Observation openers beat questions: what our own LinkedIn data shows

On our own LinkedIn outbound, an opener that states an observation about the person's work out-replied an opener that asks a question in every month we compared: 43.5% against 33.5% in February 2026, 45.2% against 33.6% in March and 51.9% against 37.3% in June. Meeting rates were equal or better. An observation gives the person something to agree with, and a question gives them a task.

This is one account, so treat it as a tested idea, not a law. It is also a rare piece of data, because most advice on openers comes from a template list, not from a comparison inside the same month on the same audience.

What did we compare?

Our agents write the first message after a person accepts a connection request. Over the year there were two common styles:

• Observation. One sentence about the person's work, with no question. Example from the case study: "Seems like your focus on scaling outreach and cost savings is making a tangible difference for your clients."
• Question. A sentence about their work, ending in a question. Example: "As someone scaling multiple AI products, how are you approaching outbound for new client acquisition?"

We then looked at reply rate by opener style, within the same month, so that seasonality and list changes affect both styles equally.

The case study publishes these rates, not the counts behind each split, so read them as direction and size, not as a precise estimate. What matters is that the gap was the same sign every time we looked.

Why would an observation win?

We do not have a controlled experiment on why, so these are readings, not findings.

A question is homework. The person has to decide whether to answer, think of what to say, and type it. An observation asks for nothing, so replying is a choice made from interest.

An observation is a proof of reading. Questions are cheap to write at scale. "What are your biggest challenges with X?" can go to anyone. A true sentence about someone's work is harder to fake and signals a person paid attention.

Questions look like a survey. On a platform where people receive many messages from tools, a generic question reads as one more.

Agreement is a smaller step than an answer. Saying "yes, that is what we do" is easy. Writing three sentences to a stranger is not.

What makes a good observation?

• True and specific. It names something in the person's profile, posts or company that you can point to.
• Short. One sentence.
• Free of a pitch. No link, no product name, no ask. The point is to start a conversation.
• Free of flattery. "Impressive work!" is not an observation. Say what you noticed, not how you feel.
• Connected to a reason you are writing. If someone engaged with a post, say what was in it.

An illustration of the difference, not a quote from a campaign. A weak observation: "I see you work in sales. Great profile!" A stronger one: "Noticed you moved from agency work to running growth at a seed stage product, a different problem to solve each week."

What about questions?

Questions are not bad. Three conditions make one work.

1. It is specific to their work. "How are you handling outbound for the new product line?" beats "how do you approach lead generation?".
2. It is easy to answer in a line. One question, one thing.
3. The audience expects it. A peer asking a peer is natural. A vendor asking a stranger is not.

In the case study, the question opener quoted above was specific to two products the person was building, and he booked after follow ups. The comparison above is about typical openers at scale, not the best question.

What should you test?

Run the split yourself. A short method:

1. Take one list and split it at random into two groups of at least 100 each.
2. Write one observation opener and one question opener about the same person set, using the same details.
3. Keep everything else the same: sender, days, follow ups.
4. Measure replies and then meetings, and compare after four weeks.
5. Keep the winner and test the next variable, such as the length or the first word.

Do not read a one week gap on 40 people as a result. Differences of 10 points on hundreds of people mean something. Differences of 4 points on 30 do not.

Can a question ever beat an observation?

On our own account the comparison went the same way in all three months, so on that audience the answer is rarely. The cases where a question may do better are the ones with a real reason to ask: a person who has publicly asked for advice, a role where answering questions is the job, or a topic where you cannot say anything true without asking. Treat those as tests, not as exceptions to assume.

How do you write an observation from a profile?

A short routine you can run on each person, or give to an agent as a rule.

1. Find one concrete thing: a recent role change, a post, a product, a hire, a market.
2. Say what you noticed in the person's own terms, in one sentence.
3. Add nothing: no pitch, no link, no compliment.
4. Check that the sentence would be false for most other people on the list. If it would be true for all of them, it is not an observation.

What does a bad observation look like?

• Too generic. "Seems like you are growing fast." True for half the list.
• Wrong. The agent states something that is not on the profile. One false line costs more than a blank opener, so keep a rule that the observation must come from a source the agent can name.
• A hidden pitch. "Noticed you still use spreadsheets for outreach" is a sales line dressed as an observation.
• Over personal. Reading a family detail or a location out of a profile feels like surveillance.

How should you count the result?

Count replies per person messaged for each opener, then replies that are more than a pleasantry, then meetings. The case study publishes rates but not the counts, and you should keep counts for your own test. With 100 people per group, a 10 point gap is 10 more replies, which is the smallest gap we would trust. A gap of 3 replies is noise.

What does this change in your campaign?

Make the observation the default opener for a new list, write a rule that it must name something the agent has seen, and keep a question variant as the control. Review the first twenty drafts in approval mode, because the failure to avoid is a fluent but empty first line. Then follow it with a proposition that fits the person (a value proposition that gets read).

What is the next experiment?

The data above comes with a gap we are open about. Only 5 of about 30,000 of our openers mentioned the post that put the person on the list. Most of the leads came from an intent signal, and we wrote to them as if we had not seen it. The obvious next test is an observation that names the post. The playbook for a specific opener on a no note invite describes the shape, and the acceptance rate post covers the step before it, getting the connection accepted.

For the whole picture of what moved results on this account, from the audience to the follow ups, read what 27,737 outreach messages taught us.