What personalisation is not
“Hi Sarah, I see you’re the Head of Sales at Acme.” Sarah knows her name and where she works. This sentence transfers no information and signals that a machine assembled it. It is worse than no personalisation, because it announces the attempt while failing it.
Personalised images with a logo dropped onto a coffee cup were clever in 2021 and are now recognised as a tool feature. Same problem, one layer up.
The four levels
Level 0, merge fields. Name, company, job title. Baseline. No effect.
Level 1, the observable fact. Something true about their company that took a small amount of looking: they are hiring three SDRs, they opened a Dublin office, they moved to a new CRM. Better, but often still applies to a hundred companies.
Level 2, the individual detail. Something about that specific person: a post they wrote, a job they left, a talk they gave, a view they hold. This is where personalisation starts working, and where most people stop trying because it takes real time.
Level 3, the format. A voice note or short video, in your actual voice, addressing them by name. This outperforms everything above it in our data: a video first touch replied at 40.4 percent against 28.8 percent for text-only, with 57 percent more meetings. A recorded human voice cannot be faked at scale by someone who is not bothering.
Most advice tells you to get to Level 2. The gap between Level 2 and Level 3 is bigger than the gap between Level 0 and Level 2.
Where Level 2 details come from
In rough order of how well they work:
- Something they posted or commented on in the last 90 days
- A role change in the last six months, theirs or their boss’s
- What their company’s job ads say they are building
- A shared employer, school or mutual connection you can name
- A company announcement: funding, expansion, product launch, award
- A podcast, webinar or panel they appeared on
The last one is underused and enormously effective, because almost nobody listens to the podcast.
Doing this without spending ten minutes a prospect
The honest tension is that Level 2 requires reading, and reading does not scale by hand. Forty properly researched prospects a day is somebody’s entire job.
Two ways through it.
Narrow the list. If you can only research 40 people a day, message 40 people a day and make them the right 40. A tight list beaten well outperforms a wide list sprayed. This is the option available to everyone and the one most teams reject because the numbers feel too small.
Automate the reading, not the deciding. A model can read a profile, recent posts, the company page and job ads in seconds and extract the one usable detail. It cannot decide who is worth messaging or what to offer them. That split is exactly how the ICP filter works: it reads every profile, scores fit, skips the ones that do not qualify, and writes the opener from what it found.
Then change the format
Once the opener is specific, the biggest remaining lever is not more words. Record yourself once, thirty seconds, and send a personalised voice note or video to every prospect who accepts. Same research, different medium, several times the reply rate. There is a script structure that works.
How to tell if yours is working
Two numbers. Acceptance rate above 35 percent means your targeting and opener are landing. Reply rate above 20 percent means the message is doing its job. Below those, the problem is almost never the wording of the second paragraph, which is where most people spend their editing time.