# Personalisation at Scale Is a Research Problem, Not a Writing Problem

*Nikhai Jaysen · September 3, 2026*

> Why merge-field personalisation fails on LinkedIn, which research signals actually lift acceptance and reply rates, and what to automate versus leave to a human.

Source: https://www.spaceblanket.ai/blog/linkedin-personalisation-research-problem

Everyone knows a merge field when they see one. Real personalisation is not better sentences, it is better inputs — and that is an automation problem worth solving.

Everyone can spot a merge field. Hi {first name}, I saw you work at {company} and thought I would reach out. The recipient has seen that exact shape a thousand times, and the pattern itself communicates that no human looked at their profile.

The usual response is to write better sentences. That is the wrong layer. The message is only as specific as the information behind it, and most outreach fails because there is no information behind it at all.

## Three tiers of signal

Not all personalisation is equal, and the difference is not effort — it is how hard the detail would be to fake.

### Tier one: directory data

Name, title, company, industry, headcount. Available on every profile, used by every competitor, and worth close to nothing. Referencing it signals that you ran a filter, not that you did research. Use it to route someone into the right segment, never as the personalised element itself.

### Tier two: recent activity

A funding announcement, a job change, a new office, a post they wrote, a hiring push in a particular function. This is genuinely specific and it is what most good outreach uses. It has a shelf life — a reference to a funding round from fourteen months ago is worse than no reference, because it reveals the research was stale.

### Tier three: observable symptoms

The tier almost nobody reaches. Something visible from outside the company that indicates the problem you solve is present right now. Four open support roles on the careers page. A demo request form with no calendar booking. A help centre that has not been updated in two years. A pricing page that mentions a feature the product no longer has.

Tier three works because it is not flattery, it is observation, and it implies you have thought about their business rather than their profile. It also does something the other tiers cannot: it makes the reason for the message obvious in the first sentence.

## Where automation genuinely helps

The research step is where volume kills manual approaches. Reading four hundred careers pages, checking four hundred help centres, and noting which companies changed something recently is not a job a founder can do weekly, which is why most outbound quietly degrades to tier one.

This is exactly the kind of work automation handles well: gathering, summarising, and flagging. A pipeline can check each account against a defined set of signals, pull the relevant detail, and present it as a short research note per prospect.

What automation should not do is decide what is worth saying. A model given free rein to write each message produces text that is fluent, plausible, and frequently wrong in ways that are embarrassing in a first touch — mistaking a company for a competitor with a similar name, congratulating someone on a round that was actually a down round, or referencing a post they were quoted in critically.

The pattern that holds up is narrower: automation gathers signals and drafts within a structure a human approved for that segment, and a human reviews anything that will be sent to a named account you actually care about.

## The connection request is a different problem

On LinkedIn, personalisation has to survive a much harder constraint before the sequence even starts. A connection request has very little room, and its only job is to earn acceptance — not to sell, not to explain the offer, not to book anything.

The most reliable requests state the specific reason for connecting and stop. Founders consistently over-write here, trying to compress the pitch into the request, which converts worse than a plain sentence that reads like it came from a person.

Acceptance rate is also the number that tells you fastest whether the targeting is right. If acceptance is poor, the problem is almost never the request text — it is that you are approaching people with no reason to know who you are, which is a segmentation issue.

## What the sequence should actually do

Once someone accepts, the temptation is to pitch immediately. The better sequences spend the first message establishing why this person specifically, then give the recipient something concrete before asking for time.

Three to four messages over two to three weeks, each with a genuinely different angle, is the shape that works. The failure mode is a sequence where message three is message one with more urgency, and message four opens with the words just following up. That is not persistence, it is a countdown the recipient can feel.

## Account safety is a real constraint, not a setting

LinkedIn restricts accounts that behave mechanically. Conservative daily action limits, varied timing rather than fixed intervals, and gradual ramping are not optimisations — they are the ceiling the channel operates under.

Being honest about this matters when planning: LinkedIn will not produce email-scale volume, and any tool promising it is describing a way to lose the account. The channel's advantage is response quality on accounts you have deliberately chosen, not reach. That is also why it pairs well with email rather than replacing it, which we cover in [running LinkedIn and email as one sequence](/blog/linkedin-email-one-sequence).

## The thing worth measuring

Most teams measure reply rate and stop. Reply rate includes the polite no, the not right now, and the please remove me, all of which look identical in a dashboard.

Track positive reply rate separately, and track it by segment rather than by message variant. That single change tells you whether your research is landing or whether you have simply found a more efficient way to reach the wrong people. We build the targeting, research, and sequencing together as one system — [get in touch](/contact) if you want to see what that looks like for your accounts.

## Frequently Asked Questions

### Does personalisation actually improve LinkedIn outreach results?

Specific personalisation does. Inserting a first name and company name does not, because every recipient has seen that pattern thousands of times and reads it as automation. What lifts acceptance and reply rates is referencing something that could only be true of that person or company.

### Can AI write personalised outreach messages?

AI is very good at the research and synthesis step and much weaker at judging what is worth mentioning. The reliable pattern is to use automation to gather and summarise signals, then use a tightly constrained template that a human has approved for the segment, rather than letting a model free-write each message.

### How many touches should a LinkedIn sequence have?

Typically three to four messages spread over two to three weeks after the connection is accepted, each adding a genuinely new angle. Sequences that run longer usually repeat themselves with escalating urgency, which reads as pressure rather than persistence.

### Is LinkedIn automation against the platform rules?

Aggressive automation risks account restrictions. The practical constraints are conservative daily action limits, human-like timing rather than fixed intervals, and no scraping behaviour that produces obvious machine patterns. Treat account safety as a hard limit on volume, not a setting to tune upward.

## Related
- [All articles](https://www.spaceblanket.ai/blog)
- [Contact](https://www.spaceblanket.ai/contact)