AI can write a perfectly competent homepage in seconds, which is partly why so many websites are beginning to sound alike. Ask a language model to write copy for a project management tool, accounting platform, analytics product or recruitment company and the result will usually be polished, grammatically clean and broadly sensible. Benefits appear in the right places, calls to action sound familiar and almost every sentence feels reasonable enough to publish.

Read enough of those websites and the similarities become harder to ignore. Different companies start using the same rhythms, the same broad claims and the same frictionless language about simplifying workflows, empowering teams, unlocking insights and helping customers work smarter. Good writing and distinctive writing are not the same thing, and AI is very good at producing copy that clears the first bar without necessarily reaching the second.

AI is very good at producing the reasonable answer

Language models produce language by learning patterns from enormous amounts of existing text. A broad instruction such as “write homepage copy for my SaaS product” leaves most of the important decisions to the model, including how a SaaS homepage normally opens, which benefits deserve emphasis, which vocabulary sounds professional and how a call to action should be phrased.

Common patterns become useful answers when very little company-specific information is available. Category conventions fill the gaps because the model has seen countless examples of software companies talking about collaboration, productivity, visibility, efficiency and growth. Safe language is often the logical result of a safe prompt.

Research suggests the effect extends beyond the occasional cliché. A 2026 study published in Nature Human Behaviour analysed more than 880,000 texts and found that LLM-assisted rewriting reduced variation in writing complexity across datasets and models. Meaning was often preserved while the way people expressed it became more similar.

Research into creative writing has found similar patterns. Across three preregistered studies involving 2,200 college admissions essays, human-written essays collectively introduced more new ideas as the sample grew than GPT-4-generated essays did. Attempts to increase the model's creativity improved diversity without eliminating the gap. Read the study.

None of the research means every AI-generated sentence is identical, nor does it mean using AI automatically makes writing worse. Similar models receiving similar prompts with similar context are simply more likely to gravitate towards familiar ways of expressing the same ideas.

Generic copy is often completely fine

Imagine a software company giving an AI model one sentence to work from:

Our software helps operations teams keep track of work.

A reasonable rewrite might look something like this:

Streamline your operations with a powerful platform that helps teams collaborate more effectively, improve visibility and work smarter.

Very little is technically wrong with the rewrite. Very little also gives those words a reason to belong to one particular company. “Streamline”, “powerful”, “collaborate”, “visibility” and “work smarter” can plausibly describe hundreds of business software products because the original brief never gave the model enough detail to say anything more specific.

A richer brief changes the job:

Operations managers use us when jobs are being tracked across email, spreadsheets and Slack. We pull every open job into one place and show what is waiting, who owns it and what is overdue. Our customers hate bloated project management software, so we deliberately keep the product simple.

More specific source material creates room for more specific copy:

See every job that is stuck before someone has to ask about it. One place for what is open, who owns it and what is already overdue, without another sprawling project-management system.

Better output did not require making the model more “human”. Better input gave the model something distinctive to preserve. Product context, customer frustration and a clear point of view narrowed the range of plausible answers until generic category language became less useful.

A brand voice cannot come from three adjectives

Brand prompts often describe the desired voice as “professional, friendly and conversational”. Those words provide some direction, but thousands of companies could reasonably choose the same three adjectives. A model still needs to make most of the meaningful decisions itself because the prompt says very little about what separates one company's writing from another.

Useful brand guidance contains constraints as well as aspirations. A company might explain technical ideas without technical vocabulary unless the terminology is unavoidable, avoid describing its product as “powerful”, prefer dry understatement over enthusiasm, lead with examples before claims because its audience distrusts marketing language, or deliberately keep the founder's longer conversational sentence structure instead of forcing every thought into punchy fragments.

Specific choices give a model boundaries. Generic adjectives describe an overall mood, while constraints explain how the brand behaves when someone actually has to write a sentence. Better guidance also defines what does not belong, which can be just as valuable as describing what does.

“Make it sound human” solves the wrong problem

Generic AI copy often leads to another generic instruction: “make this sound more human”. A model might respond by shortening sentences, introducing contractions, loosening the rhythm and replacing formal language with something more conversational. Humaniser tools take a similar approach, often concentrating on patterns associated with generated text.

A more natural result can still sound nothing like the company publishing it. Removing obvious AI mannerisms only answers whether the copy resembles a broad idea of human writing, while brand voice asks a much narrower question about whether the copy resembles a particular organisation.

Our earlier article, Why “Humanising” AI Copy Isn’t Enough, looks at the same problem from the editing side. A law firm, skate brand, cybersecurity company and neighbourhood bakery should not all become “friendly, punchy and natural” in exactly the same way. Universal humanisation simply replaces one form of convergence with another.

A more useful review asks whether the copy sounds like the brand behind it. Style, vocabulary, assumptions, confidence, humour, examples and sentence rhythm all matter more once the goal changes from appearing human to remaining recognisable.

Distinctiveness comes from information, not decoration

Surface-level personality is easy to add. A joke can make a paragraph feel less formal, shorter sentences can make a page feel faster and a few banned phrases can remove the most obvious AI clichés. Decorative changes can help, but they rarely create a distinctive voice on their own.

Consider two analytics companies selling similar services:

We help growing businesses make smarter decisions with data-driven insights tailored to their needs.

Compare that with:

Most businesses already have enough dashboards. We help you work out which number actually deserves your attention.

Sentence structure explains only part of the difference. The second company has expressed a belief about the category, rejected the assumption that customers need more reporting and revealed something about how it approaches the problem. A competitor could copy the sentence, but the thinking behind it gives the brand far more material to build on.

Strong brand voice guidance captures those kinds of beliefs. Good documentation tells a writer which language fits while also preserving recurring arguments, attitudes and ways of explaining the product. AI becomes much more useful once the source material contains ideas that cannot be guessed from the category alone.

The best AI brief contains things the internet cannot guess

AI already knows how SaaS companies usually describe SaaS. Generic instructions about sounding confident, concise or benefit-led add relatively little because the model has encountered similar instructions countless times.

Company-specific information is far more useful. Customer complaints before purchase, language used on sales calls, assumptions the team disagrees with, phrases customers naturally use, claims the company refuses to make and examples that repeatedly help people understand the product all reduce the model's need to invent a generic marketing position.

Compare a conventional prompt:

Write a confident and conversational homepage for an analytics company. Keep it concise and benefit-led.

A stronger brief could say:

Our customers are marketing managers who already have dashboards. Their problem is not access to data. They have too much of it and still need to explain what changed at the Monday meeting. We do not sell “real-time insights” because most of their decisions do not need to be made in real time. We focus on finding the few movements worth investigating and explaining why they happened.

The second brief gives the model an audience, a recurring situation, a category disagreement, rejected terminology and a clearer definition of value. Fewer gaps remain for generic analytics language to fill.

Messy source material can therefore be more valuable than a beautifully engineered prompt. Notes from customer calls, founder explanations, support conversations and existing copy can contain more brand signal than another page of instructions telling AI to be authentic.

AI works better when a voice already exists

AI can be genuinely useful for marketing teams. Drafting alternatives, restructuring rough ideas, questioning weak claims and turning notes into publishable prose can all become faster without handing the entire writing process over to a model.

Problems begin when AI has to invent the brand at the same time as it writes the copy. A prompt containing little more than a product category and feature list forces the model to borrow familiar category patterns because almost nothing else is available. Polished output can emerge quickly, but much of the polish belongs to the category rather than the company.

A better workflow starts with the material worth preserving. Beliefs, examples, vocabulary, proof, audience language, preferred levels of certainty and explicit boundaries all give the model something to work within. Revision then becomes an exercise in sharpening an existing voice instead of generating a plausible one from scratch.

AI detection asks the less useful question

Debates about AI-written copy often focus on whether a detector could identify how the first draft was produced. Authorship tells us surprisingly little about whether the final page does its job. A human can write vague and interchangeable marketing copy, while an AI model can produce sharp and brand-specific writing when the source material and constraints are strong enough.

Copy quality becomes easier to judge by looking at the result itself. Brand consistency, clarity, specificity and distinctiveness can all be inspected without knowing who typed the first draft. A sentence that could appear unchanged on five competitor websites probably needs work regardless of whether it came from ChatGPT, an agency or someone inside the company.

Revisi is built around that distinction. Rather than trying to prove whether AI wrote a page, Revisi reviews the copy that actually ended up there and looks for signals such as brand fit, human sound, specificity and distinctiveness. The useful outcome is not an accusation about authorship, but a clearer view of where a website has started to flatten into category language and which parts deserve attention. Run a page through Revisi to see where the copy needs work.

AI will keep getting better at producing polished writing, and trying to identify generated text from surface-level quirks will become less useful as the models improve. Brand voice creates a more durable standard because the benchmark belongs to the company rather than the model.

The goal is not to make AI writing harder to detect. The goal is to make your writing harder to confuse with everyone else's.