Why AI copy sounds the same across every brand

Try this on your own category. Open five competitor homepages, cover the logos, and read the first screen of each aloud.

I have done this in eyewear, in outdoor gear, in skincare and in coffee, and the result is the same every time. Three of the five are interchangeable. Usually there is a promise about quality, a promise about ease, and an adjective pair doing the emotional work: confident but approachable, premium yet accessible, considered, never complicated.

None of it is badly written. That is the problem I want to take apart, because "AI writes bad copy" is not true and it lets everyone off the hook. AI writes good copy. Good, in the sense of correct, fluent and appropriate. It is the appropriate that costs you.

[Internal link: this is the companion to Lab Note 01 - what AI can and can't do in brand strategy. That piece draws the line. This one explains why the line is there.]

Why AI content sounds generic: probability, not laziness

A language model works by predicting likely continuations. Given everything it has read and the words so far, it picks what most plausibly comes next, over and over.

Sit with what "most plausibly" means when the words so far are "write homepage copy for a premium sunglasses brand".

The likely continuation is the one that has appeared most often in that context across an enormous amount of text. Which is to say: the thing the most people already wrote. Probable means typical. Typical means someone got there first, and then a few thousand other people did too.

So the output is not a failure of the tool. It is the tool working exactly as designed, on a request that gave it nothing to work with. Ask for the most likely sunglasses homepage and you will get the most likely sunglasses homepage - which is, by construction, the one your competitors are also closest to.

This is worth being precise about because it changes what the fix is. If the problem were quality, a better model would solve it. The problem is not quality. Each generation of these tools is more fluent than the last, and more fluency is more typicality, which makes the sameness slightly worse rather than better.

Fluent is not the same as distinctive

Here is the trap that catches people, and it caught me for a while.

When you read AI copy for your own brand, it reads well. It flatters back. It says the things you would have said, in a slightly tidier register than you would have managed on a Tuesday afternoon. Nothing triggers the internal alarm that says this is wrong, because nothing in it is wrong.

The alarm you need is a different one, and it does not go off on its own. It is: could a competitor publish this sentence tomorrow without changing a word?

Run that test on the last thing you shipped. Most brands fail it on the hero, the About page and every product description. And failing it is not a copywriting problem - it is a business problem, because in a category where five brands say the same thing, the only remaining way to choose between them is price. You have argued yourself into a price war through the medium of nice adjectives.

What distinctiveness actually is

Three things, none of which a model can invent for you.

The specific. Not "designed for durability" but the actual condition it was designed for. Tahan Outdoors makes gear for tropical camping - heat, humidity, heavy rain - because most outdoor equipment is engineered for continental climates and fails in a Malaysian downpour. That sentence is not available to a brand in Colorado. Specificity is a moat made of facts.

The odd. The thing that would not survive a committee. Vibes Shades sells fashion eyewear on the argument that sunglasses should not carry the markup that retail shelves and distributors add. That is an unusual thing for a fashion brand to lead with - it is an argument about supply chains on a page about looking good - and it is unusual precisely because it is a real position rather than a pleasant one.

The refusal. What you will not say. This is the one everybody skips and it is the strongest of the three, which is the rest of this piece.

Three inputs a model cannot generate for you

Verbatim customer language. The actual words in your reviews, DMs and support tickets. A model can process them brilliantly once you hand them over. It cannot know them, because they were never public.

Founder judgment. Which of two true things matters more. Whether this brand will ever discount. What it would be embarrassing to claim. These are positions, and positions come from someone who has something at stake.

The banned list. The words this brand does not use, written down. This is proprietary by definition, because it is the negative space of the other two.

Constrain, don't prompt

The standard advice is to write better prompts, and I think it is close to backwards.

A brief made of adjectives - confident, premium, human, approachable - asks a model to produce the average of confident, premium, human and approachable. You have described the middle and then expressed surprise at arriving there.

A brief made of prohibitions does something different. Each ban removes a region of likely output. Enough of them and the model has nowhere left to land except somewhere specific, which is the only place your brand was ever going to be.

Never use the word premium. Never apologise for the price. The customer is a Mover. Shop is Access.

That is not a style guide. That is a set of walls, and the walls are what make the shape.

What a brand voice file actually contains

Every Lab89 client gets one before a word of copy is written. Four sections, and three of them are restrictions.

Look at the lower-right block. Overly casual or gimmicky. Artificially luxurious. Verbose. Trend-chasing. Apologetic about price disruption.

Five refusals. Between them they rule out most of what a model would produce for a fashion eyewear brand, and they took a founder about twenty minutes to decide.

Two details in that extract matter more than they look.

The LOCKED tags mean a decision was made on a date and is not reopened every quarter by whoever is writing that week. Consistency is not a personality trait, it is a filing system.

And Add to Cart stays Add to Cart. A brand that renames the checkout button has started enjoying itself at the customer's expense. Knowing where the glossary stops is part of the glossary - and it is the kind of judgment call that only a person with something at stake will make correctly.

[Template capture: the blank version of this file, the one we fill in with every new client. Four sections, roughly an hour to complete, and it makes everything downstream better - including the AI. A Founder Decode pack is coming alongside it - the prompt, the full question set and the Gap Map sheet - so you can run the vision layer on yourself too.]

The honest version of the argument

None of this is anti-AI. I use these tools every day and the expression layer is genuinely faster than it was two years ago.

But the value has moved. It is not in the generating, which is now free and which everyone has. It is in the constraining, which is slow, which requires someone to hold an opinion, and which nobody can do on your behalf.

The brands that will read as distinctive in three years are not the ones that used less AI. They are the ones that did the unglamorous work of writing down what they are not, and then held the line.

What is on your banned list? If the answer is that you do not have one, that is the whole piece in a sentence.