Sound Like Somebody

Adobe went and asked 850 American small business owners a question I've been chewing on for two years, and the answer they got back is the most useful piece of marketing research I've seen all year.

The question, roughly: when everybody has access to the same software, what's left to compete on?

More than half said taste. The sharper eye. Not budget, not tooling, not volume. Judgment.

The rest of the numbers in that survey, run with Morning Consult, are where it gets interesting. Seventy nine percent said a clear point of view is what makes people follow them. Nearly one in three said they'd rather post nothing at all than post something that sounds off brand. Sixty six percent are already using AI for marketing.

And then the number that made me put my coffee down. Only six percent published their last AI draft untouched. Forty eight percent reworked it heavily.

Six percent. Ninety four out of a hundred owners took what the machine gave them and changed it before it went out. Half of them changed it a lot.

That's not a story about AI being bad at writing. AI is fine at writing. That's a story about a whole population of business owners discovering, independently and without anyone telling them, that the default output doesn't sound like them, and that sounding like them is the entire point.

So today I want to talk about how to build that discovery into a process instead of rediscovering it every time you open a blank page and feel vaguely dissatisfied with what comes back.

Quick context, because it changes the stakes.

Anthropic started watermarking Claude's text output on August 2, worldwide, to satisfy the EU AI Act. Substack teamed up with Pangram to flag AI generated content on its platform last month, and its CEO Chris Best publicly called out what he called Claudefishing, which is a genuinely good coinage for passing off machine writing as your own. Suno said it'll mark tracks made on its platform after a run of legal challenges. The infrastructure for detecting machine involvement is getting built out fast, from several directions at once.

You could read all that as a threat. I read it as a market signal.

Here's the thing. Detection is not the reason undifferentiated content fails. Undifferentiated content was already failing. It failed before anybody could detect anything, because readers have always been able to tell the difference between someone who knows something and someone assembling sentences about a topic. Detection just makes the failure legible faster.

Which means the strategy doesn't change. It just gets more urgent, and the payoff for doing it right gets bigger, because the pile of forgettable content you're competing against is growing every day.

Let me get concrete about what "sounds like you" actually means, because it's a squishy phrase that people nod along to and then can't act on.

There are exactly four categories of material that no model can generate, no matter how good the prompt, because they don't exist in the training data. They exist in your life.

Numbers from your own operation. Not industry benchmarks. Yours. What your average project actually costs to deliver, how long your sales cycle really runs, what your churn looked like the year you raised prices. A model can tell your reader what the industry average is. Only you can say "we ran this for eleven months and here's what came out the other end."

Names and specifics. The client who called on a Saturday. The vendor who missed the deadline. The town, the number, the exact thing that was said. Specificity is unforgeable, and it's the fastest way to prove you were there.

Scars. The thing you tried that didn't work, described honestly, including the part where you were wrong for longer than you should have been. This is the single most underused asset in small business content and I think I know why. It feels like weakness. It reads as credibility. Every reader has been wrong about something and they trust people who admit it.

Opinions with a cost. Not "consistency is important." An actual position that would annoy somebody, that you'd defend in a room, that costs you a slice of the audience who disagrees. If your content couldn't offend anyone, it also can't convince anyone. Those are the same muscle.

Now here's the move. Those four things become your inputs, not your edits. You don't write a machine draft and then sprinkle personality on top. You supply the four things first, and let the machine handle structure, transitions, and cleanup around them.

That inversion is the whole method. Most people use AI to generate the substance and then fix the voice. Do it backwards. Supply the substance, let the machine do the joinery.

Telling a model to "write in my voice" produces a caricature. It'll grab three surface tics and turn them up until you sound like a parody of yourself. You know this because you've tried it.

What works is specimens plus rules.

Pull five pieces of your own writing that you genuinely like. Not your most successful posts, your favorite ones. There's a difference and the favorites are more diagnostic. Emails count. A long text to a business partner counts. Anything where you were writing fast and not performing.

Read them back and write down what you notice about the mechanics. Actual mechanics, not vibes. Sentence length and how much it varies. Whether you open with a story or a claim. Whether you use questions. How you handle transitions. Whether you use the reader's name, or "you," or neither.

Then write down the rules. Mine include: never open with a definition. Contractions always. One idea per paragraph and paragraphs stay short. No word that I wouldn't say out loud to a person. Kill any sentence that could open somebody else's article.

That's your voice file. Specimens plus mechanics plus rules. It goes at the top of every content prompt you write, and it does about eighty percent of the work that "write in my voice" fails to do.

Update it quarterly. Your voice moves, and a file from a year ago will pull you backward.

Here's how a piece actually gets made in my shop, start to finish.

I supply the raw material by talking. Not typing. Voice memo, five to eight minutes, no structure, just what I actually think about the thing including the parts I'm unsure about. That recording contains all four unforgeable categories because I'm just talking and that's how people talk. The numbers come out, the client story comes out, the thing I got wrong comes out, and the opinion comes out with the heat still on it.

The machine gets the transcript, the voice file, and one instruction: organize this, do not add to it. That constraint matters more than anything else in the prompt. The moment a model starts adding, it adds generic material, because generic material is what it has when it doesn't have you.

What comes back is a structured draft made entirely of my own material. It's usually about seventy percent there.

Then I do the last thirty percent by hand, and it's always the same three passes. First pass, I put back the thing the model smoothed out, because there's always one blunt sentence that got sanded into something polite. That sentence was the point. Second pass, I read it out loud and fix every place my mouth stumbles, because if I can't say it, nobody can hear it. Third pass, I cut the weakest fifteen percent, which is almost always in the middle where I was explaining something the reader already understood.

Total time, about forty minutes for something like this. And the finished piece is mine in every way that matters, including legally, including ethically, and including the way that actually counts, which is that a reader who's read me before would recognize it in a lineup.

I want to spend a minute on the read aloud pass, because when I tell people about this process it's the step they skip, and it's the step that does the most work.

Machine writing has a specific failure mode that's hard to see on the page and impossible to miss in your mouth. The rhythm is too even. Sentences arrive at a consistent length. Clauses balance. Every paragraph resolves neatly. It reads like a metronome and human speech is nothing like a metronome.

Real talking has short punches. Then it runs long because the person got going and didn't want to stop and the thought kept unspooling past where a careful writer would have put a period. Then it stops.

You cannot fix that by asking a model to vary its sentence length. I've tried every version of that prompt and what you get is variation that's itself uniform, which is somehow worse. You fix it by reading the thing out loud and letting your own breath tell you where it's wrong.

Do it standing up. I have no defensible reason for this beyond that it works better and I've stopped questioning it.

I don't want this to read as "AI can't help with content," because that's not what I'm saying and it would be a waste of a tool that's genuinely good.

Here's where it's excellent, no reservations. Structuring a mess into an outline. Generating fifteen headline options so you can pick one and reject fourteen. Cutting a long piece down to a specific word count, which it does better and less painfully than I do. Adapting one piece into five formats for different platforms. Catching the sentence you wrote three times in slightly different words. Building the tedious connective tissue between two sections you care about.

And the distribution layer, which is pure logistics and has no voice in it whatsoever. Scheduling, formatting per platform, timing. I run mine through Buffer and the whole thing takes about ten minutes a week now, which is ten minutes I used to spend copying and pasting like an idiot.

The rule that governs all of it is simple. The machine does the parts that would be identical no matter who ran the business. You do the parts that only make sense because it's your business.

Formatting is identical for everyone. Your opinion about your industry is not.

Come back to the six percent for a second.

Ninety four percent of owners edit their AI drafts before publishing. Almost half rework them heavily. That's not a small tweak on the way out the door. That's a population of people who have all independently concluded that the raw output isn't good enough to carry their name.

They're right, and they figured it out without a framework, just by looking at what came back and feeling something was off.

What I'd add is this. That instinct is expensive if it stays an instinct, because you burn it on every single piece, fixing the same problems over and over with no system. Turn it into a process and it costs you once. The voice file, the four categories, the three passes. Build them once and the instinct becomes infrastructure.

And one more thing about the nearly one in three who said they'd rather post nothing than post something off brand. I have complicated feelings about that. The discipline is admirable. But silence isn't a strategy, and I suspect a lot of that group is sitting on good material they've talked themselves out of publishing because it didn't feel polished enough.

Polished is not the bar. Recognizable is the bar. Publish the rough thing that sounds like you over the smooth thing that doesn't, every time, and it isn't close.

This week, do one thing. Record a five minute voice memo about something you actually have an opinion on in your business. Don't script it. Transcribe it, hand it to a model with instructions to organize and not add, and see what comes back.

I think you'll be surprised how much of the good stuff was already in your mouth.

See you tomorrow.

Jordan

The AI Newsroom is written by Jordan Hale. This issue contains affiliate links to tools I actually use. If you sign up through them I may earn a commission at no extra cost to you.