Brand Voice Playbook

A brand voice is the consistent character of a company’s writing. It is only useful when it can be checked: if you can’t point to a sentence and say it broke the voice, you have a mood, not a voice. So write it as rules a sentence can fail, put them where your writers and your AI tools read them, and check every draft against them.

Most brand voice work produces a document people admire once and never open again. That used to be survivable, because the same two or three people wrote everything and carried the voice in their heads. It isn’t any more. Your copy now comes from four people and three models, and the models don’t carry your voice anywhere. They carry the average of your category.

Three things changed at once. More people write: founders, marketers, sales, a freelancer. AI drafts a growing share of first versions. And the output ships faster than anyone reviews it.

The second one is what breaks voice. A language model writes likely text for what it’s asked. With no rules, that’s the way most companies in your category already write. Every draft pulls toward the middle, and your voice is the part that isn’t in the middle. There is experimental evidence of the pull: in one study, short stories written with ideas from a language model were rated more creative, but they were more similar to each other than stories written by people alone^[1]. I’ve written about why AI-written content doesn’t sound like you in more detail.

So the question changed. It used to be “how do we sound?” Now it’s “does everything we publish still sound like us?” That’s an operations question, and it needs an operations answer: rules, a place they live, an owner and a check. This playbook covers how a company writes. The promise, positioning and visual identity that the voice expresses are the subject of the Brand Foundations Playbook; settle those first.

Pull twenty pieces you’ve published across channels: site pages, emails, posts, a proposal. Read them in one sitting and mark two kinds of sentence. The ones that could only have come from you, and the ones that could have come from anyone.

The first set is your voice. The second set is the first draft of your list of things to avoid. Derive everything that follows from this evidence, not from how you’d like to sound. A voice written from aspiration produces rules nobody can apply, because nobody has ever written that way.

More than four and the voice has no centre. For each attribute, write what it means in your writing and, the part that matters, what it isn’t. “Candid, not self-deprecating.” “Plain, not simplistic.” “Precise, not pedantic.” The “isn’t” is what stops an attribute being read as licence for its worst version. Mailchimp’s public style guide is built the same way: “weird but not inappropriate, smart but not snobbish”^[2].

This is the step most voice work skips, and it’s the one that makes everything else enforceable. An adjective describes a voice; it can’t be broken. A rule can.

Give every rule a sentence that fails it and a rewrite that passes; those pairs are the do and don’t examples your writers and your AI tools learn from. Where a public standard exists, borrow it: the 25-word limit is the one GOV.UK gives its own writers, who are told to split sentences of more than 25 words^[3].

The test is simple: could a new writer, given only the rule, tell whether a sentence passes? If not, keep rewriting the rule.

Words to avoid is the list a check enforces mechanically, so keep it literal: single words and short fixed phrases. Three kinds belong on it. Category filler that turns up in every competitor’s copy. Claims you can’t defend. And words that are wrong for you specifically, which are the entries a competitor couldn’t copy.

The positive list matters just as much, and it’s the one people leave out. Give a model a banned list and nothing else and it routes around the ban into a different flavour of generic. The phrases you do use are what turn avoidance into a voice. Think of how you describe what you do, how you open, and the line you repeat on sales calls. Model makers give the same advice for prompts in general: Anthropic’s guide to prompting its models says to tell the model what to do instead of what not to do^[4].

One number: the highest reading grade any block of copy may score. The usual measure is the Flesch–Kincaid grade, which the Brand Voice Checker computes from sentence length and syllables per word. I set most B2B copy at grade 8 to 10; that is my practice, not a published standard. Set it higher if you genuinely write for a technical audience, and write down why. Expertise is a weaker reason than it sounds: GOV.UK’s guidance notes that people with higher levels of literacy prefer plain English too^[3].

This is the field most often missing. The work of capturing a voice rarely asks for it, and the work of enforcing one depends on it. Leave it unset and an automated check either never fires or fires on every dense paragraph, and your editors learn to ignore it.

Voice holds on a landing page. It breaks in four places, and they’re exactly where readers pay closest attention:

• Apologies: an outage, an error, a missed commitment.
• Pricing: quoting, raising, defending.
• Bad news: a delay, a refusal, a result that missed.
• Saying no: to a prospect you can’t help.

The voice stays; the tone moves. Nielsen Norman Group describes tone on four dimensions: funny or serious, formal or casual, respectful or irreverent, enthusiastic or matter-of-fact^[5]. Where you sit on them changes how much readers trust you. In its test of the same pages in different tones, a playful version of an insurer’s page was rated friendlier but less trustworthy than a serious one, while a casual version of a bank’s page beat its serious twin on friendliness, trust and willingness to recommend^[6].

For each hard moment, set where the tone sits, and write how you sound and what you never do. Nobody writes a good apology at speed, so decide it in advance.

A voice guide in a shared drive governs nothing an AI tool writes. Put the rules, meaning attributes, do and don’t, phrases, words to avoid and the reading-level ceiling, in the context file your AI tools load: the project instructions, custom instructions or system prompt they read before every draft, and the file an agent reads before it starts. Then every draft starts from them, and a check can quote back every violation with the paragraph it’s in. Send the same block to every freelancer and agency, so people and tools work from one text.

Pair the rules with an AI tool policy. It should say which tools are approved, what never goes into a prompt, and what gets reviewed before anything ships.

Run every draft through a check before it reaches a reviewer, and keep two severities. Hold the line between them. Blocking is a rule the brand wrote for itself: a banned word, copy over the ceiling. Those fail. Advisory is everything else: long sentences, hedges, a paragraph that reads flat. Those get noted, never failed. The moment a check starts failing content on taste, writers stop trusting it, and an ignored check governs nothing.

Know what the check can’t do. Literal matching catches “leverage” and misses “harness the power of”. The check scales the mechanical part of review, the same way every time, so a person can spend their attention on the rest (Figure 7). The Brand Voice Checker quotes each finding in the block it sits in, and marks a banned word inside quotation marks as possibly discussed rather than used.

Name one person who owns the voice. Review a sample of published work every month: which rules are broken most, and why. When a writer and the guide disagree and the writer is right, change the guide. A voice guide that never changes is being ignored, not obeyed.

Judge it on your own numbers, not a benchmark: there is no credible industry figure for “on-brand”. Three measures are worth tracking from the first month:

• Blocking violations per published piece. This should fall as writers and prompts absorb the rules.
• The share of drafts that pass the check first time.
• Which rules are broken most often. A rule broken constantly is either unclear or wrong.

If the first number falls while the same rule keeps topping the third list, fix the rule, not the writers.

What changes by stage?

What mistakes should I avoid?

• Adjectives instead of rules. “Be bold” can’t fail, so it can’t be enforced (Step 3).
• A banned list with no preferred list. The model routes around the ban (Step 4).
• Taste in the blocking tier. The check starts failing good copy and gets switched off (Step 8).
• Rules written from aspiration. Nobody can match a voice nobody has ever written in (Step 1).
• Reading the score instead of the findings. I ran a literal check against my own voice guide and it failed on every banned word the guide discusses. Read what each finding quotes before you act on it (Step 8).
• One tone for every moment. A cheerful price-rise email reads as spin; set the tone for each hard moment in advance (Step 6).