Personalising at scale without inventing anything

Personalisation works when the message contains something you could only know by looking, and fails when it contains something you could have merged from a spreadsheet. The difference is not effort or word count, it is whether a fact about the recipient was observed or retrieved, and readers tell the two apart instantly.

The harder half of this is the part nobody discusses: at any scale above a handful of messages, the risk stops being blandness and becomes invention. A message that confidently states something untrue about a company is worse than no message.

Why merge fields read as automation

Because they are the only variable thing in a sentence whose shape never changes. A reader who has received four messages built on the same template can see the slot, and the name sitting in it proves a list exists.

The test is mechanical: take your sentence and swap in a different company's details. If it still reads fine, it was a merge field, however specific the words looked.

What counts as evidence

Evidence is a fact that took a look to obtain and that the recipient can verify. Three tiers, roughly in order of how much they are worth:

• Something about their performance. What their posts reach, whether anyone replies, how their launch actually did. Strongest, because they usually know it and rarely hear it said out loud.
• Something about a specific artefact. A page on their site, a post they published, a job listing they wrote. Verifiable and clearly not from a database.
• Something about timing. That an announcement was three weeks ago and nothing followed it. Weakest alone, strong in combination.

What does not count, regardless of how it is phrased: industry, headcount, funding stage, location, tech stack, the founder's job title. All retrievable in bulk. All therefore proof of a list rather than of attention.

The invention problem

Every scaled personalisation system, whether it is a junior researcher at 2am or a language model, has the same failure: when there is nothing interesting to say about a company, it produces something interesting anyway.

The output is fluent, specific and wrong. It congratulates a company on a product it did not launch, references a post it did not write, cites a number from a different company. And it fails worse than a generic message, because a generic message is merely ignored while a confidently incorrect one tells the recipient exactly how it was made.

The check that stops it

Every factual claim in the message has to trace to something that was actually fetched. Not "probably true", not "consistent with what we know". A specific artefact that was retrieved.

In practice that means separating the two steps. First collect facts about the company and store where each came from. Then write using only those facts. A writing step that has access to general knowledge about a category will use it, and what it produces will be plausible rather than true.

The corollary is the rule most people will not accept: when there is nothing specific to say, the correct output is no message. A pipeline that always produces one is a pipeline that invents, because it has been given no other option.

Personalisation that is not about the recipient

One underused kind: being specific about yourself. "We have run this for four crypto projects and the reply rate landed between 8 and 12 percent" is not personalised to them at all, and it is more credible than any merge field, because it is checkable and nobody puts a real number in a template they are ashamed of.

How much is enough

One observation. Not three.

Messages containing several personalised facts read as a dossier and trigger the exact surveillance feeling you are trying to avoid. One specific, useful, correct observation is the entire budget, and the rest of the message should be about what you would do.

The summary you can apply today

Take your last outbound message. Delete every sentence that would survive being sent to a different company. If what remains is nothing, the message was a template with decoration. If what remains is one sentence that is checkably true and slightly uncomfortable to read, that was the message, and the rest was packaging.