Straight answer: Your AI content sounds generic because you’re feeding it generic instructions, usually a brand voice document full of adjectives like “friendly, professional, approachable.” Feed it your own old emails, transcripts and posts instead, the messy real ones, and it starts sounding like you within a day. No new tool required.
Worth reading next: Why Your AI Content Sounds Like Everyone Else’s (And the One-Page Fix).
The brand voice document is the problem, not the solution
I sat through a call last spring with a client, I’ll call her Jenny because that’s not her name, who runs a five-person accountancy firm in Leeds. She’d paid a branding agency just under £4,000 for a “tone of voice guide.” Twenty-two pages. Words like “warm but authoritative” and “clear, never jargon-heavy.” She’d handed that document to ChatGPT and asked it to write her monthly client newsletter.
It came back sounding like every other accountancy newsletter on earth. Competent. Bland. Forgettable. The kind of thing you’d write if you’d never met a client in your life.
Here’s the bit nobody selling brand voice workshops wants to say out loud: adjectives don’t teach a machine how you sound. “Warm and approachable” tells the AI nothing about the fact that Jenny always opens her emails with a one-line joke about the weather, that she uses “look” as a verbal tic before delivering bad news, that she never uses the word “solutions.” A tone document is a summary of your voice written by someone else, in their words, not yours. Feeding that to an AI is like describing a friend’s laugh to a stranger and expecting them to do an impression.
What I fed it instead
I asked Jenny to send me everything she’d written herself in the last year that hadn’t been touched by an agency or a copywriter. We ended up with:
- 38 of her own client emails, the ordinary ones, not marketing emails
- Transcripts of two webinars she’d run, pulled from Zoom’s auto transcript
- Fourteen LinkedIn posts she’d written herself before she gave up posting
- Her actual voicemail greeting, typed out word for word
That’s roughly 16,000 words of raw, unedited Jenny. I pasted the lot into a single document and gave it to ChatGPT with one instruction: “Study how this person writes. Note sentence length, the words they repeat, what they never say, how they open and close. Then write the next newsletter in that exact voice, on this topic.”
First draft, ninety minutes of work total, sounded closer to her than the £4,000 document had managed in three months of trying.
The actual step by step
- Gather 10,000 to 20,000 words of your own unedited writing. Old emails, WhatsApp voice note transcripts, Slack messages to your team, comments on your own posts. The rougher the better.
- Paste it into one document, or upload it as a file if you’re using a tool that supports that.
- Ask the AI to reverse-engineer your voice first, before you ask it to write anything. Get it to list sentence patterns, filler words, what you never say.
- Check its list against reality. Correct it out loud, “no, I never say ‘in today’s world’, delete that.”
- Only then ask it to draft something new, referencing the voice notes it just made.
- Save that voice profile as a standing prompt or a custom instruction so you’re not rebuilding it every time.
That last step matters more than people realise. Most people who try this once, get a good result, then forget to save the profile and go back to typing “write this in a friendly professional tone” a week later. The whole point is that the profile persists.
What changed once we did this
Jenny’s newsletter open rate had been sitting at 19%, which is respectable but not exciting for a list of existing clients. Three newsletters written from the voice profile later, it was at 31%. Nothing else changed. Same list, same send time, same subject line format roughly. The only variable was that the content stopped reading like a template.
Her LinkedIn posts, which she’d stopped writing because they felt like a chore, started getting the kind of comments that only happen when people recognise you in the writing: “this is so you, Jenny.” That’s the tell. Nobody ever comments “this sounds so branded.”
If you want to see what a brand built and defended its own voice looks like over years rather than months, the way Airbnb built a brand that wins is worth studying, they’ve never sounded like a generic travel company, and that consistency is a deliberate choice made post by post, not a document filed away after a workshop.
Where this gets uncomfortable
Here’s the part most posts about “AI brand voice” skip over. If your real, unedited writing is boring, this trick will produce boring content, just more efficiently. AI can copy your voice. It cannot fix a voice that has nothing to say. I’ve had clients send me their raw emails expecting magic, and the honest feedback was that the emails themselves were the problem, hedged, over-formal, allergic to a strong opinion. No amount of voice-matching prompting fixes that. It just makes the blandness scale faster.
The other uncomfortable bit: this only works if you’re willing to let the AI see your writing at its worst. People are precious about their raw drafts, the typos, the run-on sentences, the bit where you trail off mid-thought. That mess is exactly the data the model needs. Sanitise it first and you’re back to feeding it a brand document in disguise.
There’s also a limit on how far you can push this before it stops being useful. If you try to make one voice profile cover your formal client proposals, your casual Instagram captions, and your investor updates, you’ll get something that’s a compromise between all three and fits none of them well. I keep three separate voice profiles for my own writing: one for long-form posts like this, one for LinkedIn, one for client emails. They share a family resemblance but they’re not the same document.
Where people usually go wrong applying this
The most common mistake is skipping the “check its list against reality” step. People paste in their writing, ask for a draft, and accept whatever comes out without correcting the model’s assumptions first. If you don’t tell it ” I never use exclamation marks, delete every one,” it’ll happily invent quirks that aren’t yours and you’ll end up sounding like a slightly-off impression of yourself.
The second mistake is doing this once and never updating the profile. Your writing changes. Six months after Jenny built hers, her firm had picked up two new junior staff and her emails had shifted, more delegation, less first-person “I’ll sort this.” The voice profile needed refreshing. Treat it like a living document, not a one-off project.
And a smaller point that trips people up: this works best on platforms where you can save custom instructions or build a persistent assistant, rather than starting from a blank chat every time. If you’re rebuilding the context from scratch in every conversation, you’ll get inconsistent results and conclude the method doesn’t work, when really the setup was wrong.
Why this matters more now than it did two years ago
Readers and clients have got noticeably better at spotting AI-flattened writing, the same rhythm, the same three-item lists, the same “in today’s fast-paced world” openers. That’s not paranoia, it’s pattern recognition, and it happens because so much AI content is written from the same generic prompt: “write a professional blog post about X.” Everyone using that exact instruction produces siblings of the same piece.
The businesses that stand out right now aren’t the ones avoiding AI, they’re the ones feeding it something specific enough that it can’t produce the average version. Look at how brands with a distinct point of view get built, the way Figma built a brand that wins on a opinionated, unmistakable voice rather than committee-approved neutrality. That’s the target. Not “does this sound professional” but “could a competitor’s AI have written this exact sentence.” If the answer is yes, you’ve fed it a generic brief.
A quick word on tools versus prompts
People assume the fix here is a better tool, a paid plan, a “brand voice AI” product. It isn’t. I’ve done this with the free version of ChatGPT and with a basic Mailchimp draft as the output, and it worked exactly the same as it does on paid platforms with fancier memory features. The variable that matters is the input document, not the subscription tier. Spend your money on time to gather good raw writing, not on the tool that processes it.
If you’d rather someone build this system for you and set it up across your actual channels, that’s the kind of hands-on work I do with clients through AI implementation coaching, but you don’t need to hire anyone to try the basic version this week.
How to know it’s working
Forget engagement metrics for the first test. Send the draft to someone who knows you, a business partner, a long-standing client, your other half, without telling them AI wrote it, and ask “does this sound like me.” That single question is more useful than any open rate. Jenny’s business partner read the third AI-assisted newsletter and asked her when she’d found time to write it herself. That’s the bar.
Once you’ve cleared that bar, the metrics tend to follow. Comments that mention you by name rather than the topic. Replies that continue the conversation rather than just reacting to it. People forwarding the thing to a colleague with “you should read this, it’s exactly how they talk.” Brands that get talked about this way tend to have built an unmistakable identity over a long stretch, the same way a company like Instagram grew a brand people recognise instantly, through relentless consistency rather than a single clever campaign.
Frequently asked questions
Do I need a paid AI tool for this to work?
No. A free ChatGPT account handles the whole process, from analysing your writing to drafting new content in your voice. The result depends on the quality and volume of your own writing that you feed it, not on the subscription tier.
How much of my own writing do I need?
Aim for 10,000 to 20,000 words minimum, roughly 30 to 40 emails or a couple of transcribed calls plus a dozen social posts. Less than that and the model has too little pattern to work from and will fall back on generic phrasing.
What if my own writing is honestly a bit dull?
Then the AI will produce dull content faster, because it copies patterns rather than adding personality that isn’t there. Fix the source writing first, get a bit more opinionated and specific in your real emails and posts, then rebuild the voice profile from that improved material.
How often should I update the voice profile?
Refresh it every four to six months, or sooner if your role, team or audience shifts noticeably. Voice drifts as your business changes, and a profile built a year ago on old writing will slowly stop matching how you talk now.
Related reading: The AI Notetaker Sitting In On Your Client Calls Might Be Breaking Your NDA and The AI Receptionist Test Every Small Business Owner Should Run Before Launch.
For the bigger picture, see my full guide to AI marketing.
