DELIVERABILITY

Cold email deliverability, properly

Last updated 11 September 2026

Read this before asking us to diagnose an inbox. It explains how we work out whether a problem is the domain, the copy, the tenant, the infrastructure or the inbox type - and how to avoid creating the problem in the first place.

We provision the infrastructure and get the authentication right before you send. Everything after that is sending practice, and this is what we have learned works.

The ten rules

Follow these and most deliverability problems never start. If results still drop afterwards, you will have the structure needed to work out what actually went wrong.

Reply rate matters, but economics decide it

Reply rate is the metric everyone watches and it is not the one that decides whether cold email works. The real question is what you pay to acquire a customer and what that customer is worth.

A campaign at a 0.5% reply rate on high volume can be the cheapest acquisition channel in the business. A campaign at 3% that costs three times as much to run can be worse. Both facts are uncomfortable if you are optimising reply rate in isolation.

Two times the reply rate at three times the cost is not a win

Google Workspace inboxes often produce better reply rates than Microsoft 365 when the recipient is on Google - and they cost several times more per inbox. Sending 2,000 a day at twenty-five per inbox needs around eighty Google inboxes, before data and sequencing. The same volume on Microsoft 365 at three per inbox is more mailboxes but dramatically less money.

Which is right depends on your market size. With a small addressable list, go where reply rates are strongest. With a large one, cheaper volume usually wins on total output even at a lower rate per send.

Match the method to the market

A market of twenty thousand companies and a market of ten million should not be run the same way. Pick the approach based on how many real prospects exist and how much you need to protect the list.

Small market, or client work

Around twenty thousand targetable companies. Your job is to protect the list and the domains while testing new reasons for people to reply.

• One to two cold emails per inbox per day, with five to six warmup messages alongside.
• Consider two sending groups alternating weekly so domains get cooldown.
• At least three structural variants, heavy spintax, new variants constantly.
• Do not go past step two in a sequence - most spam risk sits in the extra follow-ups.
• Twenty thousand prospects at two touches is forty thousand sends, so you can loop the list roughly monthly. Spend that month building a genuinely new angle.

Mid-sized market

Roughly 150,000 to 200,000 prospects. Aim to work through the market about once a month.

• Two to three cold emails per inbox per day.
• For a 100,000-prospect market, build toward 4,000-6,000 sends a day so you can loop monthly.
• That is around twenty orders of a hundred inboxes.
• Group rotation is usually unnecessary here unless deliverability has been poor.
• New variants every week, still no further than step two.

Large market, volume economics

Hundreds of thousands of prospects or more - recruiting, financial services, insurance, broad productised offers.

• Three cold emails per inbox per day, scaled by adding orders rather than pushing inboxes.
• Often a single touch, looping the list every one to three months.
• Domains will burn. It may not be worth replacing one until it is below a 0.5% reply rate over a month.
• Add variants daily, and make that one person's named responsibility rather than everyone's occasional task.
• Mix registrars and TLDs rather than running an identical footprint everywhere.

The practices that prevent problems

Stop writing emails that look like cold emails

Most cold email reads exactly like cold email, which makes it trivial for both filters and recipients to classify. Getting your message across without sounding templated is most of the work.

A real variant changes the shape - different opening, different length, different paragraph order, different CTA and sign-off. It changes the angle, framing the same offer as a different reason for making contact. Rewording a sentence and calling it a new variant leaves the same fingerprint in place.

Heavy spintax, on top of real variants

Three structural variants each carrying deep spintax means ten thousand sends are not ten thousand copies of three emails. Fifteen spintax groups with four options each is over a billion possible versions of a single line.

It is still not enough on its own. Copy fatigue moves faster than it used to, so keep adding new structural variants weekly rather than treating spintax as a substitute for them.

Send less per inbox

One order is a hundred Microsoft 365 inboxes on one domain. At three cold emails per inbox per day that is 300 sends a day from a single order, and that rate is deliberately conservative - Microsoft rewards restraint more than Google Workspace does, where twenty-five to thirty per inbox is normal.

Be realistic about the offer. Most senders overestimate how distinctive theirs is, and the market has usually seen something similar recently. Starting low means you find that out before generating spam complaints.

Warmup is provider-specific

The composition of your warmup pool matters more than its size. If eighty percent of your prospects are on Google but only twenty percent of your warmup traffic reaches Google inboxes, your effective Google warmup is twenty percent - not the number you think you are running.

Match the provider mix of the warmup pool to the provider mix of the people you are actually emailing. If you cannot, accept the lower effective warmup and send less.

Test redirects rather than assuming

Redirects on sending domains are argued both ways. Settle it with your own data: with ten domains, run five with redirects and five without. If the redirect group degrades first, you have isolated a real pattern rather than inherited an opinion.

Engagement keeps domains alive

Engagement is the strongest signal you control. When recipients dislike an offer, or have seen it many times, they ignore or report it, and domains burn faster. The clearest evidence is local-language markets, where senders using otherwise questionable practice keep domains healthy because people genuinely reply.

Lead with something real rather than a call request. A weak CTA - "can we hop on a call", "is this of interest" - gets ignored. A specific, genuinely useful resource earns replies, and replies protect infrastructure.

• Weak: "I have a plan for how you could run Facebook ads."
• Better: "We collected what the top twenty manufacturing firms are doing to improve retention. Want me to send it over?"
• Better: offer two free ad creatives so they can see the output before committing to anything.
• Better: send a short list of ranking gaps rather than asking for a call first.

Spam words are overstated, but not irrelevant

Word choice is rarely the sole reason a message lands in spam, and it does make some variants flag faster - particularly when the email already reads like a pitch. Run copy through a checker and remove the obvious offenders before scaling a variant.

The better test is reading it aloud. If it sounds like a sales email, it will be treated like one. The difficult part is sounding like normal one-to-one correspondence while still only attracting replies from people who actually want what you sell.

Diagnosing a reply-rate drop

A drop only becomes useful once you know which variable it follows. The things you are separating are domain reputation, campaign copy or offer, tenant quality, infrastructure pattern and inbox type.

You only get that signal when the same domain runs across several campaigns and each campaign draws from a mixed pool of domains. If every domain sits in exactly one campaign, a drop is unattributable - you cannot tell whether the domain burned or the copy did.

• Build a pool large enough that each campaign contains inboxes from several domains.
• Run genuinely different copy, offers or segments across separate campaigns.
• Spread each domain across multiple campaigns, and keep some inboxes as untouched holdouts.
• Read the shape of the drop rather than the size of it.

The point is overlap without total dependency. A domain should appear in more than one campaign but not necessarily every campaign, which gives you comparison data while keeping part of the pool clear of a risky campaign. A hundred inboxes on a domain is enough supply to do this properly.

Different problem, different fix

This is why isolating the cause comes first: the remedy changes completely depending on what broke, and rebuilding everything is usually the wrong response.

Frequently asked questions