Affinero

Blog

How a style profile keeps AI content in your voice

GENERATED · 751 words · EN→FI ✓

Editorial illustration: metal letterpress type sorts arranged in the concentric whorl of a fingerprint

Paste a topic into any competent language model and you get back prose that is clean, grammatical, evenly paced, and completely anonymous. It reads like every other company's blog because it was assembled from the average of every other company's blog. That is the complaint people are making when they say AI content sounds like AI content.

The interesting part is that this is a briefing failure, not a model failure. Ask a good freelance writer to produce a post about your product with no brief and thirty minutes of context, and you will get something equally anonymous. The model is just faster at being generic.

A style profile is a document, not a setting

There is no toggle for voice. A style profile is a real artefact, built once from a company's existing material and revised as the archive grows. Ours are assembled by reading whatever a client already has — site copy, past posts, sales decks, and the answers their team gives in email, which are usually the most honest writing a company produces.

What goes in it is more specific than adjectives. "Professional but approachable" is not a style profile; it is a horoscope. Useful entries look like sentence-length ranges, whether the company says "we" or uses its own name, whether it addresses the reader as "you", how it handles numbers, which technical terms it uses unglossed because its readers already know them, and which industry words it refuses to use.

That last category does more work than any other. A list of banned phrases is the fastest way to make generated prose stop sounding generated.

Voice is mostly made of refusals

Most of what makes writing recognisable is what it will not do. A company that never opens with a rhetorical question, never uses "leverage" as a verb, and never ends a post with a call to action already sounds unlike ninety percent of its competitors, before anyone has said anything clever.

This is good news, because refusals are easy to specify and easy to check. Positive instructions like "be insightful" cannot be verified. "Never write a sentence longer than thirty-five words" can be, mechanically, on every draft.

We keep the refusal list longer than the aspiration list, deliberately.

The refusals also age well. A company's ambitions for how it wants to sound tend to drift with whoever is running marketing that year, but the things it would be embarrassed to publish stay remarkably stable. Ask a founder what they never want to see on their own site and you will get a fast, confident, specific answer. Ask them what they want it to sound like and you will get thirty seconds of silence followed by an adjective.

The profile has to be argued about

The failure mode of style profiles is that they get written once by whoever had time, approved by nobody, and then quietly diverge from how the company actually talks.

The version that works is the one a founder or a head of marketing has disagreed with in writing. Real objections — "we would never say that", "that is too soft", "our customers are engineers, stop explaining what an API is" — are worth more than a week of analysis, because they encode judgment that is not visible in the published archive.

We deliberately produce a sample post early and invite that argument, because the argument is the raw material.

What a style profile cannot do

It cannot give you something to say. This is the limit worth being blunt about.

A profile controls register, rhythm, and vocabulary. It does not supply the observation that makes a post worth reading — the thing your salespeople know because they have had the same conversation two hundred times. That has to come out of the company. When a blog reads as hollow despite sounding exactly right, the profile is working and the subject list is not.

It also cannot survive a company that has genuinely changed. Profiles need revisiting when the positioning moves, and a profile that has not been edited in two years is describing a company that no longer exists.

The test

Take a post, strip the logo, and show it to somebody who knows your business. If they cannot tell it is yours, the profile is not finished. If they can tell but they wince, the profile is accurate and the subject was wrong.

Both are fixable. What is not fixable is a blog that nobody would recognise as belonging to anyone.