Part four of the series AI Content: 10 Strategies That Work. A new part every Friday.
In practice, personalisation usually means dropping a name and an industry into the text. That is addressing, not personalisation.
Addressing is not personalisation
Inserting a first name and a company name into an email is something a system can do while knowing nothing whatsoever about the recipient. The text stays the same for everyone, only two words change in it. The reader notices, because the rest of the message says nothing about their situation.
It is more useful to work with what the reader is dealing with at that moment, what part of it stings, and what they do not believe. Those three things can be described in a brief far more precisely than demographics, and they have a much larger effect on the result.
The difference shows in two briefs. „Write for middle-aged women“ versus „write for someone who has already tried two similar solutions and been let down by both, so they are watching for the catch from the first sentence and will appreciate you naming it yourself“.
The second brief points the model towards a completely different text. The tone changes, the order of the arguments changes, and so does the opening. And notice that it contains not a single demographic detail.
What belongs in a description of the audience
An audience description worth handing to a model contains five things. What that person is doing at the moment they run into your content. Which word they use for the thing and which word you use. What specifically they are afraid of. What they have already tried. And which objection stops them.
The last item is the most valuable and at the same time the easiest to get hold of, because you already have it in customer emails, in comments, in call transcripts and in sales notes. You do not have to invent it, you just have to copy it out.
A text written against a real objection from a real email is incomparably stronger than a text written against an invented persona. An invented persona tends to raise objections that are easy to refute, because you invented them yourself.
A practical exercise that takes an afternoon: take the last twenty sales emails, write out the sentences where the customer hesitates, and put those into the brief verbatim. The model gets material it has nowhere else, and you can tell from the result immediately.
Where personalisation hits its ceiling
The more you tailor content to what the audience already believes, the more you confirm what they already think. In the short term that raises engagement, because agreement is pleasant to read.
In the long run it turns you into a source that moves nobody anywhere, and therefore a source nobody shares and nobody remembers. Content that only confirms works like a mirror, and nobody recommends a mirror to a friend.
A practical guide: in a text meant to move someone, start where the reader is standing but finish one step away. Not ten steps, because the reader will refuse. One.
The tokenisation tax English-language guides stay quiet about
Personalisation in a language other than English carries a technical cost you will not find in English-language handbooks. A May 2026 study of tokenisation across 24 European languages measured that Czech needs on average 2.28 tokens per word, while English needs 1.23. That is roughly 1.86 times as many, and Czech is far from the worst case in that sample.
In practice this means three things. A prompt and an output in such a language cost you almost twice as much through an API as their English equivalents. You exhaust the context window faster, so less source material fits into a single brief. And you hit subscription limits sooner, which shows up precisely when you are using the tool at full tilt.
For a one-off article none of this matters. For an automation generating thousands of personalised variants it is a budget line you need to calculate in advance rather than discover on the invoice.
The way around it is to run the structure and the research in English and switch to the target language only for the final text. At the cost of the result then sounding like a translation, which is a different problem and one we will return to in the part on the division of labour between a human and a model.
How to check your personalisation
The cheapest quality test for a personalised text is to read it through the eyes of the person it was written for and look for the places where they would say „that does not apply to me“. Every such place is either for cutting or for rewriting.
The second test is even simpler. Try swapping the audience description for a different one and read the text again. If it works just as well, it is not personalised. It is written generically with a name added.
The third test belongs to the sales team. Give them the finished text and ask whether they would send it to a specific customer they are currently negotiating with. The answer „I would send it, but I would cut this part“ is the most valuable feedback you can get on a piece of content.
Previous part: The prompts that make a difference, and the fact-checking without which none of it matters. Next Friday: Visibility in the year people stopped clicking.
The whole series: AI Content
1. Why most advice about AI content no longer holds · 2. A publishing plan that assumes volume does not work · 3. The prompts that make a difference, and the fact-checking · 4. Personalisation that is more than a different salutation · 5. Visibility in the year people stopped clicking · 6. Where AI ends and a human begins · 7. Images, and how to spot a marketing myth · 8. Measurement that tells you something · 9. One piece of content, several channels, without copy-paste · 10. Law and ethics, specifically, as they stand in August 2026 · 11. Build a system, not a pile of articles · 12. What to take from the whole series