What is a good reply rate for cold email? (2026 benchmarks)

A good cold email reply rate is whichever number your own list produced last quarter, measured the same way twice. No standards body, regulator or peer-reviewed dataset publishes a cold-email reply benchmark, so every figure in circulation is a vendor describing its own customers, with its own definition of a reply and no correction for who chose that vendor in the first place. That is not a standard you can be held to. What you can do is build the instrument honestly: fix the denominator, strip the machine-generated replies, separate positive from negative, and know the width of the error bar before you act on a difference.

Why no honest answer to this question opens with a number

Start with what would have to be true for a public benchmark to exist. Someone would need a defined population of senders, a random sample from it, a shared definition of a reply, a shared denominator, and an audit. None of those exist for cold email, and the first one may not be constructible at all — there is no register of businesses that send cold email from which to draw a sample.

What circulates instead is aggregate data from sending platforms. Each figure describes the subset of teams that chose that platform, completed onboarding, kept paying, and had their data included. Every one of those steps is a filter correlated with performance. A platform whose customers are mostly agencies running broad lists and a platform whose customers are mostly founders mailing 40 people a week will report different numbers for the same word, and neither is wrong about its own population.

Then the definitions diverge. Some platforms count any inbound message on the thread. Some count only messages from the original recipient. Some strip out-of-office replies, some do not. Some divide by messages sent, some by messages accepted. Two platforms reporting "6% reply rate" can be measuring quantities that differ by a factor of two before anyone writes a word of copy.

This is not a reason to stop measuring. It is a reason to stop importing. The rest of this post is the instrument.

What counts as a reply, before you divide anything

A reply is a message you received that a human at the target account composed in response to yours. Every clause in that sentence is a decision you have to implement.

Thread association. The reliable mechanism is header-based. A conforming client sets In-Reply-To to the Message-ID of the message it is answering, and appends that ID to the References chain, both defined in RFC 5322. Matching on those headers is exact. Matching on a subject line that begins Re: is not: subjects get edited, localised clients prepend AW: or SV: or RE: in other scripts, and an unrelated message can collide with a common subject. If your reporting quietly falls back to subject matching when headers are missing, you have two different metrics sharing one label.

Sender identity. A reply from [email protected] on behalf of the person you mailed is a genuine human response and should count. A reply from mailer-daemon@ is not. A reply from a different domain entirely is usually a forward, and whether that counts is a judgement you should write down once rather than re-make per campaign.

Direction. Your own follow-up landing in the same thread is not a reply. This sounds too obvious to state, and it is a real bug in home-made reporting built on top of an IMAP folder.

The four things that get counted as replies and are not

Each of these inflates the number, and each inflates it unevenly across segments, which is worse than inflating it uniformly.

Out-of-office and other automatic responses. RFC 3834 defines the Auto-Submitted header field for exactly this purpose. Its permitted values are no, auto-generated and auto-replied, and the RFC states that automatic responses "SHOULD NOT be issued in response to any message which contains an Auto-Submitted header field ... where that field has any value other than 'no'". So the test is mechanical: any inbound message carrying Auto-Submitted: with a value other than no is not a reply. Microsoft systems layer on X-Auto-Response-Suppress, documented in Microsoft's MS-OXCMAIL protocol specification with values including OOF, DR, RN, NRN and All; its presence on a message is a strong signal you are looking at machine traffic. Neither header is universal, so also treat a message whose envelope sender is the empty reverse path as automatic — that is the path RFC 5321 reserves for notifications, and no human composes from it.

Bounces that arrive as messages. A delivery status notification is a message in your inbox with a subject line and a body. If your reply detection is inbox-based rather than header-based, hard bounces raise your reply rate. That produces the exact inversion of reality: the worse the list, the better the copy appears to be performing.

Challenge-response and spam-filter confirmations. Some receiving systems mail the sender a confirmation request before delivering. These arrive threaded, from a human-looking address, and are pure noise.

Your own aliases. Forwarding, catch-all routing and shared mailboxes can loop a copy of your own send back into the thread. It happens rarely and it is a nightmare to notice, because it inflates by a small constant.

Strip all four before dividing. If you cannot strip them, say so in the report, because a rate that includes auto-replies is not comparable with one that does not — and holiday periods will move it for reasons that have nothing to do with your writing.

Which denominator you divide by changes the answer more than your copy does

Four candidate denominators, in the order they shrink:

• Contacts enrolled. Includes people the sequence never actually mailed because they replied at step one or were suppressed. Almost never what you want.
• Messages sent. Everything handed to the mail server.
• Messages accepted. Sent minus everything that returned a permanent failure. The receiving server issued a 250 at the end of the SMTP transaction, which is defined in RFC 5321 as the server taking responsibility for the message.
• Messages delivered to a human-reachable mailbox. Accepted, minus soft bounces that never cleared. This is unknowable from the sending side with any precision, which is a good reason not to build a metric on it.

Use accepted, and define it explicitly as sent minus hard bounces. The argument is that a hard bounce contains no information about your copy: nobody read it and decided not to answer. Leaving hard bounces in the denominator means a batch with 12% invalid addresses reports a reply rate roughly 12% lower than the same copy sent to a clean list, and you will spend a week rewriting a first line to fix a data problem. The distinction between soft and hard bounces is what makes that subtraction safe to automate.

Whatever you pick, store it as a field next to the number. A rate quoted without its denominator cannot be compared to anything — including the same team's measurement from three months earlier, taken before somebody changed the report.

Positive-reply rate is a separate measurement, not a subset you can eyeball

Total reply rate and positive reply rate answer different questions and can move in opposite directions. Broadening a target list usually raises total replies, because more of the recipients feel obliged to say no. Tightening it usually raises positive replies while lowering the total. A team optimising the blended number will systematically choose the worse list.

Positive-reply rate needs a written classification rule, applied consistently. The rule that survives contact with reality has three buckets and one tiebreak:

• Interested — asks a question, requests a call or materials, forwards you internally, or gives a specific future date.
• Not now — declines with a stated reason or timeframe, and does not ask you to stop.
• Refusal — asks you to stop, expresses annoyance, or reports the message.
• Tiebreak — if two people would classify a reply differently, it goes in the lower bucket. Ambiguity is not interest.

The third bucket is a metric in its own right and the one most teams never build. A campaign whose positive replies rose while refusals rose faster is not a campaign that improved. Autocloz classifies inbound replies in the unified inbox across all five channels so the same taxonomy applies to an email reply, a LinkedIn message and a WhatsApp response rather than three incompatible ones.

How wide the error bars on your reply rate actually are

This is the part that changes behaviour, because the intervals are far wider than people assume. A reply rate is a proportion estimated from a sample, and the appropriate interval for a proportion with a small numerator is the Wilson score interval rather than the textbook normal approximation, which misbehaves badly near zero.

At 95% confidence, for an observed 6% reply rate:

• 12 replies from 200 delivered messages gives an interval of 3.5% to 10.2%. That is 6.7 percentage points wide. Your "6%" is consistent with a true rate under four and with a true rate over ten.
• 30 replies from 500 narrows it to 4.2% to 8.4%.
• 60 replies from 1,000 gives 4.7% to 7.6%.
• 120 replies from 2,000 gives 5.0% to 7.1%, about one point either side.

Run it at the scale most small teams actually operate. Three replies from 40 delivered messages reads as 7.5%, and its 95% interval is 2.6% to 19.9% — a span of more than 17 points. Any decision taken on that number is a decision taken on noise. The interval, not the point estimate, is what you should be quoting internally. The free cold email benchmark calculator does this arithmetic on your own figures without a spreadsheet.

Autocloz's free plan covers 5 users and 10 mailboxes with per-step reply and positive-reply reporting included, so the denominator and the classification are the same across every campaign — start free if you want the number and its interval rather than the number alone.

How many sends a comparison needs before the difference means anything

Testing two subject lines is a comparison of two proportions, and the sample size it needs is unintuitive. At a two-sided 5% significance level with 80% power, using the standard normal approximation for comparing two proportions:

• Separating a true 3% from a true 6% needs about 750 delivered messages per variant.
• Separating a true 4% from a true 6% needs about 1,900 per variant — roughly 3,800 in total.
• Separating a true 5% from a true 6% needs about 8,200 per variant.

That last figure is the one worth sitting with. A one-point difference at these rates is not detectable at any volume a small outbound team will reach inside a quarter. Most subject-line tests in cold email are underpowered by an order of magnitude, which means their results are close to coin flips dressed as findings, and re-running the loser next month will often "win". If you are going to test, test things big enough to see: offer, segment and channel mix change outcomes at magnitudes a test can actually resolve, which is the case for structuring the test around the variables that move rather than around wording.

One further trap. Checking the result repeatedly and stopping when it looks significant inflates the false-positive rate well above the nominal 5%, because each look is another chance to cross the line by luck. Fix the sample size before starting, or use a method designed for repeated looks. Do not fix it afterwards.

Why a blended reply rate can point at the wrong variant

Segment mixing can reverse a result, and this is not a hypothetical. Take two variants, each sent to 1,000 delivered addresses across two segments.

• Segment one, a warm-adjacent list. Variant A gets 30 replies from 300 sends, which is 10.0%. Variant B gets 45 from 600, which is 7.5%.
• Segment two, a cold enterprise list. Variant A gets 14 from 700, which is 2.0%. Variant B gets 6 from 400, which is 1.5%.
• Blended. Variant A gets 44 from 1,000, which is 4.4%. Variant B gets 51 from 1,000, which is 5.1%.

Variant A beat variant B in both segments and lost the aggregate, purely because B was sampled more heavily into the easy segment. This is Simpson's paradox, and in outbound it happens whenever enrolment is not balanced across segments — which is most of the time, because enrolment follows list availability.

The defence is to compare within a segment and never across, and to define segments before the test rather than after. Segments defined after seeing the data are a different and worse problem.

What to do with the number once the instrument is honest

Three uses, in descending order of value.

As a trend against yourself. The same segment, the same denominator, quarter over quarter. This is the only comparison where the confounders mostly cancel, and it is the one that tells you whether the programme is improving.

As a trigger for diagnosis. A reply rate that drops sharply while positive-reply rate holds usually means your list broadened. A reply rate that drops while bounce rate climbs is a data problem. A reply rate that drops while both hold steady is worth checking against placement before copy, because a message filed in spam is a send that no human could answer, and open rate will not tell you — open tracking is a far less reliable instrument than the reply number you now have.

As one input to a channel decision. Reply rate on email and connect rate on the phone are different units and cannot be compared directly; the arithmetic that makes them comparable runs through cost per qualified conversation, not through the raw percentages. Teams evaluating multi-channel platforms against single-channel senders — the sort of comparison on the Instantly comparison page — should be checking whether both channels report into one denominator or two.

What a reply rate does not tell you, and what Autocloz does not measure

A reply rate is a measure of response, not of revenue. It says nothing about deal size, sales cycle or whether the interested replies convert, and a programme can raise it while lowering pipeline by targeting people who enjoy replying. Track it alongside a qualified-opportunity count or it will optimise itself into irrelevance.

It also cannot separate targeting from copy. Both move the same number, and no amount of statistics untangles them from observational data — only a controlled test where one is held fixed does, and that costs the sample sizes above.

Autocloz reports reply and positive-reply rate per step in campaign analytics, applies the classification consistently, and excludes automatic responses from the reply count. It cannot tell you whether a reply was sincere. It cannot detect a reply sent from a personal address on a different domain unless the headers connect it to the thread. It does not publish an industry benchmark to measure you against, because it has no way to construct an unbiased one and would rather show you an interval than a number with false precision. And no measurement here changes the underlying rate — the instrument tells you what happened; the list, the offer and the placement decide it.