Open five competitor blogs in your category and read the first paragraph of each. There is a decent chance they open the same way: a rhetorical question or a short scene-setting sentence, then a line about how the landscape is changing, then a promise that this article will break it down. Different companies, different writers, one voice.
This is not plagiarism and mostly not laziness. It is what happens when a market full of people asks a small number of models for the same thing and accepts roughly the first answer. The model is doing exactly what it was built to do — produce the most probable next sentence — and the most probable sentence is, by definition, the one everyone else is also getting.
What actually converges
It helps to be precise about where the sameness lives, because it is not mainly in the words.
It is in structure. Three sections, each with three points, each point a bolded phrase followed by two sentences of elaboration. It is in rhythm — a consistent medium sentence length, few very short sentences, almost no very long ones. It is in the moves: open by naming the problem, acknowledge the counterargument, close on a forward-looking note that commits to nothing.
And it is in a small vocabulary of connective tissue that models reach for constantly. Anyone who reads a lot of this output starts spotting it within a sentence or two, and increasingly that includes your customers, who are also using these tools and know what the default sounds like.
The practical consequence is not that readers think “this is AI.” It is quieter and worse: nothing registers. The page is read, understood and forgotten, because there was nothing in it shaped differently enough to catch on anything.
Sameness is a positioning problem first
Before blaming the tool, it is worth checking whether the tool is only exposing something that was already true.
If your product page could have its logo swapped for a competitor’s and still read correctly, that is not a writing failure. The copy is generic because the positioning is generic, and the AI just produced generic copy faster than a human would have. Plenty of brands survived on human-written text that was equally interchangeable; it simply took longer to produce and therefore felt more deliberate.
The useful test is to write down the three things you would say that a direct competitor could not say honestly. Not values — specifics. Who you deliberately do not serve. The thing you are worse at and are fine with. The opinion about your market that would start an argument at a conference. If that list is hard to write, no prompt will save the copy, and the work to do is positioning rather than tooling.
If the list is easy to write and none of it appears in your published pages, then it is a tooling problem, and it is fixable.
What the model does not have
A model can imitate a style. It cannot supply the material that makes writing specific, because that material is inside your company and not on the internet.
It does not know what the customer said on Tuesday, or the three objections your sales team hears every week, or the feature you built and then removed, or the number from your own dashboard, or the mistake you made in your first year that shaped how the product works now. It does not know the way your founder actually talks, and it does not know your industry’s small internal disagreements — the ones practitioners care about and outsiders have never heard of.
This is the whole lever. Generic input, generic output; there is no prompt that manufactures particularity from nothing. But specific input changes the result immediately, because the model is a decent writer and a poor source. Give it your material and it will arrange it well. Give it a topic and it will arrange everyone’s.
Practical ways to sound like yourself
Five things, roughly in order of how much they pay back.
Put a real fact in the first hundred words. A number from your own data, a dated event, a named customer situation, a thing you tried that failed. This is the fastest way to make a page unmistakably yours, and it also happens to be what makes it worth reading.
Build a voice sample rather than a voice description. Collect three or four pieces of your own writing that sound right, and paste them in as the reference every time. Adjectives like “conversational, confident, not corporate” get you the model’s idea of those words. Your own paragraphs get you your own cadence.
Take a position in every piece. The default output hedges, because hedging is the safest average. If a sentence in your draft could be signed by any company in the category, either sharpen it or cut it. One clear opinion per page is usually enough.
Break the structure on purpose. Vary section length. Let a paragraph be one sentence. Cut the summary conclusion, which contributes almost nothing and is the most recognisable tell in the format. Small disruptions to the shape do more than word-level editing.
Edit last, not first. Do not polish sentences in a draft whose structure is the problem. Decide what the piece argues, cut whatever does not serve it, then fix the prose. Most AI drafts improve most by deletion.
None of this requires a different tool. It requires your material and about twenty minutes, which is the part that does not scale and never will.
Where sameness is completely fine
The argument has limits, and pretending otherwise would be the same kind of overclaiming.
Plenty of marketing text is functional and nobody wants it to have a voice. Transactional emails, product specifications, category pages, alt text, ad variations tested against each other, meta descriptions, internal documentation. Here the default output is not a compromise — it is correct, and spending editorial effort on it is waste. Much of what AI SEO tools and AI tools for social media marketing do well lives in this territory.
The distinction worth holding is between text that carries the relationship and text that carries information. Information can be average. The pages someone reads while deciding whether to trust you cannot, and there are usually fewer of those than people think — a homepage, a few product pages, the things you actually want forwarded.
Spend the human hours there and let the tool run everything else. That is a better answer than either refusing the tools or handing them the brand.
The market’s default voice will keep getting more uniform, because the incentive to publish faster is not going away. Which quietly improves the return on sounding like a specific company with specific opinions — the same advantage it always was, just cheaper for everyone else to forgo. AI marketing tools vs marketers goes further into which parts of the job are actually moving, and AI tools for content marketing covers the production side of it.
