How do we build a brand style guide our team and our AI will both follow?

A brand style guide your team and your AI will both follow has three layers, and the order matters more than the contents. Layer one is what you believe and who you're for. Layer two is how you talk, format by format. Layer three is how you look. Most companies ship layer three in full, layer two as a handful of adjectives, and layer one not at all. That's why nobody follows the guide you already paid for, and it's why an AI tool can't use it even when somebody pastes the whole thing into a prompt.

What actually belongs in a brand style guide?

Start with the layer that requires judgment. Anybody can copy your hex codes by looking at your website. Nobody can guess what you believe by looking at anything. Layer one is the part only you can write: your best-fit customer in plain language, the problem you fight, the belief you hold that your three closest competitors would argue with, the words you own, and the words you refuse to use. That layer is your narrative identity, and it sits underneath everything you will ever publish. The difference between brand identity and narrative identity is the whole ballgame here, and most style guides only document the first one.

Layer two turns that belief into writing rules, one format at a time. A cold email, a case study, a homepage section and a conference abstract need different lengths, different openings, different closes. A voice spec built format by format gives a writer something to follow in a way a tone paragraph never does. Layer three is the visual system, and it absolutely belongs in the guide. It just doesn't belong first.

Why does nobody follow the style guide you already have?

Because it answers questions your team wasn't stuck on. A new marketer doesn't freeze over which blue to use. They freeze over whether to name the competitor, how to describe the problem in the first line, and whether the customer story leads with the number or the person. Open your current guide and go looking for those answers. They're usually nowhere in it.

The second reason is that adjectives can't be checked. "Confident, approachable, human" sounds like a standard right up until three writers turn in three different pages and every one of them is defensible under those words. A rule survives an argument. "Never open with the company, open with the reader's problem in their words" either happened or it didn't, and anyone can tell which. Write rules you can fail.

The third reason is ownership. A guide written by an agency and handed over at the end of a project has no one accountable for the next argument. Somebody has to be allowed to settle what the company believes, and it has to be somebody above marketing, or the guide gets reopened by every new hire with a strong opinion and a fresh mandate.

Why is a PDF style guide worse now that your team writes with AI?

The old guide was enforced by a person. Somebody senior read the draft and quietly fixed the parts that drifted. That check has mostly disappeared, because the first draft now comes out of a model and it arrives polished. Polish used to mean somebody cared enough to sweat the words. Polish is free now, and it hides drift instead of revealing it. What you lose when your content sounds like everyone else compounds one asset at a time, and nobody catches it in review because nothing looks wrong.

There's a mechanical problem underneath the cultural one. A design PDF can't be loaded into anything. If your rules live in a file none of your team's tools can read, they're decoration. The fix is to keep the guide as plain text, owned by one person, versioned like code, and readable by the machines your team actually writes with. Companies who push this far enough end up serving those rules over an MCP server so every tool reads the same version on every request instead of whatever somebody pasted in last quarter.

How do you test whether your style guide is followable?

Three checks, and you can run all of them this afternoon.

The cold-start test. Hand the guide to somebody who joined three weeks ago and ask for a one-page customer story. No questions allowed. Whatever they had to invent is a hole in your guide, and the holes cluster in the same place every time: what to claim, what to name, what to refuse.

The paste test. Drop the entire guide into ChatGPT or Claude and ask for a cold email to your best-fit buyer. Read what comes back, then cover the logo and ask who wrote it. If the answer could be any of your competitors, the guide is missing layer one and no amount of tone coaching will patch it.

The disagreement test. Find one sentence in the guide that a competitor couldn't lift word for word and put in theirs. If every line is true for everybody in your category, you documented the category instead of the company. That's the most common outcome, and it's the quietest one, because a category-true guide reads perfectly reasonable to everyone who approves it.

What do we see when we audit homepages against this?

We've now run 47 Brand Signal Score audits on B2B homepages. The rubric scores narrative clarity, trust, AI-readiness and conversion, and it never grades design at all. The median homepage comes in at 50 out of 100, and 23 of the 47 score under half.

The useful part is where the top gap lands. On 12 of the 47, the single biggest gap is that no enemy and no status quo is ever named. On nine more, it's differentiation by belief rather than by feature list. On six more, it's owned, distinctive language. Add those together and a conviction gap is the number one finding on 27 of 47 homepages, 57 percent of everything we've scored. The thing companies do best, by a distance, is naming who the product is for in the hero, which shows up as a strength on 21 of 47. People document the audience and the look. The belief goes undocumented, and then it goes unsaid. You can run the same audit on your own homepage in a couple of minutes.

What does a followable guide look like in practice?

The companies that get this right stop treating the style guide as a deliverable and start treating it as a source file with an owner. It opens with the narrative identity. It moves into writing rules per format with real before-and-after examples, because an example settles arguments that a rule alone won't. It carries a kill list of words and constructions the company refuses, which is the single most useful page for anyone writing with a model. It finishes with the visual system. Then it gets loaded into the tools rather than filed in a brand folder.

That same document is what makes an AI Brand Twin possible, and it's the reason messaging holds together when the whole team writes with AI. One file, one owner, one version, read by people and machines alike.

If you'd rather watch this get built than read about it, the AI Workforce Clinic walks through this exact setup with a live company each session, and the first one is free. Come sit in on one.

What this means for you

You probably don't need a new style guide. You need the two layers that were never written down, added underneath the one you already have. Take an hour, open a blank file, and answer four questions in plain language. Who is the best-fit customer? What problem do we fight? What do we believe that our competitors would argue with? Which words do we refuse to use? That file is worth more to a new hire, and to a language model, than the forty-page PDF sitting in your brand folder.

Do that, and the machines writing for your company start writing from your conviction instead of from the average of your category. Your team stops guessing, and your buyers stop reading a page that could have come from anyone.