The short version: AI content sounds the same everywhere because everyone is using the same default prompts on the same underlying models, and the fix isn't a cleverer prompt, it's building a proper voice file with real examples of your own sentences, your banned words, and your worst habits included. It takes about three hours to build once and it's the difference between content that sounds like you and content that sounds like a LinkedIn post generator set to "professional."
The tell is always the same three words
I can spot AI-written LinkedIn posts in about four seconds now, and so can you, you just haven't noticed you're doing it. "In today's fast-paced world." A rhetorical question followed by "Here's the thing." Three bullet points where the middle one restates the first. An em-dash doing the work a full stop should be doing. It's not that the writing is bad exactly, it's that thousands of small business owners are all publishing the same sentence structure because they typed a near-identical prompt into a near-identical model.
I did this myself for about four months in 2023. I was rebuilding my business after a rough patch and I leaned on AI hard for first drafts because I simply didn't have the hours. The output was fine. Grammatically clean, on topic, technically correct. And it read like nobody. A client I'd worked with for six years messaged me and said "this doesn't sound like you." She was right, and that was the moment I stopped treating prompting as the problem to solve and started treating my own voice as data I hadn't given the model yet.
Why "write in my voice" as an instruction doesn't work
Most people's attempt at fixing this is typing "write in a warm, conversational tone" into the prompt box. That instruction is almost useless because "warm and conversational" describes about eleven million websites. The model has no reference point for what makes your warm different from the next person's warm.
What the model needs is examples, not adjectives. Not "be punchy," but forty of your actual punchy sentences pasted in so it can see the rhythm. Not "avoid corporate jargon," but a literal list of the fifteen words you personally never use, sitting right there in the prompt or the project file.
This is the same principle that makes certain brands instantly recognisable no matter which channel they're on. Look at how differently Duolingo built its unhinged, slightly unhinged mascot-led voice compared with how Calm built a deliberately soft, slow, almost sparse voice for an app selling the opposite energy entirely. Neither brand got there by telling a copywriter to "sound calm" or "sound fun." They got there by defining, in writing, exactly what they would never say, and feeding every piece of content through that filter for years.
The voice file: what goes in it
Here's the document I now keep at the top of every AI content project, and I rebuild it every few months as my writing shifts. It's not a brand guidelines PDF full of adjectives. It's a working file with seven sections.
- 1. Forty real sentences. Pulled straight from things I've written, emails, LinkedIn posts, a talk transcript. Not summarised, not tidied up, copied exactly as I wrote them, typos and all.
- 2. A banned words list. Mine currently has 23 words on it including "use," "," "unlock" and "delve." I built this list by running my last 40 posts through the AI and asking it to flag every word it had also seen in 100 other business blogs that month. Humbling exercise.
- 3. Sentence length notes. I write short sentences a lot. Then a longer one. That rhythm has to be named explicitly or the model defaults to uniform medium-length sentences, which is the single biggest giveaway of AI writing.
- 4. Three bad examples. Actual AI output I've rejected, pasted in with a note saying "do not write like this." This matters more than the good examples. Models correct toward what you show them to avoid almost as strongly as toward what you show them to copy.
- 5. My actual opinions. A running list of positions I hold that most people in my field don't say out loud. Content without an opinion in it is the second biggest tell.
- 6. Specific facts only I'd know. A client name, a number, a year, a place. AI defaults to vague because vague is safe. You have to hand it the specifics.
- 7. One line about what I'm allergic to. Mine says: no em-dashes, no rhetorical questions used as a paragraph opener, no "let's dive in."
That's the whole document. Mine runs to about 900 words. I paste the relevant chunks into whatever tool I'm using at the start of a session, whether that's ChatGPT, Claude, or a client's own custom GPT, and the difference in output is immediate and obvious to anyone who knows my writing.
The uncomfortable part nobody selling AI tools wants you to hear
Here's the bit that isn't fun to say. Building this file means going back and reading your own old content closely enough to notice your habits, which is uncomfortable for most business owners. It means admitting that a lot of what you've published in the last two years was already fairly generic before AI even touched it. And it means accepting that no software feature is going to do this step for you.
Every AI writing tool now advertises some version of "trains on your brand voice in seconds," usually by scanning your website and generating a style summary automatically. I've tested six of these. They all produce the same output: three adjectives and a paragraph about tone that could describe almost any small business. The automatic version skips the exact thing that makes it work, which is your specific sentences and your specific rejections, not a machine's summary of what your writing is "like." The vendors won't tell you this because "upload your site and we'll handle it" sells better than "spend three hours doing manual curation," even though the manual version is the only one that holds up once you read the output back.
What this looks like across different types of business voice
The reason I keep coming back to how established brands do this is that it's the clearest proof the method scales past personal blogs into full companies with teams of writers. Hotjar's marketing leans on a plain, slightly self-deprecating, data-first voice that would fall apart if a writer tried to make it sound "exciting." GoPro's entire content voice is borrowed from its customers' own footage and captions, which is a version of the same trick, feed the system real raw material instead of an approximation of it. Even a utility-first product like IFTTT keeps its content voice deliberately flat and functional because that flatness is the brand, not a failure to be more "engaging."
None of these brands got their voice from a single prompt. They got it from a documented, maintained reference that every piece of writing gets checked against, which is exactly what a small business voice file is doing at a much smaller scale.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
How to build yours this week
Block three hours, ideally not on a Monday when your inbox is worst. Here's the order that works:
- Pull your last 15 to 20 pieces of writing, whatever you've got: LinkedIn posts, emails, a talk you gave, even good text messages to a client.
- Read them out loud. This sounds silly and it's the single most useful step. You'll hear your own patterns in a way you never spot by reading silently.
- Copy 40 sentences you like into a document, no editing.
- Ask an AI tool to list every generic business phrase it recognises in your own recent drafts. Turn that list into your banned words list.
- Write down three opinions you hold that you haven't said publicly because they felt too blunt. Put them in the file anyway.
- Generate one piece of content using the file and one without it, then read both back to back. The gap will be obvious within two paragraphs.
- Update the file every eight to ten weeks. Voice shifts as you do, and a stale file is almost as bad as no file.
If you run this for a client business rather than yourself, the file needs input from whoever talks to customers day to day, not just the founder. I built one for a manufacturing client last year and the most useful sentences in the whole document came from a sales call transcript, not from anything the marketing team had written.
Where this saves money
The commercial case is simple once you've done it once. A voice file built cuts editing time on AI drafts from roughly 40 minutes per post down to about 10, because you're no longer rewriting every sentence, you're tweaking a handful. Across 100 posts a year, that's the difference between 66 hours of editing and 16, which at even a modest £40 an hour is around £2,000 saved, before you count the value of content that converts because it sounds like a real person made a decision, not a system produced an average.
If you'd rather have someone build this system with you and set it up across your content, your emails and your sales scripts in one go, that's one of the core things I do as an AI implementation coach, and it's a very different job from just "using ChatGPT better."
Frequently asked questions
How long does it take to build a voice file for AI content?
About three hours the first time, done in one sitting. Refreshing it after that takes closer to 30 minutes every couple of months, mainly swapping in newer examples and removing phrases you've since retired.
Can I just ask AI to analyse my website and copy my style automatically?
You can, and most tools offer this, but it produces a generic summary rather than a working reference. It tends to output adjectives like "friendly" and "professional" rather than the specific sentences, banned words and rejected examples that change the output.
Does this work for a whole team, not just one founder's voice?
Yes, but the file needs input from more than one person, ideally whoever writes customer emails and does sales calls, not only the person who runs the marketing account. A single founder's voice file built without that input tends to sound right only in the founder's own posts.
How is this different from a normal brand tone of voice document?
A tone of voice document usually describes writing in adjectives, things like "confident, warm, direct." A voice file for AI content is built from real sentences, real rejected examples and a literal banned words list, because AI models respond to concrete examples far better than they respond to abstract descriptions.
Related reading: Why Your AI Meeting Notetaker Might Be Breaking the Law (And Killing Your Sales Calls) and Monzo Marketing Strategy: How They Built a Brand That Wins.
