How many follow-ups should you send? (the data-backed answer)

Three to five follow-ups over two to three weeks is a reasonable default, and it is a default rather than an answer because the correct number is an output of three constraints you can measure: how much complaint budget you have, how reliably your system detects a reply and stops, and how disciplined your suppression is. Get those three right and six touches is safe. Get them wrong and two touches will damage a domain. The interesting work is in the constraints, not in the count.

Why the number is an output, not a choice

Every extra touch does two things at once. It adds a chance of a reply from someone who meant to answer and forgot, and it adds a chance of a complaint from someone who has decided you are a nuisance. Both scale with the same volume. The question is which one grows faster on your list, and that depends on the list rather than on outbound in general.

Here is the arithmetic, on made-up but clean numbers so the shape is visible.

Take 1,000 leads. Suppose the first email produces a 2% reply rate — 20 replies — and each subsequent touch produces roughly half the incremental replies of the one before: 10, 5, 3, 2. Five touches total 40 replies against 1,000 leads, and the fourth and fifth touches together contribute 5 of them. Now suppose complaints run at a flat 0.1% per delivered message. Touch one on 1,000 sends produces one complaint. Five touches produce close to 4,800 delivered messages after replies and bounces drop out, and about 5 complaints.

The ratio is what to look at. Touch one produced 20 replies per complaint. Touches four and five together produced roughly 2.5 replies per complaint. The sequence is not becoming unprofitable, but its marginal quality is falling by an order of magnitude, and the complaint side is the one with an external ceiling attached to it.

That ceiling is not metaphorical. Google requires senders of more than 5,000 messages a day to Gmail accounts to keep the spam rate reported in Postmaster Tools below 0.30%, a requirement in force since 1 February 2024. Microsoft began applying comparable authentication requirements to high-volume senders to its consumer domains on 5 May 2025. A complaint rate is a budget with a hard limit, and touches spend it.

So the decision rule is: add touches while your measured complaint rate has headroom, and stop when it does not. Not "five, because an article said five".

What stopping on reply actually requires

"Stop on reply" is the sentence every tool prints and the behaviour that is genuinely hard, because "reply" is three different events that need three different responses.

A genuine human reply must stop the sequence. In Autocloz this sets the enrolment to stopped with reason replied, stamps an end time, and promotes the lead's status so it surfaces in the queue that needs a person. A new human reply also clears any prior "reviewed" marker, so a second message from someone you already looked at reappears rather than staying dismissed.

A hard bounce must stop it and suppress the address. The address goes onto the workspace suppression list, and for an unambiguous hard bounce or a complaint it also goes onto a platform-wide list, because repeated complaints are treated by mailbox providers as a reputation event rather than a per-sender one. A confirmed bounce then blocks that address at any age, through a single shared predicate used by both the enrolment gate and the send gate — so the two can never drift apart and let a dead address back in.

An out-of-office or auto-responder must stop nothing. This is the one most systems get wrong, and it is expensive in the opposite direction. The recipient is temporarily unavailable, not disengaged. Treating "I am on leave until the 14th" as a reply retires a live lead permanently, and the person never hears from you again. Autocloz classifies inbound intent and leaves the enrolment running in that case, so the next step still fires on schedule.

There is a fourth case worth naming because it is where the classification is genuinely ambiguous: the delayed delivery notification. Gmail sends a delivery status notification with a 4.x.x code for a transient delay, and it looks like a bounce to any coarse classifier. Suppressing a valid recipient on a transient failure is the worse error, so the correct bias is to record it, classify the severity, and only suppress on a genuine hard bounce or complaint. If your tool suppresses on anything labelled "bounce", check what it does with delay notifications before you trust your suppression list.

Behind all of that sits a second layer. The reply cooldown defers the next step when the lead replied within a configured window — 24 hours by default, adjustable up to 720. It exists because the primary stop cannot cover every path: manual sends outside a sequence, an enrolment somebody restarted by hand, a reply that arrived through a channel the classifier does not watch. Belt and braces on the most embarrassing failure in outbound is proportionate. One queue holding replies from every channel is what makes the human half of this fast enough to keep up with the automated half.

Spacing: the decision rule at each gap

The gaps matter more than the count, and there is a rule rather than a table.

Between touch one and two: one to two working days. Long enough that your first message has scrolled out of the visible part of the inbox, short enough to read as diligence. Never inside the same working day; a same-day bump is the clearest possible signal that a machine is sending.

Between two and three: three to four working days. This is the gap that carries the most incremental replies in most sequences, because it is where people who genuinely meant to answer surface.

Between three and four: five to seven days. The tone shifts here. Anything before touch three can be a reminder. From four onward you need a new reason to be in the inbox, not the same reason repeated.

Before the breakup: seven to ten days. The gap is part of the message. A breakup sent three days after the previous touch reads as a tactic; sent after a longer silence it reads as what it says it is.

Two rules that override all four numbers. Skip weekends and the recipient's public holidays, which means the gap is in working days rather than calendar days. And never let two channels fire on the same lead on the same day — an email and a LinkedIn message arriving within an hour is not multichannel, it is a swarm.

Autocloz's free plan covers 5 users and 10 mailboxes with per-channel caps and the reply cooldown active from the first campaign — start free and set the cooldown before you build the second touch, not after somebody gets double-messaged.

A worked cadence, with the reason for each touch

Five touches over about eighteen working days. The reason column is the part that matters; copy the reasons, not the days.

• Touch 1, day 0 — the specific observation. One reason you are writing to this person rather than a similar one. If you cannot write that sentence, the lead does not belong in the sequence.
• Touch 2, day 2 — the shorter version. Same ask, half the words, no new argument. This catches people who saw the first message at a bad moment. Reply to the same thread.
• Touch 3, day 6 — a different angle. A new reason, not a repetition. A relevant example, a specific number, an answer to the objection you expect. This is where a new thread with a new subject is defensible, because you are genuinely raising something new.
• Touch 4, day 11 — the low-friction ask. Drop the meeting request. Ask a question they can answer in one line: is this you, or should I be talking to someone else? Wrong-person routing is one of the more common useful outcomes of a sequence and touch four is where you should ask for it.
• Touch 5, day 18 — the breakup. State that this is the last message, make the exit clean, and mean it. How to write a breakup email covers the version that works and the version that is a fake ultimatum.

Three variants worth knowing. For a high-value, small list, stretch the same five touches over six weeks and write each one by hand. For a very large list, cut to three touches and put the saved volume into a better list instead. And for a channel switch — a call or a LinkedIn message between email touches — treat it as a replacement for a touch rather than an addition, or you have doubled your contact frequency while telling yourself you diversified. Multichannel sequence templates set out the versions that hold that rule.

The mechanism underneath is worth knowing if you are building this rather than describing it. A sequence in Autocloz is an ordered list of steps, unique per campaign and ordinal, each with a wait in seconds and a kind — email, call, LinkedIn invite or message, SMS, WhatsApp template, or a flow-control step. Branching is a condition on a completed step: if_replied or if_not_replied with a within_hours window and an action of stop or skip to a named step, plus conditions on opened, clicked, bounced, connection accepted and message viewed. That palette is what lets a cadence be a decision tree rather than a fixed list, and what a sales sequence actually is covers the difference from a drip campaign. If you want to lay one out before building it, the cold email sequence builder will draft the steps and gaps.

The failure modes that produce over-contact

Over-contact is almost never a decision. It is one of five bugs.

The same person on two lists. Two campaigns, two enrolments, two independent cadences, and a recipient receiving eight emails in a fortnight from a company that believes it sent four. Deduplicate across lists before enrolment, not within a list.

A reply to a mailbox nobody watches. If your reply detection reads one mailbox and the sequence rotates across ten, replies to the other nine never stop anything. Confirm that inbound is wired for every sending mailbox, not just the first one you connected.

The auto-responder treated as a reply, or the reply treated as an auto-responder. The first loses the lead; the second keeps messaging someone who answered. Both come from a classifier with no visible output, which is why the classification should be inspectable on the message rather than implicit in the enrolment status.

A retry loop that looks like a cadence. A step that fails to dispatch and retries is not a new touch, but if the failure happens after the message left, it is. Autocloz caps consecutive dispatch failures on one enrolment at eight and then stops the row with an explanation rather than retrying forever, which is the behaviour to look for: bounded retries, and a visible reason at the bound.

Suppression that is scoped too narrowly. An unsubscribe recorded against one campaign rather than the workspace means the same person is contactable from the next campaign. Suppression belongs at the highest scope that is legally and commercially correct — the workspace at minimum, and for hard bounces and complaints, wider.

The counterweight to all of this is making the exit easy. RFC 8058, "Signaling One-Click Functionality for List Email Headers", published by J. Levine and T. Herkula in January 2017, defines the List-Unsubscribe-Post header that lets a mail client issue an HTTPS POST and, in the RFC's own words, know that the request "can be fully processed by the mail sender's system" without further interaction. Google requires bulk senders to support it for marketing and subscribed messages and to include a visible unsubscribe link in the body. CAN-SPAM separately requires a working opt-out and a valid physical postal address, and the FTC's guidance sets the window for honouring an opt-out at 10 business days. Ten business days is the legal ceiling; anything other than immediate is a choice you are making.

What changes when ten people share a list

At one person, the cadence is a setting. At ten, three things change.

Frequency becomes a property of the contact, not the campaign. Each rep obeys their own cadence and the recipient experiences the sum. The control that works is a workspace-level minimum gap between any two touches to the same person on any channel, enforced at dispatch rather than in a document.

Ownership of the stop becomes ambiguous. A reply arrives to a shared mailbox and three people see it. Route it to one owner — in Autocloz, the reply notification goes to the member who owns the sending mailbox rather than broadcasting to everyone — so somebody is responsible for the follow-up rather than everybody assuming somebody.

The complaint budget becomes shared and invisible. Your spam rate is a domain-level number. One person's aggressive seventh touch spends the budget the whole team is sending against, and nobody sees it until placement drops. Review it as a team metric on a schedule, not when something breaks. What to send after no response is the individual version of this question; the team version is a volume-and-reputation question, and if you are comparing tools on how they enforce these limits, the Autocloz and Saleshandy comparison covers the sequencing side directly.

What a sequence cannot decide for you

It cannot tell you whether the list was right. A five-touch cadence against the wrong people produces five times as much evidence that they are the wrong people.

It cannot stop on a reply it never sees. Reply detection requires inbound to be connected for the mailbox that sent the message. If it is not, every gate described here is inert and the sequence runs to completion regardless — a silent failure with no error attached to it.

It cannot classify intent perfectly. The separation of human reply, hard bounce, delay notification and auto-responder is a classifier, and classifiers are wrong sometimes. The design bias is to keep sending on ambiguity and stop on certainty, which means an occasional message to somebody who thought they had answered.

It cannot make touch six as good as touch one. Nothing can. The incremental reply from a late touch is real and small, and the honest reason to send it is that it is cheap, not that it is effective.

And it cannot substitute for a reason to be in the inbox. Five messages saying the same thing in different words is one message with four retries. If you cannot write a distinct, honest reason for each touch, the correct number of follow-ups is the number of reasons you have.