What a positive reply rate actually measures

Positive replies divided by total replies. That is the formula wherever a vendor writes one down, and it is worth having in your head before you compare your number to anybody's benchmark, because the metric sitting next to it on the same dashboard is not divided by the same thing.

Plain reply rate runs on four different denominators depending on who published it. Emails sent. Emails delivered. Total replies. Unique opens. All four are live right now on vendor pages, which is most of the reason two published cold email benchmarks can sit a factor of seven apart while describing the same activity.

The five-minute version, if you only do one thing with this: open your own sending tool's help documentation, search it for the reply-rate formula, and find out which of those four it divides by. Then you will know whether the benchmark you read this morning is a comparison or a coincidence.

The formula, and who publishes it

Instantly states it directly on its own blog: "Positive reply rate = positive replies ÷ total replies × 100." Smartlead's help centre gives the identical formula for the metric on its dashboard. Nobody found in this research disputes that shape.

The agreement extends to what "positive" means, roughly. Apollo defines it by exclusion, and the exclusion is the useful part. A total reply rate, it says, "includes out-of-office messages, unsubscribe requests, and negative responses", and none of those advance pipeline. lemlist counts a reply as positive if it continues the conversation, asks a question, or requests the resource you offered, and explicitly excludes out-of-office, unsubscribes and "not interested." The glossary at learn.thepod.fm calls it a reply "expressing genuine interest," as against any-reply rates, "which include declines, unsubscribes, and out-of-office responses."

One dissent on the denominator even here. thepod.fm's formula divides interested replies by delivered messages rather than by replies, which produces a number several times smaller from the same inbox. Two vendors say replies over replies. One glossary says replies over delivered. Google's AI Overview for this exact query manages both at once, defining positive reply rate as interested responses "divided by total delivered emails or total replies", read on 4 September 2026.

Four denominators, all live

This is the table to keep. Every row read at source on 4 September 2026.

Two names on that list belong to companies that measure their customers against a metric they never define in public. HubSpot publishes no locatable definition of reply rate or positive reply rate anywhere: no glossary entry, no blog post, no knowledge-base article, after direct search. Reply.io, whose product is named after the thing, is the same. Neither is blocked or paywalled; no search or fetch this pass located one.

And one vendor's two pages do not agree with each other. Instantly's 2026 benchmark report defines reply rate as all replies received, including follow-up responses, divided by total emails sent. Unfiltered, sent-denominated.

A separate Instantly blog post on calculating reply rate defines it as unique human replies divided by delivered emails, and instructs the reader to "exclude non-human replies: Remove auto-replies, OOO, bounces, and system notices."

Both pages are Instantly's and both are live. The point is not that anyone was careless. The point is that if one company's own two pages can define a reply rate two ways, a number you read on a third party's blog carries no information unless the definition travels with it.

What the 7.6x gap between the two big numbers is, and is not

Two figures get quoted against each other constantly. Instantly's 2026 benchmark report puts average reply rate at 3.43%, over the period 1 January to 18 December 2025, across what it describes as billions of interactions and thousands of active workspaces. Belkins puts it at 0.45%, and publishes the arithmetic: 34,393 replies from 7,530,489 emails sent, January to December 2025.

That is a factor of 7.6 on the same activity in the same year. Here is what it is not.

It is not the period. Both cover calendar 2025, both stated at source.

It is not sent against delivered. Both headline figures use emails sent as the denominator, verified verbatim on both pages. This is the explanation everyone reaches for first and on this specific pair it is wrong, which is worth knowing precisely because the difference is real elsewhere: thepod.fm's glossary and Instantly's own calculation guide both use delivered for the same metric name.

It is not positive against all. Neither number is labelled a positive reply rate by its publisher. Both are all-reply rates.

What is left is two things, and neither can be measured from what either company publishes. Belkins strips auto-replies and bounce notifications out of its reply count and says so. Instantly's benchmark-report definition does not say either way. And the populations are not comparable: Instantly is averaging a self-selected mass of its own customers, a population its own report describes as running from an unstated floor up through a top quartile at 5.5% and top performers above 10%, while Belkins is reporting one agency's managed book.

Both are real candidates. Neither is quantifiable, because no source states what share of its own counted replies are auto-generated. That number does not exist in public for either vendor, so this article does not estimate it, and any page that assigns percentage points of that gap to a cause has invented them.

Why the auto-reply question is not a technicality

One company publishes the missing shape, on its own population. Sales.co classified 61,770 replies out of more than two million emails sent between 2024 and early 2026.

Auto-replies were the largest single category, by a wide margin over anything a human wrote on purpose. This is Sales.co's own book and it cannot be transplanted onto Instantly's or Belkins' numbers. It does establish that the mechanism is big enough to swallow a benchmark whole, which is why "does this include out-of-office" is the first question to ask about any reply rate and not the last.

Smartlead reaches the same conclusion from its own platform data and says it more bluntly: out-of-office responses are "the most common false positive in reply rate reporting."

Note also what 14.1% positive against 29.9% negative does to the intuition behind this metric. In that set, a reply was more than twice as likely to be a rejection as an expression of interest. A raw reply rate counts those identically.

How to read any benchmark in thirty seconds

Four questions, in this order, and if the page cannot answer the first two the number is decoration.

What is the denominator: sent, delivered, opens, or replies. Are auto-replies and out-of-office in the numerator or not. What is the countable sample size, where "billions of interactions" is not one and "7,530,489 emails" is. And what population produced it, because one agency's managed campaigns and a self-serve tool's entire customer base are different subjects.

Run those against your own dashboard before you compare. If your tool divides by unique opens and the benchmark divides by sends, you are looking at two numbers that share a name and nothing else.

One provenance note, because it shows how a number travels. Woodpecker publishes the same 3.43% and is straightforward about where it comes from: "The platform-wide average cold email response rate in 2026 is 3.43%, based on Instantly's benchmark analysis of billions of cold email interactions," with the figure linked to Instantly's report. One vendor's average of its own customers becomes another vendor's platform-wide average, correctly cited and one step further from the thing it measured. Two more republications without the link and it is a fact about the industry.

The one number worth tracking instead

If you only keep one, keep positive replies over total replies, and label it. It is the only ratio in this whole set that holds still when your bounce rate moves, when open tracking breaks, or when you change how much you send, because both halves come out of the same inbox.

Then accept that it will look small, and that the thing which moves it is not the metric. The largest published effects in this category are about who received the email rather than how the reply was counted, which is the subject of our piece on who actually replies to cold email. The same pattern of vendors measuring different things under one word shows up in how long a cold email should be, where four datasets disagree about the direction of the effect. And whether your denominator is sent or delivered is decided by what your bounce rate is doing, which is worth knowing before you pick one.

Every definition and figure in this article was read at source on 4 September 2026, on the publisher's own domain: Instantly's benchmark report and calculation guide, Belkins' response-rate study, Smartlead's help centre and blog, lemlist's benchmark documentation, Apollo, Gong, Woodpecker, learn.thepod.fm and Sales.co. The AI Overview was read on the same date and is a live search feature rather than a page, so it may read differently by the time you check it. Where a vendor publishes no definition, this article says so rather than supplying one.