How to Define a Brand Voice AI Can Actually Use

The fastest way to ruin a brand voice project is to start with a tone-of-voice document. Someone writes "professional yet approachable," everyone nods, and six months later the blog still reads like a press release. Defining your brand voice only works if you start from what your brand has actually written, not from words that describe a feeling.

That's the part most guides skip. They'll tell you to pick adjectives, maybe build a mood board, land on a tone. None of that gives a model, or a new hire, anything to do. A voice definition has to be specific enough to steer output, and steering requires rules and examples, not descriptions.

Start from real writing

The first step in defining your brand voice is extraction, not invention. Pull three to five articles you've already published, the ones that sound most like you at your best, and read them for pattern rather than content.

Ignore what each piece is about. Look at how it moves. Where do sentences end? Which words keep showing up? Does it use "we" or "you"? Is there a joke anywhere, and what kind? Every answer becomes a line you can act on: sentences run short, contractions are constant, the brand explains before it sells.

We've watched this play out with our own product. Voice extraction from three to five real articles consistently beats extraction from a polished tone document, because brands describe their voice aspirationally and their published writing shows the real one. A founder will tell you the brand is "warm and confident." Then you read their actual emails and they're clipped, a little dry, allergic to exclamation points. That's the real voice. The tone document was wishful thinking.

If you genuinely have nothing worth mining yet, the workaround is to collect writing from outside your brand that sounds like what you're aiming for, and treat it the same way. Just be honest with yourself that you're choosing a voice at that point, not documenting one you already have. Most established brands don't need this shortcut. They need to stop trusting the mood board over the archive.

This is also where the bigger argument in brand voice in the AI era becomes concrete. Voice as a competitive advantage only works if it's specific enough to reproduce, and specificity starts with the writing you've already shipped.

Make the rules concrete

A usable voice guide replaces adjectives with rules a model can follow without interpreting anything. "Professional yet friendly" tells a writer, human or otherwise, nothing to do, because it's true of almost every brand and instructs none of them.

Concrete rules look different. Sentences stay under twenty words on average. The brand says "customers," never "users." No exclamation points, full stop. "Leverage" and "synergy" are banned outright. Each of these is testable: you can look at a paragraph and say definitively whether it followed the rule or broke it. That's the whole distinction between a voice guide that works and one that gets ignored. Testable beats aspirational, every time.

Do and don't examples do more work than any amount of description. Show a generic sentence next to the version your brand would actually write. "We are committed to leveraging cutting-edge solutions to optimize outcomes for our valued customers" next to "We build tools that get the job done, and we'll tell you plainly when one isn't the right fit." The gap between those two sentences is exactly what a model needs to see to stop defaulting to the generic one.

Banned phrases matter more than most teams expect. Not because a stray "in today's fast-paced world" ruins a piece, but because these phrases are gravitational. Left unchecked, a model drifts toward the average of everything it's read, and the average sounds like nobody. A ban list is a way of naming the drift and cutting it off before it starts.

Use examples that constrain output

Description tells a model what your voice is like. Examples show it what your voice actually does, and the difference matters more than most teams assume going in.

Pick two or three short passages from your own writing, the ones that nail the voice without any editing. Attach them to the guide as anchor examples. A model asked to write "in a warm, expert tone" will produce something bland and forgettable, because warmth and expertise are broad categories with a thousand valid interpretations. A model shown three actual paragraphs of yours has something to imitate, and imitation is a much narrower, much more reliable target than description.

This is also where our own experience got uncomfortable. Readers flagged three tells in our early blog posts before we tightened anything: a colon followed by exactly three parallel items, sentences built on "it's not X, it's Y," and verbless fragments dropped in for punch. They were right on all three counts. We'd been generating from description and general tone guidance, not from real anchor passages, and the gap showed. Adding actual examples of paragraphs we'd stand behind helped more than any adjective list had, though we still catch these patterns in review; anchor passages narrow the gap, they don't close it. The model didn't get smarter. It had something specific to copy instead of a vibe to guess at.

The practical version of this is a short section in your guide labeled examples, holding three to five real paragraphs pulled straight from your best published work. No polishing them for the guide. Use them exactly as published, tells and all, because the tells are part of what makes them real.

Keep one source of truth

Splitting your brand voice across multiple documents doesn't average out into a compromise. It creates conflict, and whichever document gets consulted first wins by accident rather than intent.

We learned this the expensive way. At one point we were running three parallel voice systems on the same project: an inferred tone field the system built automatically, a separate tone-of-voice document someone had written by hand, and a style guide with its own rules layered on top. They disagreed with each other constantly, in small ways that added up. Generation would grab whichever source it found first, which meant the same brand could sound three different ways depending on which document happened to load. Consolidating down to a single authoritative source, one style guide with the voice profile extracted straight into it, fixed more voice complaints than any prompt rewrite we tried before or since.

The lesson generalizes past our own setup. Whatever tool or workflow you're using, decide on one file that holds the vocabulary, the banned phrases, the do and don't pairs, and the anchor examples. Everything else defers to it. If a new tone document gets written later, it either replaces the old one entirely or it doesn't get used, because a model asked to reconcile two conflicting sources will pick one inconsistently every time, and you won't be able to predict which. This single-source discipline is most of what brand voice guardrails are built to enforce at scale.

One more thing worth saying plainly: some rules shouldn't live in the prompt at all. We ship an AI writing product with a hard no-em-dash rule, and our own model still breaks it often enough that reminding it stopped being worth the effort. We strip dashes in code now, after every generation, deterministically. If a rule is truly non-negotiable, enforce it mechanically rather than trusting a model to remember it every single time. Prompts are guidance. Code is a guarantee.

A voice guide is strong enough when a capable writer who has never seen your brand could read it and produce an on-brand paragraph without asking you a single follow-up question. If they'd still need to ask what tone to use or which words to avoid, the guide isn't finished. That test also happens to be a fair way to judge whether your content is legible to AI systems that read and cite it, a question covered in more depth in Generative Engine Optimization.