How Long Should a Cold Email Be? We Found a Cliff at Microsoft

Most advice on cold email length is about persuasion. We had opinions on that too, until September 2026, when we found out a lot of our emails were never seen at all.

Four of our own campaigns were live and the replies were lower than the lists deserved. Eight inbox placement tests had run overnight. Read as a blended number they looked fine, somewhere between 59% and 79% inbox. Split by mailbox provider, every test of the real copy said the same thing: Google 98% to 100% inbox, Microsoft 0%.

So we paused the campaigns and ran a ladder: the same pitch, one change at a time, to find out what Microsoft was reacting to. It turned out to be length. This post is that data, what it does and does not prove, and the steps we now run on every email before it goes out.

What we tested, and how

An inbox placement test sends your email to a panel of seed mailboxes the testing tool controls, then reports which folder it landed in at each one. We used Instantly's placement testing with a panel of 84 seed inboxes, roughly two thirds Google and one third Microsoft.

The setup for the ladder:

• 42 of our warmed sending mailboxes, across four domains, all the mature ones we had.
• The same 84-seed panel for every body.
• The same subject line on every body, so only the body changed.
• One variable per body. Plain text versus HTML, brand name removed, bullets turned into prose, one idea instead of three, prospecting jargon stripped, and length.
• Each result is counted in sender-and-seed pairs. A first-round body produced about 700 pairs, 210 to 260 of them at Microsoft. Second-round bodies produced about 190, with 61 to 72 at Microsoft.

The full record lives in our internal campaign notes. Every number below comes from it, from our copy rules file, or from the length gate that now runs in our sending code.

The results: the same email at six lengths

Here is every body from both rounds, lined up by word count. Word counts include the greeting and the signature, because that is what the filter sees.

Three things jump out.

Microsoft falls off a cliff. 77 words still landed more than half the time. 84 words landed one time in ten. That is a seven word difference and a 46 point drop. Somewhere between 77 and 89 words, Microsoft stops trusting the email.

Google does not care. The Google column wanders between 81% and 98.5% with no link to length. The two lowest Google results came from bodies where we changed the wording and formatting.

The control repeated. The 173 word live email scored 0 of 225 at Microsoft in round one and 0 of 72 in round two. That repeat is what lets us put both rounds on one axis.

What it was not

The reason we trust the length reading is that the other suspects were tested one at a time and each one failed to move Microsoft.

• Domains, authentication and warm-up. The same four domains delivered the 58 word placeholder to Microsoft 90% of the time, from the same mailboxes in the same hour they sent the long email to spam. Infrastructure that bins one body and delivers another is not the problem.
• HTML. The identical body sent as plain text scored 1 of 228 at Microsoft. No change there. (It did help at Google, more on that below.)
• Our brand and the ask. Removing "Bleed AI" from the sign-off and swapping the call to action scored 0 of 263.
• Bullets. The same sentences written as prose scored 0 of 231 at Microsoft and got worse at Google.
• Spammy vocabulary. This was the leading suspect, and it was wrong. A 153 word version with every "target", "B2B SaaS" and "hiring on LinkedIn" removed scored 3%. A 54 word version that kept all of those words scored 85%.

One smaller effect did survive. At similar lengths, dense, specific commercial copy did worse than vague copy: 84 words of our real pitch scored 9.5% while 77 words of filler scored 55.9%. The lesson we took: spend the word budget on the idea and cut everything around it.

How much to trust this

We want to be straight about the limits, because placement data is easy to over-read.

• It is one ladder, run on one day (2026-09-01), from our mailboxes, to one seed panel, with one subject line.
• The second-round samples are small. 61 to 72 Microsoft pairs per body. The effects are large enough that this does not change the direction (85% against 0% is not noise), but the exact percentages would move on a rerun.
• A seed panel is not your buyer's mailbox. Corporate Microsoft tenants run their own rules on top of Microsoft's. Our own operating notes treat a placement test as a smoke test.
• Other senders can see something different. A different sender ran a real placement test on an email of about 145 words, from their own domains, for a warm-list campaign, and measured 94.4% at Microsoft and 100% at Google. A length rule is a model fitted to our ladder. A direct measurement of your own email beats it every time.

So treat the numbers as a strong default and a reason to test. In our own sending we hold the line, because the cost of being wrong is a campaign nobody sees.

Step 1: find out how Microsoft-heavy your list is

The limit only costs you what your list is made of. A list that is 3% Microsoft can carry a long email. A list that is 50% Microsoft cannot. Here is what we measured on our own lists in September 2026:

Three client lists we measured in September and October 2026 ran between 51% and 57% Microsoft. In our data the young software founders were the most Google-heavy list, and the lists of more established businesses were the most Microsoft-heavy. Do not guess yours from that, measure it.

How to check yours:

1. Pull the MX record for each domain. Any DNS lookup works (nslookup -type=mx domain.com on the command line, or a free MX lookup site for spot checks).
2. Classify it. If the MX host ends in mail.protection.outlook.com, that domain's mail runs on Microsoft 365. If it ends in google.com or googlemail.com, it is Google Workspace.
3. Treat security gateways as unknown. Hosts like pphosted.com (Proofpoint) or mimecast.com sit in front of the real mailbox, which is very often Microsoft. You cannot see behind them.
4. Count unknowns as Microsoft. An unknown that turns out to be Microsoft costs you the whole send. An unknown that turns out to be Google costs you nothing by being short.
5. If you send through Instantly, every lead already carries an esp_code field (1 is Google, 2 is Microsoft). Export your leads and count.

Then pick your ceiling. This is the rule our sending code enforces before any campaign can be created:

Treat these as ceilings. The only lengths we measured landing well at Microsoft were 54 and 58 words. We aim for about 65.

Step 2: count words the way the filter does

Most people undercount their emails. We did, for months. Three habits fixed it.

Count the signature. Most sequencers render the signature inside the body. Ours ("Taha, Founder @ Bleed AI") is 5 words, and it was missing from every count we took before 2 September 2026.

Count the worst case. If you personalise with variables, a single variable can swing a lot. One of our intro lines ranged from 18 to 38 words inside one campaign. If you check the median, the long tail of your list goes to spam. Take the longest value each variable holds on your list. If you use spintax, count the longest option in every group.

Count every step. All our attention went to email 1. Email 2 of the same sequence rendered to about 97 words and had never been counted, so two thirds of the sequence was over the line while we fixed the first third.

Step 3: cut it to fit, with a word budget

The cut is easier if you give every part of the email a budget before you write it. Here is the one we use for a three-idea email:

We set this budget too loose twice before we measured properly, first at 25 to 35 words per idea, then at 14 to 18. Both produced emails our own ladder puts in the spam folder.

Some cuts that keep coming up:

• Drop the throat clearing. "I came across your company and noticed that..." is 8 words of nothing.
• Say the company name once.
• Cut the "because" clause from each idea unless it is the whole point.
• Shorten the call to action. "Want me to send it?" does the job of "Would you be open to me sending over the full write-up I put together?"
• One proof point is enough.

What we would not cut: the line that says what you offer. More on why below.

Before and after: our own email

This is the real email from that campaign, with the prospect's name and company swapped out. It is 173 words with the signature, and it landed 0 of 72 at Microsoft.

And this is the cut. Same offer, same company facts, same jargon, same sign-off. It is 54 words and it landed 52 of 61 at Microsoft (85.2%).

What changed: the long description became one clause, three ideas became one, and the "because" stayed only where it carried the point. Nothing about the vocabulary changed, which is exactly why this pair killed the jargon theory for us.

If you need all three ideas, it can still be done inside the budget. Our rewrite kept three ideas at 8 to 12 words each, and it reads tighter than the original.

Does cutting cost you replies?

This was the worry we heard most from our own team, so we checked it against every email 1 we have sent across six sending workspaces.

• 24 clean same-time A/B tests where a shorter and a longer email 1 went out in the same week to an even split: about 37,000 people in total. The shorter variants replied at 0.62%, the longer at 0.66%. Statistically no difference (p = 0.63). Shorter won 11, longer won 9, 4 were ties.
• Very short questions bring the wrong replies. On one list, a 27 word bare question got 2.5 times the reply rate of the longer email, and zero of its 9 repliers said yes. Most were "we don't use that". The longer email's two good replies were referrals to the right person.
• A personal detail mattered more than length. In one same-offer, same-week pair, a short email with a real personal number replied at 0.67% and a short email with nothing personal at 0.23%.

So the rule we work to now: stay under the ceiling, and inside it write 45 to 70 words that name what you do, carry one personal detail and ask one question. Do not cut all the way down to a bare question with no offer in it.

One more free win: send plain text

The plain text test did nothing at Microsoft, but it was the only change in round one that clearly beat the control at Google: 92.1% to 98.4%, on the identical body. Turn HTML off in your sequencer for cold steps. It costs nothing and it is one of the few placement changes we have seen that is both real and free.

The checklist we run on every email

We moved the length rule into code after writing it in a doc and watching the next batch come out at 86 to 97 words anyway. If you run your own sending, a spreadsheet column with the worst-case word count of each step does most of the same job.

If you would rather have someone else hold the line on this, it is part of how we run outbound for clients: the list is checked for Microsoft exposure, every step is gated on worst-case length, and placement is read per provider before anything scales. You can also see it working on your own market first with a free two week trial.