Getting AI to sound like your business rather than like AI

Everyone can spot it now

The tell is not vocabulary, it is the absence of specifics. Copy that could apply to any business in the sector reads as automated because nothing in it required knowing anything about you.

Customers notice this faster than most companies expect, and the damage is not embarrassment. It signals that nobody local was paying attention, which is precisely the opposite of what most of this content is trying to convey.

Prompting harder for a warmer tone does not fix it, because tone was never the problem.

Specificity comes from research, not instruction

A model asked to write about a service in a town it knows nothing particular about will produce something generic, because generic is the only honest option available to it.

Give it material first. What this business actually does, which services it sells, what is true about this area, what the customer said. Then the writing step has something to be specific about.

This reordering, research before drafting rather than during, is most of the difference between output that reads as informed and output that reads as filler.

Voice is data, not a prompt

Tone instructions buried in a prompt drift and cannot be inspected. A better approach is to treat voice as stored, editable data: preferred vocabulary, phrasings to avoid, how formal to be, which local references are safe.

Stored that way it stays consistent across every generation, and someone can correct it when it is wrong. A prompt nobody can see is a prompt nobody can fix.

Regional variation matters more than people expect. Copy that reads naturally in one part of the country reads as slightly off in another, and slightly off is worse than plain.

Human edits are the best feedback you have

Track what people change before publishing. Those edits are a precise map of where the voice profile is still wrong, and they are far more useful than any general assessment of quality.

A phrase that gets removed every time belongs on the avoid list. A correction made repeatedly is a rule waiting to be written down.

Systems that capture this improve. Systems that do not stay exactly as good as they were on day one.