The category converged
Pick any two LinkedIn automation tools and list the differences. Cloud sending, sequence builder, a safety story about limits and IPs, a CRM connector, an AI writer bolted on since 2025. The feature tables are functionally identical and the pricing sits in the same band.
Which means the tool stopped being the variable some time ago. If everyone is sending the same shaped sequence through the same shaped platform, the software cannot be what separates a programme booking four meetings a month from one booking thirty.
The recipients noticed before the vendors did. Expandi’s platform data, drawn from 13.2 million connection requests, shows replies to connection notes falling from 3.5 percent in May 2025 to 2.2 percent in April 2026. A 37 percent relative decline in a year, while acceptance rates held roughly steady. People are still accepting. They have stopped bothering to reply, because the note reads like the last nine they got.
AI made this worse rather than better. Generated openers converged on a recognisable shape some time in 2025, and once a reader can spot the shape, the personalisation stops proving anything. It was only ever evidence that a human spent time on you. That evidence is now cheap, which means it is worthless.
So what does separate results?
We have run 41 client programmes since 2018, 389,890 prospects and 15,018 meetings, and tested fifteen individual variables properly: one change at a time, prospects randomised inside each account so industry, geography, seniority and company size stay constant on both sides, minimum 2,000 prospects per arm.
Three things carried real weight.
1. Targeting. Worth 4x, and it happens before you write anything
Across 6,870 enterprise prospects, an identical sequence produced a 13.0 percent reply rate on a list built from job titles and 51.9 percent on a list where we had verified, per person, that the specific problem sat inside their remit.
Four times the reply rate. Same words, same sender, same week. The only difference was the work done before the send button.
Nothing else we tested comes close. The biggest copy-level effect in the whole dataset, naming the prospect’s own product or a direct competitor, was a doubling from 17.1 to 34.5 percent. Real, and half the size.
This is the finding that annoys people, because it is slow, unglamorous and cannot be bought. Reading a profile and deciding whether a problem really belongs to that person is the least interesting task in outbound and the highest paid.
2. Speed. Worth 6x, and it is not a sales problem
Replies answered within 24 hours converted to a meeting at 21.4 percent. Replies answered after 48 hours converted at 3.6 percent.
Nearly six times, on replies you have already paid to generate. Measured across 10,120 inbound replies and every one of the 41 accounts.
A reply is not interest, it is a window. For a few hours the person has an active thought about the problem you named. Two days later the thought is gone, the context has to be rebuilt, and the message that felt worth sending now feels like a commitment they did not intend to make.
This is an operations failure rather than a selling one, which is why it persists. The inbox is never the thing on fire that morning.
3. Media. Worth 40 percent, and it is the only one with a moat
A personalised video in the first touch replied at 40.4 percent against 28.8 percent for text-only, across 8,385 prospects, and produced 57 percent more meetings. A voice note at step two lifted meetings from 4.1 to 5.0 percent for a fraction of the production cost.
The number matters less than the durability. It held across all three accounts and showed no decay over eleven weeks, which is unusual. Most format effects in outbound fade within a month as the tactic spreads.
Here is why we think it lasted, and it is the central argument of this piece.
Video is expensive to fake, and that is the entire point. Text personalisation was valuable precisely as long as it was costly. Once a model could produce a plausible personalised opener for nothing, the signal stopped meaning anything, because the recipient knows the cost has gone to zero. Thirty seconds of you talking to one named person cannot yet be produced for nothing, which is why it still reads as effort.
The honest caveat: we cannot tell you what happens when everyone does it. Our eleven weeks says it has not decayed yet. It says nothing about a market where video is the default. Anyone claiming otherwise is guessing.
What hyper-personalisation actually means
The phrase has come to mean more sentences about the prospect. Our data says that is exactly backwards.
| Variant | Reply rate |
|---|---|
| Longer message with solution detail | 18.6% |
| Two and a half sentences, question-led, no pitch | 44.5% |
| Company-led copy | 10.8% |
| Question-led copy | 33.2% |
| Over 700 characters | 11.2% |
| 240 to 420 characters | 38.6% |
Every row rewards taking words out. Generated personalisation adds them: a clause about their post, a clause about their funding, a clause about their role, each one standing between the reader and the question that would have made them reply.
Useful personalisation, on this evidence, is three things:
A verified reason this person specifically should care. The 4x.
One concrete detail you could only know by looking. Their product, a competitor, something they shipped. One, not four. Two specifics read as research; four read as a dossier.
Media that proves a human was involved. Because the written proof no longer works.
That is it. Everything else is words the reader has to get through.
What this means if you are choosing a tool
Choose on whether it helps with those three things, and be honest that most of the category helps with none of them.
Does it stop you contacting people who do not fit, or does it happily send to whatever you upload? Does it get a reply in front of you the same day, or does it dump everything in an inbox nobody owns? Can you record once and personalise it per prospect, or is video a manual job you will do for the first twenty and then quietly stop?
If the answer to all three is no, the tool is a sending mechanism, and sending mechanisms are a commodity. Pick the cheapest one and spend the difference on the list.
The uncomfortable summary
The work that separates outbound results in 2026 is work you cannot buy: verifying the list, watching the inbox, and recording thirty seconds of video. Every one of those is boring and every one of them beats anything on a feature page.
Full method, sample sizes and all fifteen tests are on the 2026 outbound benchmarks page. Free, no gate, and you are welcome to cite it.