Human replies to B2B cold email fell from 0.55% to 0.40% of sends between May and August 2026

Findings

1. Of 736,078 prospect emails sent from May to August 2026, 2.2% drew a reply of any kind (16,090 replies), 0.48% drew a reply from a person (3,497) and 0.18% drew a positive reply (1,298).
2. 68.0% of those 16,090 replies were automated: 10,633 auto-replies and out-of-office notices (66.1%) and 303 bounce notices (1.9%).
3. The human reply rate fell every month, from 0.55% in May (568 replies, 102,525 sends) to 0.40% in August (1,021 replies, 255,675 sends), and the positive rate fell from 0.23% (236 replies) to 0.14% (354).
4. Per contact, the human reply rate was 0.62% in May (551 replying contacts, 88,919 contacts emailed) and 0.45% in August (993 and 222,619).
5. In the 14 workspaces that sent at least 5,000 prospect emails in both May to June and July to August, linked human replies fell from 0.44% of sends (969 of 219,109) to 0.37% (1,179 of 317,490), and the rate fell in 8 of the 14.
6. The 20 other workspaces that sent in July to August drew linked positive replies at 0.07% of sends (115 of 162,379), against 0.15% for the 14 (487 of 317,490).
7. First emails drew positive replies at 0.15% of sends (615 of 402,861 emails sent from June to August 2026) and follow-ups at 0.11% (250 of 230,692), so first emails, 63.6% of sends, produced 71.1% of linked positive replies.
8. 26.7% of the emails in the sending log (444,580 of 1,667,319) went to deliverability-test and internal addresses, and counting them and their replies gives a January to September raw reply rate of 1.9%, against 2.5% for prospect emails alone.
9. October to December 2025 hold 3, 7 and 45 logged prospect emails against 2,757, 2,601 and 2,801 replies, and the log holds only 52.8% to 67.3% of the emails sent from January to April 2026, so those seven months carry counts and no rates.

How to read these numbers. A rate per send counts the replies received in a calendar month per 100 prospect emails sent in that month, so 0.48% means 0.48 replies for every 100 emails sent. Follow-up emails count as sends. A human reply is any first response from a person, refusals included, and a positive reply is the part of those that showed interest. Linked rates, used for the workspace and position comparisons, count only replies matched to the email they answer and run below the monthly rates. Rates are given for May to August 2026, when the platform's sent-email log is complete, and for 1 to 22 September 2026 as a partial month.

Reply rates month by month

• All replies
• Human replies
• Positive replies

Chart data

month	All replies	Human replies	Positive replies
2026-05	2.13 per 100 sends	0.55 per 100 sends	0.23 per 100 sends
2026-06	2.26 per 100 sends	0.53 per 100 sends	0.20 per 100 sends
2026-07	2.43 per 100 sends	0.49 per 100 sends	0.18 per 100 sends
2026-08	1.95 per 100 sends	0.40 per 100 sends	0.14 per 100 sends
2026-09	1.56 per 100 sends	0.39 per 100 sends	0.14 per 100 sends

Source: Carroway Research. Replies received in each month divided by prospect emails sent that month, times 100. September covers 1 to 22 September. Sends and reply counts for every month are in monthly-rates.csv.

From May to August 2026 the sent-email log is complete, and the chart divides the replies received in each calendar month by the prospect emails sent that month. Sending more than doubled over those four months, from 102,525 emails in May to 255,675 in August, and the number of workspaces sending rose from 21 to 30. The human and positive reply rates fell in each of those months.

The raw reply rate moved less, between 2.0% (4,994 replies in August) and 2.4% (5,450 in July), because automated replies grew as a share of all replies, from 57.4% in May (1,254 of 2,184) to 72.5% in July (3,953 of 5,450) and 69.8% in August (3,487 of 4,994). July had the highest raw rate of the four months and a lower human rate than May or June, so the raw rate says little about how many people answered.

September covers 1 to 22 September. Replies to its latest emails were still arriving when the data were read, so its rates are a floor: 0.39% human (612 replies, 158,158 sends) and 0.14% positive (227).

Per-contact rates run a little above per-send rates because some contacts received more than one email in a month. In August, 222,619 contacts received 255,675 prospect emails, and 993 contacts sent a human reply, a rate of 0.45% per contact emailed. Every month's contact counts are in monthly-rates.csv.

What a reply is

Auto-replies and out-of-office notices66.1%

Human, negative12.2%

Later messages in a conversation already under way8.7%

Human, positive8.1%

Bounce and delivery-error notices1.9%

Empty, spam, test, warm-up and internal messages1.6%

Human, wrong person or other1.4%

Source: Carroway Research. Reply types from the sending platform's automated reply classifier, prospect replies only. Counts per type are in reply-types.csv.

Of the 16,090 replies received from May to August 2026, 68.0% were automated (10,936). Auto-replies and out-of-office notices made up 66.1% of all replies and bounce notices 1.9%. Human replies split into negative replies at 12.2% (1,968), positive replies at 8.1% (1,298) and wrong-person or other replies at 1.4% (231). The rest were later messages in conversations already under way (8.7%, 1,407) and empty, spam, test, warm-up and internal messages (1.6%, 250). Of the 3,497 human replies, 1,298 were positive.

A reply rate quoted without a definition can mean any of these layers. On the same 736,078 emails it is 2.2% counting every reply, 0.48% counting people and 0.18% counting interest.

Why the rate fell

Two things could lower a pooled rate while volume grows: the programs that keep sending draw fewer replies, or more of the sending comes from programs with lower rates. To separate them, we compared linked reply rates, grouped by the month each email was sent, for the 14 workspaces that sent at least 5,000 prospect emails in both May to June and July to August 2026.

The 14 workspaces sent 219,109 prospect emails in May to June and 317,490 in July to August. Their linked human reply rate fell from 0.44% (969 replies) to 0.37% (1,179), lower in 8 of the 14, and their linked positive rate fell from 0.18% (403) to 0.15% (487), lower in 9 of the 14. Every other workspace together sent 36,358 prospect emails in May to June and 162,379 in July to August. In May to June those workspaces drew linked human replies at 0.48% (175) and linked positive replies at 0.17% (61), close to the 14. In July to August, with 20 workspaces in the group, they drew 0.33% (535) and 0.07% (115).

Part of the fall happened inside programs that kept sending, and part came from the larger share of email sent by programs whose emails drew fewer positive replies. The 14 workspaces are 14 different programs with their own lists, offers and copy, so their combined fall averages changes that went in both directions.

Linked rates count only replies matched to the email they answer: the latest logged email from the same workspace to the replying address. 2,847 of the 3,497 human replies received from May to August 2026 (81.4%) could be matched. The rest came from an address the workspace had not emailed before the reply, such as a colleague answering for the contact, so linked rates run below the monthly rates.

First emails and follow-ups

1st email0.15 per 100 sends

2nd email0.11 per 100 sends

3rd email0.11 per 100 sends

4th email or later0.08 per 100 sends

All follow-ups (2nd or later)0.11 per 100 sends

Source: Carroway Research. Linked replies only. Position is the email's order among all logged emails from the same workspace to the same address. Sends, human replies and positive replies for every row are in email-position.csv.

The log records each email and does not store its step in a sequence. We numbered each email by its order among all logged emails from the same workspace to the same address, and matched each reply to the latest earlier email. For prospect emails sent from June to August 2026:

• First emails were 63.6% of sends (402,861) and drew 0.39 human replies per 100 sends (1,589) and 0.15 positive replies (615).
• Second emails were 24.4% of sends (154,548) and drew 0.36 human replies per 100 (563) and 0.11 positive (174).
• Third emails were 7.5% of sends (47,261) and drew 0.33 human replies per 100 (155) and 0.11 positive (53).
• Fourth or later emails were 4.6% of sends (28,883) and drew 0.33 human replies per 100 (94) and 0.08 positive (23).

Across all follow-ups (230,692 emails), the human rate was 0.35 per 100 sends (812 replies) and the positive rate 0.11 (250). First emails produced 66.2% of linked human replies (1,589) and 71.1% of linked positive replies (615). A follow-up drew a person to reply almost as often as a first email did, and drew positive replies less often, 0.11 per 100 sends against 0.15.

Months without rates

October to December 2025. The operator's own system marked 115,147, 121,014 and 132,767 emails as sent in those three months, and the sent-email log holds 6 of them. It holds 3, 7 and 45 prospect emails for the three months, while 2,757, 2,601 and 2,801 replies arrived. Those replies answer emails that are missing from the log, so the months are flagged small and carry no rates.

January to April 2026. The log fills in from 6 January 2026. Matched against the operator's own record of the emails its system marked sent, it holds 52.8% to 67.3% of them from January to April, against 91.9% to 98.1% from May to September. Replies to the missing emails still arrive, and dividing them by the logged sends would overstate the rates, so these months are flagged log incomplete and carry counts and no rates.

Test and internal addresses. 444,580 of the 1,667,319 emails in the log for the period (26.7%) went to addresses that are not prospects: deliverability-monitoring and inbox-placement test addresses, each of which received 100 or more emails, and the platform's own sending mailboxes. They are left out of sends, and the 1,125 replies from those addresses are left out of replies. The table shows how these counting rules move a pooled rate on the same data.

Each row removes one problem of the row above it. The first counts replies to emails the log does not hold, the second still counts test and internal emails as sends, and the third still includes the incomplete months. The last row is the basis for this note.

Method

Source. Outbound email programs run by one B2B outbound operator for its clients and itself on one sending platform, October 2025 to September 2026. 54 workspaces sent prospect emails in that time, 42 of them from May to August 2026. A workspace is one client's account on the platform, and the operator's own program is one of them. The data were read on 23 September 2026 and cover 1 October 2025 to 22 September 2026. Months are calendar months in UTC.

Sends. The platform's sent-email log holds one record per email sent, first emails and follow-ups alike. We counted every record in the period except emails to test and internal addresses. A test or internal address is any recipient address with 100 or more logged emails in the period, which catches deliverability-monitoring and inbox-placement test addresses, or any address that is itself one of the platform's sending mailboxes. The rule removed 444,580 of 1,667,319 logged emails and left 1,222,739 prospect sends. 1,970 prospect sends have no workspace recorded; they count as sends and are not counted as a workspace.

Contacts. A contact is a distinct recipient address, compared in lower case. A contact emailed in two months counts once in each.

Replies and reply types. The platform stores each inbound reply with a label from its automated reply classifier and a flag for automated replies. We used the labels as recorded and did not read or relabel replies. Replies from test and internal addresses (1,125) are left out. Each reply gets one type, with the first matching rule applied:

• Automated: labeled as an auto-reply, an out-of-office notice, a bounce or delivery-error notice, or an automated referral, or flagged as automated by the platform.
• Human: a first response from a person, whether positive, negative, from the wrong person or other.
• Positive: a human reply labeled possibly interested, question, contact later or manual referral, the four labels that define a positive reply for 2025 to 2026 in Note 02.
• Other: later messages in a conversation already under way, and empty, spam, test, warm-up and internal messages. They count in raw replies only.

definitions.csv lists every classifier label behind each type.

Rates. Each rate in this note uses one of these definitions.

• Rate per send: replies of a type received in a calendar month, divided by prospect emails sent in the same month, times 100. Both sides are counted by calendar month, so a reply to an email sent at the end of one month can land in the next. The monthly rates, the pooled rates and Exhibit 1 use this definition.
• Rate per contact: distinct contacts who sent a reply of that type in a month, divided by distinct contacts emailed in that month, times 100.
• Automated share: automated replies divided by raw replies, times 100.
• Linked rate: replies matched to an email, divided by emails sent, grouped by the month the matched email was sent. A reply is matched to the latest logged email from the same workspace to the replying address sent before the reply, and replies from any other address stay unmatched. The same-workspace comparison and the position split use this definition.

Pooled rates add the counts across months before dividing.

Position. The log does not store sequence steps. We numbered each prospect email by its order among all logged emails from the same workspace to the same address, so 1 is the first email and 2 or later a follow-up. A contact enrolled again in a later sequence by the same workspace continues the count. The position split covers emails sent from June to August 2026, so that the earlier emails of most sequences fall in months when the log was complete.

Log completeness. To test the log, we matched every email the operator's own system marked sent against it: same workspace, same address, logged within 3 days either side. Months where under 90% of those emails appear in the log are flagged log incomplete.

Month status. Each month in monthly-counts.csv has one status. Small: fewer than 10,000 prospect sends. Log incomplete: log capture under 90%. Partial month: September 2026, 1 to 22 September. Complete: May to August 2026. Only complete and partial months carry rates.

Exclusions.

• 358 logged emails with no send date and 7,570 sent after 22 September 2026.
• Emails to test and internal addresses, and replies from those addresses.
• Rates for months flagged small or log incomplete.

Limits.

• The programs belong to one operator's clients and run on one platform. Their lists, offers and copy differ, and the largest workspace sent 26.9% to 32.3% of prospect emails in each month from May to August 2026. The operator's own program is included.
• Reply types come from the platform's automated classifier and were not checked by hand.
• The test-address rule leaves in a small number of internal addresses that received fewer than 100 emails.
• Monthly rates count replies and sends by calendar month, so a month's rate includes some replies to the previous month's emails and misses some replies to its own.
• The position split matches a reply to the latest earlier email, so a reply to an earlier email in the same thread is credited to the later one.
• September is a partial month and its rates will rise as late replies arrive.

Data

The nine CSV files below hold every number in this note. monthly-counts.csv covers all twelve months with their status, and monthly-rates.csv gives the rates for May to September 2026. reply-types.csv, email-position.csv and same-workspace-trend.csv hold the three comparisons. pooled-rates.csv, log-completeness.csv and coverage.csv trace the counts from the full log to the rows analyzed, and definitions.csv lists every definition and classifier label. queries.txt holds the exact aggregation pipelines, with the database collections given neutral names. The files are free to reuse with credit to Carroway Research.

• Counts by month, October 2025 to September 2026 CSV, 12 rows, 1.0 KBOne row per calendar month: status, prospect sends, contacts emailed, sending workspaces, the largest workspace's share of sends, log capture, emails to test and internal addresses left out, and raw, automated, human and positive replies with the automated share.
• Reply rates by month, May to September 2026 CSV, 5 rows, 0.7 KBOne row per month with a complete send log, September partial: sends, contacts, replies by type, and raw, human and positive reply rates per send and per contact.
• What the replies were, May to August 2026 CSV, 7 rows, 0.4 KBReplies received from May to August 2026 by type, with each type's share of all replies.
• Reply rates by email position, June to August 2026 CSV, 5 rows, 0.5 KBProspect emails sent from June to August 2026 by their position among the emails to the same address: sends, linked replies by type, rates per send, and each position's share of human and positive replies.
• Same workspaces, May to June against July to August 2026 CSV, 4 rows, 0.4 KBLinked human and positive reply rates for the workspaces that sent at least 5,000 prospect emails in both periods, how many of them fell, and the same rates for every other workspace in each period.
• How counting rules move a pooled reply rate CSV, 4 rows, 0.8 KBPooled sends, replies and rates under four scopes, from every logged email and every reply to prospect emails in the complete-log months.
• Completeness of the sent-email log by month CSV, 12 rows, 0.3 KBEmails the operator's own system marked sent each month, how many appear in the sent-email log, and the share.
• Coverage and exclusions CSV, 20 rows, 1.7 KBLogged emails, replies and workspaces at each step from the full log to the rows analyzed, with the share of replies that link to a logged email.
• Definitions CSV, 20 rows, 3.3 KBEvery term, reply type, rate and month status used in this note, with the classifier labels behind each reply type.