What 27,737 outreach messages taught us: what worked and what did not

Across 27,737 LinkedIn conversations the audience mattered more than anything we wrote. People who had just engaged with a post on our topic replied at 54.6% and booked 4.54 meetings per 100 messaged, against 21.7% and 0.71 for a Sales Navigator search. Observation openers beat questions by 10 to 15 points, follow ups added 45% more replies, and answering in two minutes instead of nineteen changed nothing.

This is a teardown of our own outbound, written by the people who ran it. The full numbers and the method are in the case study. Here we focus on lessons, and on the parts that did not work.

What was the setup?

Since September 2025 our own agents have prospected for Sellinger on LinkedIn. They wrote to founders, heads of growth, sales and revenue leaders, and agency owners, from up to 43 LinkedIn profiles of real people: team members, ambassadors and students. The agent writes in each person's name and voice. The ask was a 15 minute showcase.

The company headline, from an owner statement on 19 October 2026, is over 1,500 calls booked in the last 12 months using on average 7.5 LinkedIn accounts. Everything below comes from the stricter subset we can reproduce from the database: one booking per person through the agents' booking link, so it is a lower bound. Excluded are one profile that ran a research style opener we retired, and threads about a different product, 2,569 people in total.

What did the funnel look like?

• 68,818 connection requests sent, 41% accepted.
• 27,737 people received at least one message after connecting.
• 10,978 replied, 40%.
• 351 showcase meetings booked through the agents, 3.2% of replies.
• 33 confirmed held, meaning a stored recording or a call our team marked as held.

The gap between 351 and 33 is its own lesson, covered in count meetings honestly: a calendar marks a meeting "completed" when its slot ends, whether or not anyone joined.

What worked?

1. The list beat the copy

We tracked reply rate and meetings per 100 people messaged by where the lead came from.

The intent signal comparison is from the October 2025 cohort, and the search figures are for the whole period. The source of the signal mattered as much as the signal. People who engaged with posts by AI SDR vendors and automation creators booked at 2.6% to 3.7% per person messaged. People who engaged with enterprise sales thought leaders booked under 0.4%. That is a tenfold spread from the audience alone, with the same agents and the same offer.

2. Observation openers beat questions

Within the same month, openers that stated an observation about the person's work out-replied openers that asked a question. In February 2026 it was 43.5% against 33.5%, in March 45.2% against 33.6%, in June 51.9% against 37.3%. Meeting rates were equal or better. The pattern is in observation openers beat questions.

3. Follow ups added replies, with a cost

The first message alone got 26.5% of people to reply. With follow ups the total was 38.4%, which is 45% more. But each touch got fewer replies and more refusals. The share of first replies that were negative went from 0.9% at the intro to 8.5%, then 14.3%, then 14.7% for the third follow up and later. We cover how to set the number in how many follow ups.

4. A one line answer to "what is this?"

When someone asked what we do, the agent answered in one line, tied it to their own words and offered 15 minutes. The booking link typically went out after three prospect messages. And 56% of bookings confirmed in the chat needed at least one more nudge after the link went out.

What did not work?

Speed was not the lever

The agents' reply time fell from about 19 minutes to about 2 over the year. The median is 2.3 minutes since May 2026, 90% within an hour, and 74% of the 48% of messages that arrive at night or on weekends are answered before the next morning. Meetings per conversation did not rise with it. Answering fast is table stakes. It stops you losing people. It does not make them want the offer.

Meetings per person fell while replies held

Reply rates stayed between 35% and 45% for twelve monthly cohorts. Meetings per 100 messaged peaked at 2.4 in November 2025 and drifted below 1 from June 2026, to 0.5 in August. Three things moved together: the mix of lists, the message style, and from July a free trial link offered to part of the traffic instead of a call. We cannot separate them from the data. A rising reply rate with falling meetings is exactly the pattern in replies are easy, conversion is audience and offer.

We rarely used the signal that put someone on the list

Only 5 of about 30,000 openers mentioned the post that put the person on the list. People engaged with a specific post, and we wrote to them as if they had not. That is the next experiment, and we say so on the case study page rather than hiding it.

Showing up

One prospect told us that if a host is late to a 15 minute call he is not interested. Our host had joined fifteen minutes late. The agent replied with one honest sentence of apology, no excuses, and he rebooked himself. The agent can book the meeting. Someone still has to arrive.

What would we do differently?

1. Pick the signal source before writing a word. Audiences matter more than adjectives.
2. Use observation openers and say something true about the person.
3. Mention the thing that triggered the signal.
4. Stop at three follow ups that each add something, then a clean exit line.
5. Measure held meetings, not booked ones.
6. When changing the offer, watch meetings per reply rather than reply rate.

How could you run the same tests on your own list?

You do not need 27,737 messages. You need a clean split and the patience to count to the end.

1. Tag every lead with its source before the first message: signal, search or list. Without the tag you cannot compare.
2. Change one thing per test. If you change the audience and the opener together, you will not know which one moved the result.
3. Measure three numbers for each group: reply rate, meetings per 100 people messaged and meetings held. Always write the denominator next to each.
4. Read the replies. The share of negative first replies tells you what a follow up costs, which a reply rate alone hides.
5. Wait for the cohort to mature. A person messaged this week has not had time to book. We compared monthly cohorts for this reason.
6. Write down the definitions. What counts as a reply, a booking and a held meeting? Ours are on the case study page.

What questions do readers ask about this data?

Is 40% a normal reply rate? We do not claim it is. It is the rate for this account, with this offer, over this period, and it counts people who replied at least once after connecting. Compare your own rate on your own definition.

Why report the 33 held meetings next to the 351 booked? Because the booked count flatters. A calendar marks a slot as completed when it ends, so only recordings and meetings our team marked as held count as held. The honest lower bound is the one worth planning on.

Why not just copy the best audience? The best signal sources here were specific to our topic. An audience that works for an AI SDR vendor will not be the one that works for yours. Copy the method of splitting by source, not the source.

How should you read these numbers?

As one account's evidence. It has its own audience, offer and definitions of a reply and a booking. Our other case studies have different audiences and definitions, and are labelled that way on purpose. Use this teardown to choose experiments for your own list, then measure them the same way: replies, bookings and held meetings, each with its denominator.