Guide
How to write brand voice guidelines that change what gets written
Most of them go in a drawer within the month. The good ones get pasted into a system prompt on Monday.
A brand voice guideline is only worth writing if somebody who has never met you can produce a passable draft from it alone. Most cannot, and the reason is not effort. It is that the document describes an impression when the reader needed a rule.
The usual failure
Written before the fact, in adjectives
Open almost any tone of voice document and the first page says the same thing: confident but approachable, expert but human, clear, warm, never salesy. Every one of those is true of your nearest competitor as well, which means none of them tells a writer what to do on Tuesday.
The deeper problem is the order it was written in. A style guide is written first and says what somebody wishes the writing were. The writing then carries on exactly as before, because nothing anybody does daily depends on the document. Within a month it is a PDF nobody opens.
That order has become expensive. Five people write for the business and produce five voices. ChatGPT, Copilot or Gemini produce a sixth, and the only quality control left is one person reading everything before it goes out. A guideline that cannot be followed without that person in the room has not solved anything.
A rule a writer cannot fail is not a rule. It is a mood.
Start from evidence
Measure what you already publish
The voice already exists. It is in the last two years of your website, your proposals and your emails, and it can be counted before it is described. Twenty-two things are measurable in any body of writing; four of them separate a real voice from a generic one, and the rest are still worth writing down.
- Sentence rhythm. Your median sentence length, your longest, and how much the lengths vary. Real voices vary a lot; generic writing clusters around one length and reads as wallpaper.
- Landing. What share of your sentences are five words or fewer. Across real businesses that runs between 15% and 29%. Copy assembled from the category average manages 0%.
- Pronoun policy. Whether you say I, we, or you, and at what rate. One fractional practice measured at 133 uses of "I" per ten thousand words and 22 of "we". A property care business by the same writer came out at 342 "we" and not a single "I". Same hand, opposite policy, and neither had ever noticed.
- Punctuation habits. Colons against semicolons, whether you use long dashes, how you handle a list.
- Borrowed vocabulary. How much of the text came out of the phrasebook every business in your sector shares. This is the measure that costs most sites their distinctiveness, and the one that AI tools push hardest in the wrong direction.
- Evidence. How often a sentence carries a figure, a date, a name or a result. Many sites contain no number of any kind, and some of those lead with the word evidence in the headline.
Do this by section, not just for the whole corpus. A site can score well overall and change costume entirely when it starts describing services. One business measured had a sentence variance of 10.8 in its origin story and 6.2 on its service pages. The whole-site number hid it completely.
The turn that matters
Turn each measurement into something a writer can fail
A measurement becomes a guideline the moment it is phrased so that compliance is checkable. Compare the two forms.
Does nothing
"Keep sentences short and varied, and write with confidence."
Nothing here can be failed. The writer already believes they are doing it.
Changes the draft
"Median sentence eleven words. Never more than fifty. One sentence in six under five words, placed where the argument turns."
Three numbers, each of which a new hire or a model can be held to.
The same move works for every measure. "Be concrete" becomes "every claim of benefit carries a figure or a named example in the same sentence". "Sound human" becomes "contractions in emails and never in proposals", if that is what the evidence shows. "Avoid jargon" becomes a list of the twelve words you actually reach for that everybody else in the sector also reaches for, and an instruction never to use them.
Prohibitions travel further than encouragement. A writer standing in for you cannot easily imitate a strength, but they can reliably avoid a listed mistake, so the section headed "what this voice will never do" is the one that gets used most.
What earns a page
The sections that do the work
- What you believe. The three or four positions your writing takes that a competitor would argue with. Most guides skip this, which is why they are so easy to copy. Nobody can copy what you actually think.
- The measurements, as targets rather than description.
- Your idiom. The phrases that are yours, quoted from your own published work, and the ones to ration before they become catchphrases.
- What you will never do. Words, claims and constructions that are wrong for you.
- Worked pairs. A generic sentence beside your version of it, with one line on exactly what changed. This is how somebody learns a voice in an afternoon.
- Awkward cases with rulings. How you price in writing. How you say no. How you deliver bad news. How you write about AI without sounding like everybody who discovered it last year. This is the section people open when they are stuck.
- A prompt pack. The whole thing compressed into the form ChatGPT, Claude, Copilot and Gemini take instruction in, because the tools are now the most prolific writer on your staff.
The last step
Install it, or it is a PDF
The step most voice work skips is the one that decides whether it worked. Rewrite three live pieces against the guideline before anybody is asked to follow it, so the rules have been tested on real work. Paste the prompt pack into whichever AI tool your team actually runs, ChatGPT, Claude, Copilot or Gemini, and check a draft against it. Then hand the document to somebody who has never met you and ask for a page.
If they produce something you would publish, the guideline works. If they produce something you have to rewrite, the guideline has described a mood, and the fix is more measurement and fewer adjectives.
One honest limit. No guideline holds forever. Models get updated, people paste over instructions, and the average of the internet pulls every draft back towards the middle. Re-run the numbers every few months; the free test takes a minute and tells you whether anything has moved.
Written before the fact, a guideline is a wish. Written after it, from the evidence, it is a specification.
Find out what your evidence says first
Paste a page you have published and get the four scored measures, the phrases you share with your sector and the sentences that are genuinely yours, in about a minute. A Wordprint is the full version of this guide, written for one business, from its own words.
Written by Ian, with Claude assisting, and scored by the smarticulate method at 92.2 out of 100 before it went up. Ian spent twenty-odd years turning thinking, his own and other people's, into something a board would act on, and built smarticulate to put numbers on it. His practice is ConsultingPartner.io: enterprise experience, right-sized. Who is he? →
More guides: ChatGPT custom instructions for your brand voice, tone of voice examples, how to measure brand voice.