Blog · Cold Outbound · October 20, 2026 · By Ahmad Zabir

What Is a Good Cold Email Reply Rate? Most Replies Are Robots

Stop judging raw reply rate. Count positive human replies per 1,000 delivered contacts, where 2.5 is good and 4.5 is strong, and measure yours this week with the audit we ran on our own campaigns.

In August our own five campaigns produced about 3 positive replies and 0 meetings from 18,524 emails. The counts point at our list and our inboxes rather than the copy, and we have not re-run the audit since. That failure is why we stopped judging campaigns on reply rate.

Of the 73 reply threads, about 77% were robots (automatic replies nobody chose to write). Below is the bar we now use, a five step audit you can run on your own campaign this week, and the evidence behind both.

A good campaign clears 2.5 positive human replies per 1,000 delivered contacts

A positive is a person who says yes in some form: interested, asks about the offer, asks for a time, or names the right buyer. Delivered contacts are the people you emailed minus the ones that bounced. Counting people rather than emails means a longer sequence cannot inflate the number.

Here is where the bar comes from. Sales.co publishes two positive figures from one dataset of 1,369,823 contacts: 0.25% of contacts month by month and 0.64% overall. The page does not say whether the overall figure counts people or reply messages, so we anchor on the month by month figure, 2.5 per 1,000, the closest public match to our unit. It counts per contact and we count per delivered contact, so the two are close, not identical. Its months run from 1.8 to 7.2 per 1,000, and the middle of that range, about 4.5, is our strong level.

If your own campaigns sit well above 2.5 across 30 or more positives each, raise the bar for your market. It is a rule of thumb, not a law.

Our two client samples sit between the good and strong levels, with the caveats in the table. The other large measurement, Saleshandy's 0.11% of delivered emails for email only, is per email, so it reads lower than a per contact figure would.

Our own campaigns scored 0.72, from about 3 positives. That is too few to judge on: the 95% range runs from about 0.25 to 2.1. The Gmail split further down, on 2,250 prospects, is the stronger evidence.

The bar needs volume. At 2.5, collecting 30 positives takes about 12,000 delivered contacts, and at 4.5 about 6,700. Thirty is where the range around a rate narrows to about plus or minus a third, which is why we use it as the minimum.

The bar also turns into money. Cost per positive is your cost per 1,000 delivered contacts divided by positives per 1,000. As a model on our own public price, as of October 2026: our cold email package starts from $3,000 a month for 15,000 to 20,000 people, which is $150 to $200 per 1,000. At 2.5 that is $60 to $80 per positive, and at 4.5 about $33 to $44. The model counts the package price only, assumes every address is delivered and leaves out your own team's time. Our cold email budget breakdown goes line by line.

We judge on positives rather than meetings because in any one campaign meetings can only come from positives. There are fewer of them, and they arrive later. Meetings take over once you have 30 or more.

Measure yours this week in five steps

You need your sending tool's export and a spreadsheet. Step 2 is the slow one, because every thread gets read.

1. Export every reply thread from your last full campaign, plus the number of contacts emailed and the number that bounced. Work from the threads, because the dashboard total mixes people and robots.
2. Tag each thread. Robots first, using the three checks in the section on what counts as a reply, then human positive, neutral, negative or hostile. A thread that opens "Out of office until Monday, limited access to email" stops at the second check: automated.
3. Find each prospect's mail provider from their domain's MX record (the public entry that says which company handles its email). Free MX lookup websites show it, and you only need to look up each unique domain once. Count away notices per provider. A provider that sends you none is probably filing you in spam, and the seed test further down checks it.
4. Divide human positives by delivered contacts and multiply by 1,000. Write the count beside it and read it against the table above.
5. Change the copy only after steps 3 and 4 look clean, and only on tests with 30 or more positives per version.

Published reply rates sit about 10x apart because they count different things

Search the question and you get a confident tier list: 1% to 5% average, 5% to 10% good, above 10% excellent. Almost none of those pages says what a reply is or what it was divided by. We read the primary pages behind the most quoted numbers on 20 Oct 2026, and of eight benchmark pages only five said what they divide by.

The spread comes from three choices.

The first is whether robots count. Belkins counted 34,393 replies on 7,530,489 emails sent in 2025, "excluding auto-replies and bounce notifications", and got 0.45%. Most platform averages say nothing about robots.

The second is the denominator (the number you divide by). Smartlead divides by contacts and puts its median sender, the one in the middle, at 0.74% across more than 850 million emails. Instantly divides all replies, follow ups included, by total emails sent and reports a 3.43% average. A person who gets four emails counts once per contact and four times per email, so the same campaign can read about four times higher per contact. Belkins itself warns that its new rates read lower than earlier editions because its 2023 study divided by emails opened: "the denominator is larger, not the performance worse". Opens are no fallback either, since Apple Mail began loading email images automatically in 2021 and an open stopped meaning a person looked.

The third is the tool's own definition. We found four live definitions on Instantly's pages alone. The newest, a September 2026 AI sales agent benchmark page, counts human replies only and warns that AI agents "classify auto-responses as human replies".

Our own five campaigns honestly report five different reply rates

The same effect shows up on our own data. Every row below describes the same five campaigns from the August audit.

The top row is about 25 times the bottom one, and both are correct. Before you compare your number with anyone's, ask what they divided by.

We read 73 of our own reply threads and about 77% were robots

Our campaigns passed the standard setup checklist: plain text, open and click tracking off, 3 emails per inbox per day from 87 inboxes on 18 domains, and SPF, DKIM and DMARC all valid (the records that prove mail comes from your domain). A checklist tests setup, not reputation, and reputation is what Gmail judged us on.

The dashboard said 75 replies. We opened the 73 threads one by one and tagged each.

That is 25 raw replies for every positive. On 73 threads the 77% has a 95% range of about 66% to 85%. That range covers sampling luck only. The tagging was one careful pass by hand that nobody re-checked, so tagging mistakes are not in it.

We never tested new copy against the same list, so we cannot rule the copy out. The counts point first at the list and the inbox, because a robot reply needs an away notice or a dead address, and a stale list has plenty of both. Spam complaints are replies too: our 3 explicit spam or legal complaints match our positives in number, and on a raw dashboard all six look the same.

Sales.co is the only public source we found that publishes this split. Divide by human replies only and the positive share rises: about 18% of our roughly 17 human threads, and about 26% of Sales.co's (our arithmetic).

Our clients show the same shape at a healthier level. Classic Programmers ran 44 campaigns, 470,293 emails to 128,798 unique leads: a 2.12% reply rate as the sending platform reports it, with 16.6% of replies positive. Fermisoft's email test reached 4,862 companies across three markets: a 3.56% reply rate, with 11.6% of replies positive. Neither has robots separated. Both are on our case studies page. One naming note: the 47.4% we quote for Classic Programmers' best campaign is a positive share (positives divided by all replies), not a rate per contact.

A reply is a message a person chose to write

If a person did not choose to write it, it is not a reply. Every thread goes through three checks in order. Stop at the first yes and tag it automated.

1. It says it is automatic. It comes from a mailer daemon, postmaster or no reply address, or it carries an Auto-Submitted header set to anything other than "no". That header is a hidden label in the email's technical details, defined in RFC 3834, the 2004 internet standard for automatic replies.
2. The first two lines announce an absence or a dead address: out of office, on leave, automatic reply, no longer with the company, mailbox not monitored, changed domain.
3. It would read the same whatever you had sent. Ticket receipts, "we received your message", anything that arrived within seconds.

If your export has no headers, checks 2 and 3 still work on the text alone.

Three edge cases cause most arguments. A robot that names a successor stays automated, but log the successor as a new lead. A one word "thanks" is human and neutral. An unsubscribe is a human negative, suppressed the same day.

If your tool labels replies with AI, hand check a sample first. Instantly's own page says those agents count auto responses as human replies.

Under 30 positives, luck and a better subject line look the same

Say you send 500 emails and get 10 replies, a 2% reply rate. Ten is a small count, and luck alone could move it a few replies either way. The true rate behind your 2% could sit anywhere from 1.1% to 3.6%, which is the 95% range (where the true rate almost certainly sits).

That range depends on how many replies or positives you counted, not on how many emails you sent. A quick rule: divide 1.96 by the square root of the count. With 10, the true rate could be about 62% higher or lower. With 30 it is about 36%, roughly a third, and that is where we start trusting a number.

The rule applies to A/B tests too. If version A gets 10 replies from 500 sends and version B gets 15, B looks 50% better, and the gap is still well within luck. Do not call a winner until each version has 30 or more positives.

Missing out of office replies showed Gmail was probably hiding our mail

On 9 Oct 2026 we split our replies by each prospect's mail provider. Google hosted means the prospect's company email runs on Google; Microsoft hosted means it runs on Microsoft 365. Same main inboxes, same campaigns, two kinds of recipient.

Microsoft hosted prospects, 1,318 people, gave 1.14% human replies and 2.66% away notices. Google hosted prospects, 2,250 people, gave 3 human replies (0.13%) and no away notices at all.

Gmail's help page for the vacation responder, written for the person going away, says: "Emails sent to your spam folder or addressed to a mailing list you subscribe to don't get your reply." Your prospect's spam folder is "your spam folder" in that sentence. Zero away notices from 2,250 people probably means the mail sat in spam.

The obvious objection is that Google hosted people set fewer away notices. So we checked a control (a comparison group): our 5 Gmail inboxes, a different set of senders, writing to 839 Google hosted prospects. They got 3 human replies (0.36%) and 9 away notices (1.07%), so Google hosted people do send them, to senders Gmail trusts.

At 1.07%, about 24 away notices were expected from the 2,250. The control rests on only 9, so its rate is uncertain. Even at the low end of its range, about 0.57%, about 13 were expected, and seeing none by luck is 0.9943 multiplied by itself 2,250 times, about 1 in 350,000 (our arithmetic). The weak point is comparability, not luck: we have not checked whether the two lists match on company size or industry.

The probable cause was ours. Warmup (sending small volumes between friendly inboxes so providers learn to trust a new address) began 4 May, and cold sending began 6 May, two days later. In the warmup reports of our sending tool, Smartlead, our shared IP sending provider (an email host whose servers many customers share) put 34% to 44% of warmup mail in spam, while our 5 Gmail inboxes scored 95% to 99%. Our client builds warm for the whole first month before any cold email. On our own campaigns we skipped that.

Run a seed test before you trust any fix

A seed test is how you check placement directly instead of inferring it. We have not run ours yet, so our cause stays probable until we do.

1. Set up a few test mailboxes you control at each provider your prospects use, for example three on Google and three on Microsoft 365.
2. From each sending inbox, send your real cold email to every test mailbox, spread across a normal sending day.
3. Record inbox or spam for every message, by sending inbox and by provider.
4. If one provider puts most of your test mail in spam, stop cold sending to its prospects from those inboxes, keep the inboxes warming, and send those prospects from inboxes that land. If both providers land in the inbox, placement is fine and the problem is the list or the offer.

Whether our own domains recover at Google, we do not know yet. None of the benchmark pages we read split replies by the recipient's provider, so this is a test you have to run yourself.

Meetings per 1,000 contacted is the audit, and ours was zero

Positives tell you each week whether to change something. Meetings per 1,000 contacted tell you, once you have 30 or more, whether the whole system works.

• 344 emails sent per booked meeting for the average rep in Gong's sample of 28M+ emails (meeting definition and window not stated)
• 0 meetings our five campaigns booked from about 4,225 contacts, Aug 2026 audit
• 0.9 meetings per 1,000 contacts: the highest rate still consistent with our zero (95% upper bound)

Source: Gong blog with Outbound Squad, Jul 2025, updated May 2026; Contwre audit of our own campaigns, Aug 2026

Per email, our upper bound is 0.21 meetings per 1,000, against Gong's 2.9. Gong's figure is for an average rep rather than a campaign, so read the gap as far below any public figure, not as precise odds.

We have no client meeting counts under this test to publish. Classic Programmers' sending platform mislabels meetings, so its log does not match our own count, and we have none for Fermisoft.

Count a meeting only when three things are true: the call happened, the person who can approve the purchase was on it, and it ended with a dated next step. Ask anyone selling you meetings to report against that test. If you are weighing agencies that sell meetings, our pay per meeting versus retainer breakdown covers what else to ask.

Questions people ask

What is a good cold email reply rate?

On raw replies there is no single good number, because sources divide by different things. Figures that remove robots or report the median sit under 1%: Belkins reads 0.45% per email sent on 7.5 million emails, and Smartlead's median sender reads 0.74% per contact. Judge on positive human replies per 1,000 delivered contacts instead, where our rule of thumb puts good at 2.5.

What is the average cold email response rate?

Instantly reports 3.43% (all replies over emails sent, 2025 data), Saleshandy 4.5% (over delivered), Sales.co 2.09% per contact and Belkins 0.45% per email with robots removed. They are not disagreeing about performance. They are counting different things, so check what each one divides by before you compare.

Does send me your proposal count as a positive reply?

Yes. It asks about the offer, so it is a human positive. Let's talk next quarter is a human neutral: log it with a date to follow up, but keep it out of your positive count until the person agrees to a time.

How long should you wait before counting replies?

Pick one window and keep it, for example 30 days after the last email, and never compare a fresh campaign with a finished one. Late replies matter: Sales.co's month by month human reply rate falls from 2.36% for its April 2025 group to 0.45% for its unfinished September 2026 group, which may partly reflect the newest contacts having less time to answer.

How do you tag replies at volume?

Tag every thread by hand until the three checks feel routine, then check a random sample each week against your tool's labels. A random 50 threads pins a robot share near 77% to roughly 63% to 86%, which is enough to catch a tool that counts robots as people. Draw the sample from the whole export, not from the first page.

How long should you warm up a new sending domain?

Our client builds never send from the main domain, warm for the whole first month before any cold email, then ramp to about 1,500 emails a day in month 2 and about 2,500 a day from month 3, which is about 20 per inbox per day. Our own campaigns started cold sending two days after warmup began, and Google probably filed them as spam.

What is a positive reply worth?

Work it back from a deal. Positives per deal is 1 divided by (the share of positives that become held meetings times the share of meetings that close). With made up rates of one in two and one in four, a deal needs 8 positives, and at $60 to $80 a positive that is $480 to $640 of package cost behind a $30,000 deal. Swap in your own rates, because these are not measured.

Before you touch another subject line, tag your last 50 reply threads. If most were written by nobody, fix the list and the inbox first. No rewrite reaches a person who never sees the email.

When you have your count, talk to us about what it says.

Book a call with Contwre · More from the blog