Products blend together when brands skip their own style.

There is a version of AI-assisted marketing that produces competent, publishable, entirely interchangeable content. The grammar is clean. The structure is sensible. The claims are defensible. And a buyer reading three vendors in one afternoon could not tell you which page belonged to which company.

Why the convergence happens

A language model returns the most probable next phrase given everything it has read. In a category with hundreds of vendors publishing similar material, the most probable phrasing is the category average.

So when a team prompts for a page about their platform, the model reaches for the sentences that already exist in that category, and returns them slightly rearranged. Every competitor prompting the same way gets a draft from the same middle.

A model trained on the middle of the market will return you to the middle of the market.

The cost is positioning, not prose

The output is usually fine. The problem is that fine is not distinguishable, and buyers do not choose between vendors on grammar. Search guidance makes a version of the same point: how the content was produced is not the issue — whether it adds anything is.

Positioning works by contrast. It requires a buyer to be able to say what you do that the alternative does not, and to remember it later. Content that lands on the category average removes the contrast, which means the buyer defaults to whatever they already knew, the incumbent, or price. It is the same mechanism as borrowing someone else's positioning: the words work, but they work for the category rather than for you.

What the model cannot supply

What makes a page yours is the material a model has no access to.

• The specific number: what your deployment actually took, what it actually cost, what actually broke.
• The argument you will defend: the thing you believe about this market that a competitor would dispute.
• The customer's own words: what they said in the call, not what the category says they think.
• The thing you decided not to do, and why. Constraints are more distinguishing than capabilities.

None of that is retrievable by prompting harder. It comes from your work, your customers, and the choices you have already made — which is the same reason AI compressed the production and not the judgment.

A test worth running

Take a page you published recently. Remove the logo, the product name, and any proper nouns. Hand it to someone who knows the category and ask which company wrote it.

If they cannot tell, or if they name a competitor, the page is doing brand-neutral work at your expense. The fix is rarely a rewrite of the whole thing. It is usually two or three insertions of something only you could have said. The same tell shows up in cold email, where a template announces itself before the pitch does.

Where sameness is the point

Not all writing is positioning, and the exception is worth naming because it is where AI earns its place outright. Internal crisis communications is the clearest case. During an incident you need the same confirmed facts in 6 formats — an all-hands note, a manager script, a customer holding statement, a support macro, a regulator-ready summary, a board update — quickly, consistently, and deliberately unremarkable.

Distinctiveness is not the goal there. Accuracy and consistency are. Divergence between versions is how an incident becomes 2 incidents.

• Draft from a single source of confirmed facts, never from the model's memory of similar incidents.
• Generate every variant from that one source, so the facts cannot drift between audiences.
• Keep the caveats and the unknowns identical in each version, including what you are not yet saying.
• Human approval before anything leaves the room. Speed is the benefit; authority is not delegable.

The judgment that stays with you is what is confirmed, what is not, and who hears what and when — which is also where the positioning consequences of breach comms get decided, long before anyone drafts the customer note.

Where AI belongs in this

This is not an argument for writing everything by hand. Models are genuinely good at the parts that should be conventional: structure, transitions, tightening, alternate headlines, and the fifth version of a paragraph you have stopped being able to read.

Treat the draft as scaffolding, not output. Deciding what deserves saying, which claim leads, and which specifics carry the argument was never delegable, and that is what buyers use to tell you apart. It is also what people-first content means in practice, and where a decision about how the company sounds actually gets spent.

Part of our work on creative & content.