If you use AI to write content and everything sounds the same, the problem is rarely the tool: it’s that you’re asking for a text without giving it your voice. Brand voice survives with a system made of four pieces: a voice guide with real examples, a brief before every draft, your judgement and experience inside the text, and a final edit with a checklist. Here’s how I set it up.
Why AI ends up sounding the same for everyone
A language model gives you, by default, the most likely version of a text. Ask for “an article in a friendly, professional tone” and you get what hundreds of brands have asked for before. And you’re not the only one doing it: according to HubSpot’s 2026 State of Marketing report, 42.5% of marketers use AI extensively for content creation and another 38% use it occasionally. At the same time, 62.7% believe brands will need more unique, human-centred content to compete with AI-generated material.
In other words, producing more no longer sets anyone apart. What does is what AI doesn’t have on its own: your judgement, your experience and your way of telling things.
Step 1: document your voice with examples, not adjectives
“Friendly, direct and professional” tells a model nothing, just as it tells a new copywriter nothing. What works is a short guide (one or two pages) with concrete material:
- Three to five of your own texts that show how you really write, so the AI imitates examples rather than descriptions.
- Do and don’t phrases: pairs of “I’d say it like this” and “I wouldn’t say it like this” for the same idea.
- Banned vocabulary: the filler words and grandiose adjectives you never want to see.
- Form rules: grammatical person, paragraph length, how you open and how you close.
This guide lives in one fixed place (a document, a Claude project with instructions, whatever you use) and is always loaded. Every time you fix a draft and think “I wouldn’t say that”, that sentence goes into the guide. Over time it becomes the most valuable part of the system.
Step 2: write the brief before asking for the draft
Most generic texts come from generic requests. Before generating anything, fill in a fixed brief. It helps me keep the thinking on my side and the drafting on the AI’s:
| Brief field | What it should include |
|---|---|
| Main question | The specific doubt the text solves, in one sentence. |
| Reader | Who reads it, what they already know and what decision they need to make. |
| Thesis | Your position on the topic. If you don’t have one yet, it’s not time to write. |
| Your own input | A real example, judgement or data point only you can add. |
| Sources | Reliable links the text may use. Nothing outside that list. |
| Structure and length | Sections, format (table, checklist) and word count. |
Step 3: AI structures and drafts; you add what only you have
With the guide and the brief loaded, I ask for a first draft and treat it as exactly that: a draft. AI is very good at organising ideas, proposing structure and polishing prose. It is not good at inventing experience that doesn’t exist, and it shouldn’t. That’s why I write (or dictate) the “your own input” part myself and let the AI weave it in.
Every fact needs checking too. Google is clear about it in its guidance on generative AI content: generative models don’t retrieve facts, they predict likely sequences of words, so information must be verified. The same page reminds you that using AI to generate many pages without adding value for the reader may violate its scaled content abuse policy. Using AI isn’t the problem; publishing without judgement is.
Step 4: edit with a checklist, not with a hunch
Reading a text and thinking “hmm, looks fine” is not quality control. I review every piece against a short checklist before scheduling it:
- Does it answer the main question in the first lines?
- Is there a clear position, or just a disguised “it depends”?
- Does every fact have a source, and have I opened it myself?
- Is there something only I could have written?
- Are any banned words or brochure-sounding phrases left?
- Could I read it aloud without stumbling?
Whatever comes out of the edit goes back into the voice guide. That way the system improves with every piece instead of repeating the same mistakes.
What to automate and what not to
My rule is simple: I automate what has no voice and keep what does. Turning an article into a social summary, proposing headline variants, organising notes into an outline or adapting a text to another language are tasks where AI saves a lot of time. The thesis, the examples, the tone of the opinions and the decision about what to publish stay with me. If you want to see how to chain tasks like these, I wrote about automations with Make and Claude.
Frequently asked questions
Can I keep my brand voice if I use AI to write?
Yes, if you give the AI material from your voice (reference texts, do and don’t phrases, banned vocabulary) and edit the result with judgement. With a generic tone instruction, the output will be generic.
Does Google penalise AI-written content?
According to its own documentation, using AI isn’t prohibited; what can violate its policies is generating content at scale without adding value for users. What matters is that it’s accurate, useful and reviewed by a person.
How much should I edit an AI draft?
Enough for the thesis, the examples and the tone to be yours. If, at the end, nothing in it is something only you could have written, it needs more editing.
How do you protect your brand’s voice when you write with AI? Tell me in the comments, or write to me through /en/contact/ if you’d like to share how your team does it.
