so you can see the number.

Do the bulk sender rules apply to cold email under 5,000 a day?

The bulk-specific rules, such as mandatory one-click unsubscribe, are triggered at 5,000 messages a day. The baseline requirements and the reputation system apply to every sender at every volume. Comply with all of it anyway; it costs an hour in DNS.

Should we buy a list?

No. Beyond the legal exposure in several jurisdictions, purchased lists contain spam traps and recycled addresses, and the reputation damage outlasts any campaign that used them.

Next step

If deliverability is degrading, start by measuring bounce and complaint rates by list segment before changing any sending configuration. The email automation engagement covers lifecycle flows and targeted outbound with holdout measurement, and includes the recommendation to send to fewer people when that is the answer.

Related: How to measure whether an AI system works · Why automations fail silently · How to evaluate an AI agency proposal · Email automation