The short version: Feeding an AI tool your old blog posts, emails, or podcast transcripts does not give you your own voice back, it gives you a smoothed-out average of what you’ve already said, and everyone else training their AI the same way ends up in the same bland middle. The fix isn’t a better prompt, it’s feeding the AI your actual opinions, arguments, and specific stories before you ask it to write anything.
Worth reading next: How to Build a Content Calendar for 2026.
What happened when I tried the thing everyone recommends
Eighteen months ago I did what half the AI courses tell you to do. I took four years of blog posts, about 340 of them, and fed chunks into a custom GPT so it could “learn my voice.” I asked it to write a LinkedIn post about pricing your services. What came back was fine. Grammatically clean, on topic, structured with a hook and three points and a call to action.
It also sounded like nobody. Not like me, not like anyone. It had the rhythm of my writing without a single opinion I’d fight for. No sentence in it would have made anyone stop scrolling, because it had sanded off every rough edge I’d ever put into a real post, the bits where I say something a bit blunt and someone disagrees with me in the comments.
That’s the bit nobody selling “AI brand voice training” tells you. The model isn’t learning your voice, it’s learning your average sentence length, your favourite transition words, your typical paragraph structure. It cannot learn your opinions from your content unless your content is already full of stated opinions, and most business blogs, mine included in some years, are full of safe, general advice with the actual view buried or left out entirely.
The uncomfortable bit: this is happening at scale, and most people can tell
I went through 40 LinkedIn posts from small business owners in my own niche last spring, all posted within the same fortnight, all clearly AI-assisted. 27 of them opened with some version of “Here’s a mistake I see all the time” followed by three bullet points and a question at the end. Not because these people copied each other. Because they were all trained on the same base models, prompted with the same “write like an expert in X” instructions, and none of them had fed the tool anything their own first. This is the same trap brands have always fallen into when they chase a formula instead of a point of view, the kind of thing I’ve written about when looking at why the Headspace marketing strategy worked while a dozen meditation apps copying its tone quietly disappeared. Headspace didn’t win because the words were soothing. It won because the founder, a former Buddhist monk, had a specific way of explaining anxiety that nobody else in the category had. That specificity is exactly what gets lost when you train AI on generic content and ask for generic output.
Readers and buyers can feel this even when they can’t name it. Open rates drop, comments dry up, and the content still “performs” on paper because it hits word count and SEO boxes, so the business owner assumes the strategy is working when it’s just cheap. Cheap to produce, cheap to read, cheap to forget.
Why “matching your brand voice” and “being specific” are not the same job
Most AI writing tools sell brand voice as a tone problem. Formal or casual, short sentences or long, British or American spelling. That’s styling. It’s the equivalent of dressing every business in the same three outfits and calling it personality. Real distinctiveness comes from content, not style. It’s the specific number you’re willing to say out loud when a competitor won’t (I charge £3,500 for a half-day AI workshop, most consultants at my level won’t tell you a figure at all). It’s the specific story you tell about the client who ignored your advice and lost six months. It’s the specific, slightly awkward opinion, like the fact that I think most “AI content calendars” are a way for agencies to bill you for volume instead of results. None of that lives in your old blog posts unless you deliberately put it there. So when you train a model on content that never had that specificity, you cannot prompt your way to it afterward. Garbage in, smooth and forgettable garbage out.
The five-step fix, in the order I use it now
- Step one: write a “positions document” before you touch any AI tool. One page, 10 to 15 statements you’re willing to argue for. Mine includes lines like “most small businesses don’t need a content strategy, they need one good story told across five channels” and “AI consultants who won’t quote a price on the phone are hiding something.” These are the raw material, not the finished writing.
- Step two: collect your actual stories, not your published ones. The client email exchange that went badly. The pitch you lost and why. The number that surprised you. I keep a running document, currently 47 entries, of things that happened in client work that never made it into a blog post because they felt too small or too specific. Those are the ones that make writing sound human.
- Step three: feed the AI the positions and stories, not the finished blog posts. This is the reversal that changes the output. Instead of “learn my style from these 340 posts,” it’s “here are 15 opinions I hold and 20 things that happened to me, help me turn these into a post about X.” The model still does the structural work, but the substance is yours.
- Step four: keep one number, one name, or one specific detail in every piece of AI-assisted content. Not “many clients struggle with this.” “A client in Tel Aviv spent £4,000 on Facebook ads in a month with a 0.4 percent conversion rate before we fixed the landing page.” Specificity is the thing generic AI output is structurally incapable of producing on its own.
- Step five: read it out loud before you post it. If you wouldn’t say it in a room to a client’s face, cut it. This single check has stopped me publishing more AI-drafted paragraphs than any prompt engineering trick I’ve tried.
Where this costs you money, not just engagement
The real damage isn’t a quiet LinkedIn post. It’s the emails, the sales pages, the proposals. I’ve seen small businesses send AI-drafted proposals to prospects that read identically to three other proposals that same prospect got that week, because everyone’s tool defaults to the same structure: problem, solution, three pricing tiers, urgency line. Buyers who are shopping around notice this fast, and it quietly signals “we didn’t think hard about you specifically,” which is the opposite of what a proposal is meant to do. This is where working with someone who builds the process with you, rather than just handing you a tool, tends to pay for itself. When I work with businesses through AI implementation coaching, the first two sessions are almost never about prompts. They’re about pulling out the positions and stories that make the business distinct, because that’s the part a tool can’t generate for you no matter how good the model gets.
A test you can run this week
Take the last five pieces of AI-assisted content you published. Cover the company name and read them cold. Could you tell they came from your business specifically, or could they have come from any competitor in your space? If you’re honest, most people find at least three out of five could belong to anyone. That’s not an AI problem, it’s a raw material problem, and it’s fixable in a week if you do the positions document. Compare it to how Disney’s marketing strategy has stayed distinct for a century even as every tool and channel around it changed completely. The stories and the specific characters never got outsourced to a formula. The delivery mechanism changed, the substance didn’t. That’s the model for using AI well: let it change how fast you produce, never let it decide what you have to say.
Where interactive and behavioural thinking helps too
One thing that works alongside a stronger written voice is giving people something to do rather than just something to read. I’ve written before about how interactive calculators convert visitors far better than static pages, and the same logic applies here. A distinct, specific piece of writing paired with a small interactive tool tends to outperform either one alone, because the tool proves you understand the specific numbers in someone’s situation rather than talking in generalities. There’s also a behavioural angle worth borrowing from Richard Thaler’s work on how people make decisions rather than how we assume they do. I covered some of this in the business lessons from Richard Thaler, and the relevant point here is that people default to whatever’s easiest to process. Generic AI writing is easy to process and easy to forget in the same breath. Specific writing takes half a second longer to read and stays with people for days.
The part that’s hard to hear
If your content has gone generic since you started using AI, the honest reason usually isn’t the AI, it’s that your content was already thin on opinion before you automated it, and the tool just made the thinness faster to produce. AI doesn’t remove your voice. It removes the need to have one, unless you insist otherwise. Coco Chanel built an entire empire on refusing the industry default at every turn, which I wrote about in the business lessons from Coco Chanel, and the throughline is that she never let efficiency replace judgement about what made the brand hers. That’s the same choice every small business now makes every time they open a chat window and ask for a draft. Use the tool for speed. Never use it for substance. The moment those two things swap places is the moment your marketing starts sounding like everyone else’s.
Frequently asked questions
Does training AI on my own website content stop it sounding generic?
Not on its own. It copies your sentence structure and word choices, not your actual opinions or specific stories, so if your original content was already general, the AI output will be a smoother, faster version of that same generality.
How much time does it take to build a distinct AI content process?
Building a proper positions document and story bank takes most business owners two to three hours the first time, then about 20 minutes a week to add new material. That upfront time is what makes every piece of AI-assisted content afterward sound like it came from a specific person rather than a template.
Is it worth paying an AI consultant just to fix voice and tone?
It’s worth paying someone if they start by pulling your specific opinions and stories out of you rather than jumping straight into prompt writing, because tone alone won’t fix generic content, and most people can’t see their own blind spots on this without an outside eye.
What’s a quick way to check if my AI-written content is too generic?
Cover your company name on your last five pieces of content and ask whether a stranger could tell it came from your business specifically, or from any competitor in the same space. If most of them could belong to anyone, the problem is the raw material you fed in, not the AI tool itself.
Related reading: Your Business Now Has an AI Reputation, and You Probably Haven’t Checked It and Starbucks Marketing Strategy: How They Built a Brand That Wins.
