Most brand voice advice assumes there's a voice guide somewhere to follow. For most businesses there isn't. If you're a founder writing your own website, a small team that inherited a site someone else wrote, or a freelancer picking up a new client, what you've got is a homepage, a pricing page, a few blog posts and an About page written in a hurry two years ago. Whatever voice exists lives in those pages, not in a document.
That doesn't mean you have no brand voice. It means your voice is inferred rather than documented. An inferred voice describes how you currently sound, read back from what you've already published. A documented voice is a decision about how you intend to sound, written down so anyone can check a new page against it, which is the idea behind the define, document and defend approach. A simple test tells you which kind you have: when two people disagree about whether a page fits, can you point to something other than taste? This post covers how to read your inferred voice, why it's risky to enforce as it stands, and how to turn it into a short documented version in an afternoon.
The trap: your site is averaging its own mistakes
When you infer a voice from your whole website, you're reading the average of everything you've published, and that includes the drift. Say the founder wrote the homepage carefully, a contractor wrote the feature pages, and the blog posts were AI drafts published with barely an edit. Your inferred voice is then a blend of all three, weighted towards whichever patterns appear most, and those often come from the pages that got the least attention. It's also circular, because the pages that drifted helped set the standard they're being judged against. That's why Revisi labels a voice read from your site as inferred until someone has reviewed and confirmed it.
Read it from your best pages, not your whole site
The fix is to choose what you infer from. Pick three to five pages you'd be happy for a new customer to read first, such as pages that perform well, pages the person who knows the business best has signed off on, or pages you've recently rewritten because the old version bothered you. Don't include or exclude a page because of how it was written. An AI-drafted page that was edited with care can sound more like the business than a hand-written page that was rushed. Leave out legal pages, old announcements and anything thrown together for a launch.
Then read those pages for patterns you can observe rather than adjectives. "Friendly" and "professional" describe how a page feels to one reader, which makes them almost useless for checking new copy. Observable patterns include typical sentence length, whether you write as "we" or keep the company in the third person, how you support claims, which words keep coming up, which words never appear, and how you sound when you're talking about price or saying no. Write down each pattern with an example sentence. To make this concrete, imagine a small invoicing app for sole traders, and compare two ways of writing up the same reading of its site.
Before: Our voice is friendly, clear and confident, with a touch of personality.
After: We address the reader as "you" and rarely say "we" in headlines. Most sentences on product pages are under 20 words. Claims are backed by a number or a named example ("cuts invoice prep from about an hour to ten minutes"), not by intensifiers. We don't use exclamation marks anywhere except the signup confirmation.
Find the contradictions and make a call
Next, check those patterns against the rest of the site, looking for places where two pages disagree. Each disagreement is a decision your inferred voice can't make for you. A common one is a homepage that speaks plainly to small business owners while the pricing page slips into procurement language.
Before (homepage): You shouldn't need an accountant to send an invoice. Set it up in ten minutes and get paid faster.
Before (pricing page): Our enterprise-grade platform delivers scalable invoicing solutions to optimise your accounts receivable workflow.
They can't both be your voice. Inference alone would split the difference, or follow whichever style turns up on more pages. Documenting means making a call. Here the homepage is closer to how the business wants to sound, so the pricing page gets rewritten to match. Write each decision down as you make it, because these calls often end up being the most useful lines in the finished guide.
After (pricing page): Every plan includes unlimited invoices and automatic reminders. The difference between them is how many people on your team can send them.
Turn observations into rules you can check
Getting from inferred to documented is mostly a matter of rewriting your notes as rules. A rule is something a person or a tool can apply to a single sentence and get a yes or no answer, and five to eight rules is usually enough. Cover sentence length, point of view, how claims are supported, words to avoid, and the awkward moments like price, limitations and errors.
Before: Our tone is confident but never arrogant.
After: State what the product does without hedging ("it reconciles your accounts", not "it can help you reconcile your accounts"), but don't compare yourself to competitors by name or claim to be the best at anything you can't show.
The second version is longer, but it's the one you can check against. If a page says "the leading invoicing platform for small business", you can point to the rule instead of arguing about whether it felt arrogant. That shared reference is what makes a regular brand voice audit possible.
Before you treat the rules as the standard, show them to one or two people who know the customers well and ask them where the rules are wrong. Their pushback is usually about the edges, such as a banned word that customers actually use or a sentence limit that doesn't suit a technical audience. Once you've agreed on a version and dated it, the voice has moved from inferred to confirmed.
Where AI fits
A brand voice doesn't have to start with a person at a keyboard. Many businesses now get most of their copy from AI drafts that someone shapes and approves, and the voice that comes out of that process is as real as a hand-written one. AI is also useful for the inference step itself, because a model can summarise patterns across a set of pages much faster than you can, and it often spots habits you've stopped noticing. The averaging trap still applies, though. Point it at your reference set rather than the whole site, and treat what comes back as an inferred voice to review.
Documented rules matter most when AI is doing the drafting. Ask a model to write "in our brand voice" without rules and it will infer a voice from whatever examples it's given, then drift towards the generic patterns it learned in training, which explains a lot of why so much AI copy sounds the same. Giving it your rules with examples improves the drafts and gives you a standard to check the output against, whoever or whatever wrote the sentence.
When an inferred voice is good enough
Not every business needs to go all the way. If one person writes or approves everything and the site is small and stable, an inferred voice held in one head can work for a long time, because the writer and the standard are the same person. The case for documenting gets stronger once a second writer joins, you bring in contractors or an agency, more of your copy starts as AI drafts, or the site grows past what one person can reread.
If you'd rather start from a draft than a blank page, Revisi reads your site and produces an inferred voice summary, clearly marked as inferred, that you can edit and confirm into a voice profile. Once it's confirmed, every page is checked against the voice you chose rather than the one your site drifted into.