Most cold email replies do not come from the first email

Why this matters

Cold email programs are built around sequences: an initial message followed by two, three, sometimes five reminders. The entire discipline of follow-up writing exists because of a widespread belief that persistence is where the reply comes from, that the first email is a coin flip and the real work happens in emails two through five. That belief shapes how teams spend their time. It determines whether a sales development rep polishes the opening line for an hour or dashes it off and saves the effort for a clever follow-up.

This study asks a narrower, checkable question: when a reply does arrive, which email in the sequence prompted it? Not whether follow-ups increase total replies, which is a different question requiring a different design, but which link in the chain a reply that already happened is attached to.

The distinction matters because the two questions get confused constantly. A team might see that a campaign generated a hundred replies across a five-email sequence and conclude that follow-ups are pulling their weight, without ever checking that 61.7% of those replies were attached to the first email. If a large majority of replies trace back to the first email, the practical implication is not that follow-ups are worthless (they may still prevent a prospect from forgetting the message, or catch someone who was traveling on day one) but that the first email is where the deliverability, subject line, and opening sentence decisions carry the most weight per hour invested.

Who should care: anyone writing or reviewing cold email sequences, anyone setting rep quotas for follow-up volume versus first-touch quality, and anyone building automated sequencing tools that decide how many follow-ups to schedule. The finding also matters to people skeptical of cold email generally, since it changes what a low reply rate actually diagnoses. A low reply rate with most replies concentrated in the first email suggests the message itself is not landing, not that the sequence is too short.

The rest of this page describes how replies were attributed to a position in the sequence, what the headline distribution looks like, whether it holds up across different kinds of replies and different outreach groups, and what a team should actually change on the strength of it.

How the measurement works

The unit. One unit is one reply. Each reply was traced back to the specific email in its sequence, the message it was sent in direct response to, and labeled with one of 4 classes: first email, follow-up 1, follow-up 2, or follow-up 3 or later. The dataset contains 1,857 such units, drawn from 1,857 rows with 0 rows excluded, spread across 45 outreach groups. A group is a distinct outreach campaign or sequence configuration; the average group contributed 41.27 replies, though as later sections show, group sizes and behavior varied.

Why attribute at the reply level, not the campaign level. A campaign-level count (how many total replies did this five-email sequence generate) cannot answer where in the sequence attention should go, because it collapses five distinct events into one number. Attributing each reply to the specific email that triggered it preserves the information a team actually needs: if someone is deciding whether to spend more time on email one or email three, they need to know how replies distribute across that specific choice, not just a campaign total.

What counts as a reply. Any inbound message a recipient sent back in the thread, regardless of sentiment, from an enthusiastic yes to an unsubscribe request, was captured as one unit and classified by sequence position. Sentiment is tracked separately as a segment (interested, neutral, not interested, unlabelled, unsubscribe or hostile) and covered in a later section. This separation matters: a study that only counted positive replies would answer a different, narrower question, since it is possible that positive replies cluster differently than hostile ones across the sequence.

Classification and its limits. No mitigations were applied to the data (0), meaning the class assignments are taken as recorded, without adjustment for possible mislabeling. The measurement window ran from 2025-12-01 to 2026-08-14. What this design cannot see: it cannot tell you whether a follow-up reply would have arrived anyway without the follow-up being sent, nor whether a different subject line on follow-up two would have pulled more replies. It measures where replies landed in the sequences that were actually run, not what would happen under a different sequence design.

The headline result

Of 1,857 replies in the dataset, the distribution across sequence position was:

The first email alone accounts for a clear majority, and each subsequent step in the sequence contributes less than the one before it: 20.9%, then 12.5%, then 4.9%. The decline is not gradual and flat, it is closer to a steep drop-off after the first message, with a long, thin tail of replies still trickling in from later follow-ups.

What this does mean. Among replies that happened, most trace back to the first email a prospect received. If a team is trying to decide where a marginal hour of editing time produces the most attributable replies, this table says: the first email, by a wide margin.

What this does not mean. It does not mean follow-ups are ineffective at generating replies overall, and it does not mean removing follow-ups would leave the first-email reply count unchanged. It is entirely possible that some replies classed as "first email" arrived after a prospect read a follow-up but chose to reply to the original thread, or that some recipients only opened the first email after a follow-up reminded them it existed, then replied there. The classification captures which email the reply is attached to, not which email caused the recipient to act. Distinguishing attribution from causation would require a design that varies follow-up presence experimentally and tracks reply timing against open events, which this dataset does not do.

It also does not mean every sequence should be shortened. A team running four-email sequences and a team running one-email blasts are not directly compared here, since this dataset describes replies within sequences that were mostly already multi-email. The finding describes where replies land inside existing sequences, not what an optimal sequence length is.

Does the pattern hold across sentiment

Replies were also segmented by sentiment: interested, neutral, not interested, unlabelled, and unsubscribe or hostile. If the first-email concentration were an artifact of, say, easy unsubscribe clicks happening early and inflating the first-email count, the pattern should look different once positive replies are isolated.

The first email leads in every segment. It is lowest among interested replies, at 52.8%, and highest among unsubscribe or hostile replies, at 87.1%. That gap is worth pausing on. It suggests recipients who intend to opt out or push back tend to do so on first contact rather than waiting through several follow-ups, which is plausible: annoyance from an unwanted first email does not usually need three more emails to crystallize into an unsubscribe click.

The interested segment shows the most room for follow-ups: 14.6%, 20.3%, and 12.2% come from follow-up one, two, and three-or-later respectively, together accounting for a larger share than in most other segments. One reading: a genuinely interested prospect may need more time, more context, or a nudge before replying, so later touches do more work for this segment than for others. Note the denominator here, 123 replies, is the smallest of the labeled segments, so single-percentage-point comparisons within it should be read cautiously. The unsubscribe or hostile segment is smaller still, at 31, meaning its follow-up 3 or later figure of 0% rests on very few replies and should not be treated as a precise estimate.

Is this a whole-population pattern or one loud group

A headline percentage computed across 1,857 pooled replies can hide a distribution where one or two outreach groups, perhaps with unusually aggressive first emails or unusually weak follow-up copy, drag the average in one direction while most groups look quite different. Before trusting the pooled number as a description of "how cold email works," it is worth checking group by group.

The dataset contains 45 groups, and the first-email share was computed separately within each of 14 groups with enough data to compute a share. No single group dominates the pooled counts: the largest group contributes 27.9% of all replies, leaving the rest spread across the remaining groups. That rules out the simplest version of the concentration concern, where one outsized group's behavior masquerades as a population-wide pattern.

Spread of the first-email share across groups:

The median group's first-email share, 69.4%, sits below the pooled figure of 61.7%, which tells you the pooled average is being pulled upward by groups with larger reply counts and higher first-email shares, not that a typical group looks exactly like the headline number. Still, even the lowest group in the distribution, at 46.6%, has the first email as a substantial contributor, and the highest reaches 89.5%. No group shows the follow-ups collectively outpacing the first email. The follow-up 2 and follow-up 3-or-later rows both have a minimum of zero, meaning at least one group recorded no replies at all from those later positions, consistent with some sequences being shorter than four emails or simply not generating late replies in that run.

The practical conclusion: the direction of the finding, first email ahead of every individual follow-up, is a property of the population of groups, not an artifact of one or two large or unusual campaigns. The magnitude, exactly how far ahead, varies enough between groups that a single team's own sequence could reasonably land anywhere between the min and max shown here.

What the class labels can and cannot be trusted to mean

This study reports 0 mitigations applied to the underlying data, meaning the class assignments (first email, follow-up 1, follow-up 2, follow-up 3 or later) are used exactly as recorded in the source system, with no correction layer, deduplication pass, or manual relabeling applied afterward. There is no independent calibration exercise reported here, such as a manual audit of a sample of replies to check whether the recorded sequence position matches the actual triggering email.

What that means practically: the numbers in this study are only as reliable as the tracking system that assigned each reply to a sequence position in the first place. Two failure modes are worth naming explicitly, because they would not be visible from the aggregate numbers alone.

Thread misattribution. If a recipient replies to an old message in their inbox rather than the most recent one, some email systems and CRMs record the reply against whichever email the reply-header points to, which may not be the email the recipient was actually reacting to. This would tend to inflate the first email's count somewhat, since first emails are the ones most likely to sit untouched in an inbox for the recipient to stumble back onto later.

Missing follow-up sends. If a sequence's later steps were skipped for some recipients (because the tool detected an out-of-office reply, or a rep manually paused the sequence, or the recipient unsubscribed after the first message), then those recipients structurally cannot generate a follow-up-attributed reply, which would also inflate the first email's share for reasons that have nothing to do with which message is more persuasive.

Neither of these failure modes has been ruled out or quantified in this dataset. Both would bias the finding in the same direction, toward overstating the first email's share, so the true concentration in the first email could be somewhat lower than 61.7% if either mechanism is present at meaningful scale. Because 0 rows were excluded, the study also is not filtering out obviously malformed or duplicate records before analysis, which is consistent with taking the tracking system's labels at face value throughout. A reader who wants tighter confidence in the exact magnitude, as opposed to the direction, of this finding should ask their own sequencing tool how it assigns replies to sequence position before assuming this dataset's pattern transfers exactly to their own numbers.

What to do differently on Monday

The finding does not say to abandon follow-ups. It says the first email carries a disproportionate share of the outcome you are actually measuring when you count replies, and that should change where scrutiny and iteration effort go.

Put review time where the replies are. If a team currently spends roughly equal editing time across four emails in a sequence, the data here suggests that allocation does not match where replies land: 61.7% of replies trace to the first email against 20.9%, 12.5%, and 4.9% for the three follow-ups combined into a shrinking tail. A/B testing subject lines and opening sentences on the first email will touch a larger share of eventual repliers than the same testing effort applied to follow-up three.

Do not read a weak first email's low reply count as evidence follow-ups will rescue it. Because most replies that happen at all come from the first email, a first email that is not landing is unlikely to be compensated for by clever follow-up copy layered on top of it. The follow-ups in this dataset are follow-ups to emails that were, on average, already working reasonably well.

Keep sending follow-ups, but budget them differently. The tail is real, not zero. Follow-up 1 alone still produced 388 replies, which is not a number to dismiss. The recommendation is not to cut sequences to one email, since this dataset cannot tell you what would happen to first-email reply counts if follow-ups were removed entirely (see the honesty note in the headline section on attribution versus causation). The recommendation is to treat follow-ups as a lower-effort, lower-expected-yield-per-message addition rather than an equal partner to the first email in the sequence.

Watch the sentiment split. If the goal is specifically to reduce unsubscribes and hostile replies, note that this segment was the most front-loaded of all, at 87.1% landing on the first email. That argues for investing extra scrutiny in first-email targeting and tone, since a poorly matched first email seems to provoke its negative reaction immediately rather than accumulating slowly across a sequence.

Re-run this analysis on your own sequences before generalizing further. The group-level spread shown earlier, from 46.6% to 89.5%, means your own campaigns could sit anywhere in that range. Treat the pooled figures as a prior, not a guarantee.

How to read the published dataset yourself

The underlying data covers 1,857 replies with 0 rows excluded, organized into 1,857 classified units across 45 outreach groups and 5 sentiment segments. Anyone auditing the finding or trying to reproduce a version of it for their own outreach should know what each layer represents before drawing further conclusions from it.

Start with the class totals. The four-row table of first email, follow-up 1, follow-up 2, and follow-up 3 or later, with counts and shares, is the foundation everything else is checked against. If you recompute this from raw data and get a materially different split, something in your extraction (probably reply-to-thread matching) differs from how this study assigned sequence position.

Check group counts before trusting group-level claims. There are 45 groups total, but the spread table only reports on 14 groups with sufficient data to compute a share for each class. A group with only a handful of replies will produce a noisy percentage, so treat any single group's reported share as an estimate with wide uncertainty unless you also know its raw count.

Cross-reference segments against classes, not against each other in isolation. The segment table answers "among interested replies, where did they land in the sequence," which is a different question from "among first-email replies, what fraction were interested." Both are legitimate questions but they use different denominators, and the fact table above gives you the first framing, not the second. If your own analysis needs the second framing, you will need to recompute it from row-level data rather than reading it off this study directly.

Treat the average group size as a rough guide, not a target. The average of 41.27 replies per group across 45 groups tells you these are moderately sized campaigns, not massive enterprise blasts and not single-digit pilot tests, but individual groups vary, as the largest single group's 27.9% share of total replies shows.

Remember what was not adjusted. With 0 mitigations applied and 0 rows excluded, this is close to a raw tabulation of the tracking system's own labels. That is a reasonable default for a first look, but it also means known failure modes in reply-to-sequence attribution, discussed earlier, have not been screened out. Anyone building on this dataset for a decision with real budget attached should independently sample a handful of raw threads and manually confirm the sequence-position label before treating the fine-grained percentages as exact.