Between 2 July and 31 August 2026 our sending profiles started 69 conversations that got a reply from the person on the other side. Twenty nine of those produced a second reply. The other forty ended with somebody typing one short line to a stranger and then going quiet, which makes the message we send at that turn the most expensive piece of copy in the whole sequence. This is what a first reply actually contains, what we wrote immediately after it, and the one message we kept leaving unsent. Every figure carries its denominator. The splits inside them run on small samples, so read them as direction and check the same cuts on your own logs before you rebuild anything.
A first reply is seventy three characters long
Across those 69 conversations, the median first reply ran 73 characters. The mean is 125, pulled up by a handful of long ones.
Seventy three characters is "Our offices." It is "Yeah". It is "hey! Sure, what about?" It is an email address with nothing around it. A reply that short carries almost no information about intent, so the intent gets supplied by whoever reads it, and the reading drives the next message. Sorting the replies by what the person was actually doing puts a floor under that guess.
The 69 first replies break down like this:
- 42 gave a partial answer, a correction, or a handoff to someone else
- 11 asked a question back
- 10 declined outright
- 6 asked for a later date
Only the 10 declines are self contained. The other 59 are an opening, and what happens to them depends entirely on the next thing that gets typed.
The steepest remaining drop sits right after the reply
The depth of those 69 conversations, counted by how many messages the other person sent:
- 69 sent a first reply
- 29 sent a second, 42%
- 17 sent a third, 25%
- 6 sent a fourth, 8.7%
Between the first reply and the second, 58% of conversations end. That is a larger proportional loss than anything downstream, and it lands on people who have already accepted a connection request and already typed something to a stranger.
The first controllable variable is whether anyone answers. We answered the first reply in 55 of the 69 conversations. In the 14 we left alone, exactly one person wrote again unprompted. Where we did answer, 27 of 55 conversations continued, 49%.
Read that gap carefully, because it flatters us. Part of the 14 we left alone were clean declines where silence is the right answer, so the comparison mixes a decision about the person with a decision about the message. What survives the caveat is the raw count: fourteen people typed something to us and got nothing back, and the cost of answering is one message.
Message one: answer the thing, then ask about their world
The 55 answers split on one property, whether they contained a question:
- answer contains a question: 15 of 27 conversations continued, 55.6%
- answer contains no question: 12 of 28 continued, 42.9%
- answer is a short acknowledgement under 120 characters with no question: 5 of 12 continued, 41.7%
Twenty seven against twenty eight is a small sample and the gap is thirteen points, so this is a direction rather than a finding. It also has a companion number that complicates the obvious lesson. Counting every outbound message across all 69 conversations, 49.6% of our messages in conversations that continued contained a question, against 56.8% in conversations that stopped. Question marks distributed across a thread do nothing on their own. The turn immediately after the first reply is where one specific question pays.
The shape that worked in our logs is the same three moves every time. Answer the literal thing they said, in one line. Concede anything they corrected, without defending it. Then ask one question about their situation that they can answer from memory.
Worked examples from the conversations that continued, with the identifying parts removed:
- They wrote "Our offices." The answer described the specific case we get called on, one hire starting in a city where the company has no room yet, and asked whether that had come up.
- They wrote that our fact about their employer was out of date. The answer was "Fair hit, that is clearly out of date, my bad", then one question about where they had landed.
- They handed over a colleague's email. The answer confirmed the exact address back to them and asked one qualifying question before using it.
Speed is worth naming because it explained nothing. Our median time from their reply to our answer was 14 minutes, and 41 of the 55 answers went out inside an hour. Splitting continuation by response latency produced no usable signal, mostly because the fastest answers in our data are automatic ones sent to closures. Answer quickly for the ordinary reason, that people forget what they wrote to a stranger. The content of the answer is what moves the number.
Count the characters each side wrote
Add up the characters each side wrote across a whole conversation and the two groups separate hard.
- conversations that stopped at one reply: 472 characters from us, 116 from them, a ratio of 4.1 to 1
- conversations that continued: 858 from us, 366 from them, 2.3 to 1
The tempting conclusion is that our messages are too long, and our data says something narrower. At the single turn right after the first reply, the two groups are identical: our answer averaged 202 characters against their 130 in conversations that stopped, and 195 against 145 in the ones that continued. The median ratio is 1.8 in one group and 1.9 in the other.
So the imbalance is cumulative. It builds one message at a time, and by the time it is visible in the totals the conversation is already over. The worst thread in our own logs ran 2,665 characters from us against 533 from them, with one question per twelve of our messages. The person finished on exactly the position they had opened with.
The rule that follows is about format, and it is easy to hold. Match theirs. A person who writes one line gets one line. A person who writes three paragraphs has invited three paragraphs. Your own explanation of what you do is worth roughly one sentence until they ask for the second one.
Message two, which mostly went unsent
Of the 55 conversations where we answered the first reply, 28 went quiet afterwards. In 5 of those 28 we sent a second message. In the other 23, our answer sat there and the conversation ended with our name at the bottom of it.
Being honest about the 5: none of them produced a reply. Five is far too small to claim that the second message works, and we are not going to claim it. What is certain is the other number. Twenty three times we had a conversation with a live human who had already replied once, and we declined to spend one more message on it.
What that message should carry is the part worth getting right. Repeating the ask adds pressure and no information, and the person has already read it once. The second message earns its place by adding something that was absent from the first: a number from a comparable account, a specific artifact you can send today, a short list of companies you would go after for them. Our own copy rules put this in the follow up canon as a hard requirement, because a follow up that restates the previous message is the machine tell that ends threads.
One related rule, learned the expensive way. A booking link goes into a thread once. Ask for the conversation in words as often as the situation warrants, put the link down a single time, and repeat it only after the person themselves returns to the subject of timing. They saw it the first time.
Deferrals need a date said out loud
Six of the 69 first replies asked for a later date. "Ping me in six months." "I am travelling for two weeks." "Ask me again in mid September."
Most of our answers to those were polite acknowledgements: understood, will do, enjoy the break. A polite acknowledgement is a soft close that both sides forget inside a week. The version that holds is one line longer. Name the date out loud, put it in a calendar where it will fire without you, and say what you will bring when you come back. In our logs the answers that named a specific month and a specific thing to return with are the ones still moving.
What to run on Monday
Five checks, all on data you already have.
- Count how many of your replied conversations produced a second reply from the other person. A share under 40% points the copy work at the message after the reply.
- List every first reply from the last month and mark whether anyone answered it at all. Ours was fourteen unanswered out of sixty nine, and it took a query to find them.
- Read your own message at that turn. It should answer the literal thing they said and ask exactly one question about their situation.
- Add up the characters each side has written in your live threads. Anything past three to one is a signal to cut before you send the next one.
- Pull every conversation that has been quiet for more than four days since your answer, and send a second message that adds a fact rather than repeating the ask.
All five run on the same habit: treating a reply as the start of a conversation with its own numbers, and measuring those numbers on the same schedule you measure sends. If you would rather have this operated for you, with the reply handling, the follow up rules and the counters already built, that is our work, and the levels are laid out on our levels and pricing page.