Key finding
Among people who accepted a LinkedIn connection request, 47.8% replied when the request carried a note against 25.5% when it did not, a gap that held in all three customer accounts large enough to compare against themselves. The effect of a note on acceptance was tested and proved inconclusive.
What we measured
LinkedIn lets you attach a short note to a connection request. Whether you should is one of the most durable arguments in outreach, and both sides are argued confidently and without data.
Reachium stores the invitation text sent with every connection request, including when that text was empty. That makes the comparison directly measurable. This study covers 180,155 connection requests sent between 8 January 2025 and 6 August 2026 and given at least 30 days to mature: 19,516 carried a note and 160,639 did not.
Nobody was randomly assigned to write notes. That single fact shapes the entire analysis, and most of what follows is the work of establishing what the comparison can and cannot support.
Methodology
Dataset window. Every figure covers connection requests sent between 8 January 2025 and 6 August 2026, pulled on 6 August 2026.
Note or no note is determined by the stored invitation text. Text that is empty, or contains only whitespace, counts as no note.
Two different units, and they are not interchangeable. A sending account is the LinkedIn account a request went out from, and is the unit the daily-volume bands are built on. A customer account is the customer the requests belong to, and is the unit the paired test compares against itself. Where this study says "account" without a qualifier below, the sentence names which one.
Acceptance rate is accepted requests divided by requests sent, within each group.
Reply rate of accepted is connections producing at least one inbound message divided by accepted connections, within each group. Replies per request sent divides the same numerator by requests sent. Both appear below because they answer different questions.
Maturity. Only requests sent at least 30 days before the 6 August 2026 pull are included, so recent requests are not counted as failures before they have had a chance to be accepted. One check below is reported on the unrestricted all-sent cohort instead, and says so where it appears.
This is observational. The groups were not randomised, so any raw difference between them could be caused by the note, or by the people who choose to write notes, or by who they were writing to. Two checks are reported below to separate those explanations. One of them changes the conclusion.
The findings
The raw comparison, matured cohort:
| Sent | Accepted | Acceptance | Replied (of accepted) | Replies per request sent | |
|---|---|---|---|---|---|
| No note | 160,639 | 44,426 | 27.66% | 25.54% | 7.06% |
| With a note | 19,516 | 4,410 | 22.60% | 47.80% | 10.80% |
Read alone, that table says a note costs 5.06 points of acceptance and adds 22.26 points to the reply rate among accepted connections (both derived from the table above). The second half of that reading survives scrutiny. The first does not, and the rest of this section is the work of showing why.
Check one: is it just sending volume?
Acceptance varies with a sending account's habitual daily invite volume. If note writers were concentrated in one part of that range, the pooled acceptance gap could be a volume gap wearing a disguise. Splitting both groups by the sending account's average daily invite volume:
| Average invites/day | Acceptance, no note | Acceptance, with a note |
|---|---|---|
| Under 10/day | 23.06% | 18.32% |
| 10 to 19/day | 30.80% | 26.67% |
| 20 to 29/day | 26.31% | 22.69% |
Both columns are acceptance rates, and this check is computed on the all-sent cohort rather than the matured one, which is why the levels differ slightly from the table above. The gap these columns show is the pooled acceptance gap, and the next section refutes it: it does not survive comparing customers against themselves. This table is not evidence that a note lowers acceptance.
What it does rule out is one specific alternative explanation. The pooled gap holds inside every band and in the same direction, so sending volume is not what produces it.
Check two: is it just which customers write notes?
It is largely this. Notes are written by a small minority of the customer accounts in this dataset. The largest single contributor to the noted group accounts for 42.99% of every noted request; the largest single contributor to the no-note group accounts for 28.02% of that group. The noted group is the more lopsided of the two by a wide margin, which is the point: the raw comparison is partly a comparison between customers rather than between techniques.
The test that separates them is to compare each customer against itself. Three customer accounts had at least 200 matured requests in both groups:
| Customer account | Note requests | No-note requests | Acceptance, note | Acceptance, no note | Reply of accepted, note | Reply of accepted, no note |
|---|---|---|---|---|---|---|
| 1 | 1,652 | 28,546 | 36.32% | 27.58% | 38.33% | 23.44% |
| 2 | 5,681 | 6,076 | 21.48% | 31.37% | 46.39% | 20.20% |
| 3 | 8,389 | 2,546 | 22.29% | 21.48% | 57.38% | 29.43% |
On acceptance the three disagree, and they disagree by a lot. Account 1 accepted 8.74 points higher with a note, account 3 by 0.81 points, and account 2 accepted 9.89 points lower (all three derived from the table above). Two of the three ran opposite to the pooled table. That is Simpson's paradox: the customers who write notes have different baseline acceptance for reasons that have nothing to do with notes, and pooling them produces a direction that none of the individual comparisons agrees on.
This study therefore makes no claim in either direction about connection notes and acceptance rate. The pooled difference is reported above because it is real in the data, and it is explained here because it is not what it appears to be. It is not evidence that a note lowers acceptance. The paired test is not evidence that a note has no effect either, given that one customer account is 8.74 points better with a note and another is 9.89 points worse. The accurate word for the acceptance result is inconclusive.
On reply rate the pattern is different, and it is consistent. All three customer accounts replied more when the request carried a note, by 15 to 28 percentage points, in the same direction and roughly the same size as the pooled figure. Removing the single largest customer workspace, which writes no notes at all, moves the no-note group to 29.64% acceptance and 24.61% reply of accepted while leaving the noted group untouched at 22.60% and 47.80%. The reply gap widens rather than closing.
The complication
Noted requests produced more replies and fewer meetings.
| Booked meetings | Of accepted | Of sent | |
|---|---|---|---|
| No note | 916 | 2.06% | 0.57% |
| With a note | 57 | 1.29% | 0.29% |
That result rests on 57 booked meetings and on customers using the platform's own booking step, so it is weak evidence. It is reported because leaving it out would make the reply finding look cleaner than it is.
What this means
Whether a connection note changes acceptance is not settled by this data. The pooled comparison points one way, the paired comparison points the other, and with three customer accounts to compare, neither is strong enough to carry. Inconclusive is the right word, and it is not a synonym for no effect: two of the three paired comparisons found a difference of roughly nine points, in opposite directions.
The reply gap is the quantity this study does establish, and it is large: noted requests were associated with 22.26 more points of reply rate among accepted connections, consistently across all three paired comparisons. What the study cannot do is attribute that gap to the note itself rather than to the sequence the note belongs to, for the reason given in Limitations.
There is a plausible mechanism, though this data cannot confirm it. A blank request asks for nothing and gets accepted on autopilot. A request carrying a note is read before it is accepted, so the people who accept it have already engaged with something. How that changes the number who accept is exactly the question this data could not answer, and it does not establish that this mechanism is what produces the reply gap either. It is one explanation consistent with the observation, offered as such.
The meeting numbers cut against that reading and should not be waved away. They are also thin enough that they could reverse entirely on more data.
In practice
An SDR whose target is connections added is being measured on the one thing this study could not settle. The data does not tell them whether to write a note. It does tell them that the acceptance number they are judged on will not answer the question either, because two of the three customer accounts measured here ran in the opposite direction from the pooled result.
A team that switched to blank requests on the strength of a pooled acceptance figure switched on the strength of a number that reversed when it was cut by customer. The reply gap is the only quantity here that did not reverse: in the three customer accounts that ran both, the reply rate among accepted connections was 15 to 28 points lower without a note. That is what the switch is measurable against.
Anyone quoting a note-versus-no-note statistic, including the aggregate table in this study, is quoting a number that changed direction when it was cut by customer. That is a general caution about outreach benchmarks rather than a fact about notes.
Limitations
The paired test rests on three customer accounts. Only three had enough volume in both groups to compare against themselves. Three is a small number of comparisons on which to overturn a pooled result, and it is equally a small number on which to confirm the reply finding. More customers running both would settle it.
Pairing does not remove selection inside a customer account. Comparing a customer against itself removes differences between customers. It does not remove the choice made within one. A customer decides which requests get a note, and almost certainly writes different follow-up messages for the sequences those requests belong to. The note is therefore confounded with the whole sequence around it rather than isolated as a variable, and this study cannot say which part of that sequence the reply gap belongs to. This is the strongest objection to the reply finding.
Nobody was randomised. Every conclusion here is observational. The two checks reported above rule out two specific alternative explanations. They do not rule out all of them.
The note group is concentrated. The largest single contributor accounts for 42.99% of all noted requests.
The two band tables in this study use different cohorts. The volume split in check one is computed on all sent requests; every other figure here uses the matured cohort. The levels are close but not identical, and the two should not be read off each other.
Note content is not analysed. Note length was tested and excluded: every length band turned out to be dominated by one or two customers, so the variable measures who wrote the note rather than how long it was. Nothing here says anything about what a note should say.
Replies are counted, not classified. No sentiment is recorded, so a reply saying "please stop" counts as a reply. If notes provoke more declines as well as more genuine interest, this data cannot separate the two, and the meeting figures above are the only hint either way.
The meeting comparison is thin. 57 booked meetings in the noted group, counted only when booked through the platform's own booking step.
The dataset is concentrated overall. The largest customer workspace accounts for 21.97% of all requests sent, the largest five for 57.66%, and the largest ten for 76.19%.