Make AI speak your brand

The Brand Context to Give AI Is What Rivals Won't Say

Six steps to get ChatGPT, Claude or a similar tool writing like your company, starting with the draft it writes before it knows anything about you.

By Ines Calloway · November 18, 2026 · 7 min read

A wooden-handled rubber stamp next to a stack of forms, each marked with the same stamped impression.
Given nothing specific about a company, an AI tool writes copy any rival in its market could use.

Most advice on how to give AI your brand context starts with uploading your brand guide. That's a fair start for tone. But most brand guides hold a purpose statement, a values list and a handful of tone words, and your competitors' guides read much the same.

So the model hands back your market's usual copy with your name on top. The model already knows what companies like yours sound like on average. What the model can't know is the part the average leaves out: the positions you hold that rivals avoid, and the proof nobody else in your market can show.

1. Start with the copy the tool writes when it knows nothing about you

Before you give the tool anything, ask it for a homepage headline and an About paragraph, naming only your market. Say a firm does commercial cleaning for hospitals. Its prompt would name hospital cleaning and nothing else.

Save what comes back. That draft is where the model drifts when your context is thin, and every later draft gets measured against it.

Research on AI-assisted writing points the same way. In a study of AI-assisted essay writing, Vishakh Padmakumar and He He found that essays written with InstructGPT, a model tuned on human feedback, were measurably less varied.

The drop came from the model's own text. What people wrote themselves stayed as varied as before, and the base GPT-3 model didn't have the effect.

The study used argumentative essays and 2023 models, so read it as a sign of which way the tools lean. Why that lean matters for a brand has its own post on brand in the age of AI.

2. Your context starts with what competitors won't say

Before you write a single tone rule, make a list.

At the top go the positions you hold that competitors steer clear of. Next comes the proof only you can show, such as a named customer, a number or the way your team does the work.

Then the part that's easiest to skip: the promises everyone in your market makes that you're willing to drop.

For the made-up hospital cleaner, a position its rivals avoid might be refusing any contract without a supervisor on site overnight. Its made-up rivals all promise "infection control excellence" and call themselves "a trusted partner," so both phrases go on the drop list.

This list is the core of what you give the model. Mission and values can come after it, since they're the part of a brand guide most likely to match a rival's.

3. A few pages of your own writing teach more than tone adjectives

Adjectives alone give a model little to go on when you train AI on brand voice. Paste in three to five pieces of your best writing as well, each labelled with what it is, say "proposal opening" or "client email."

A cloth tape measure lying across an unfinished jacket marked with chalk lines and held with pins.
A tailor cuts from the person's own measurements; your own writing does that job for an AI model.

Anthropic's prompting guide for Claude calls examples "one of the most reliable ways to steer Claude's output format, tone, and structure," and suggests three to five of them.

A few paragraphs that sound like you give a model far more to work with than "approachable, supportive, inspirational." Every competitor can claim those three words.

Barona, a Finnish working life company that employs over 30,000 people a year across Finland and the Nordics, gives one team's view. State of Brand's interview with CMO Joni Helminen describes his setup in the writer's words, and only the quoted lines are Helminen's own.

He says he doesn't have pipeline impact, ROI or quality scores, so treat Barona as one team's working system.

According to the interview, that system learns from more than 100 articles the team has published, which is this step at scale. Helminen sees pasting a style guide into a prompt as a reliable path to generic output. In his words: "AI tends to default to safe, balanced, and comprehensive, unless you sort of prompt that into it."

4. Every phrase you ban needs the line you'd write in its place

A list of banned words tells the model where to stop and leaves it guessing what to write. Anthropic's guide has a heading for this: "Tell Claude what to do instead of what not to do."

So set each phrase from your drop list beside the line you'd use. For the made-up hospital cleaner, the pairs might read:

  • "Infection control excellence" becomes "a supervisor on every night shift"
  • "A trusted partner" becomes "the same crew on your ward each week"

5. Store the context where the tool reads it on every request

Save the context as standing instructions so every new chat starts with it: a Project in Claude, or in ChatGPT, a custom GPT or the custom instructions setting.

An open wooden index-card drawer with brass label holders and tabbed dividers separating the cards.
Filed once by subject, the reference cards are ready for each request without being pasted in.

If you're working out how to train ChatGPT on your brand voice, put your step-two list, your writing samples and your phrase pairs into one of those two places, so each chat starts from them.

Inside those instructions, put long reference documents at the top and your request at the end. Anthropic's guide says queries at the end can improve response quality by up to 30% in its tests, especially with complex, multi-document inputs. That's the vendor testing its own model, and "up to" is the best case.

Helminen's setup shows this step in practice: his context sits in separate files the AI draws on. There's one for the brand, one for tone of voice, one per service line, and ideal customer profiles broken out by business unit. Each sub-brand gets its own file as well.

He also admits the first version held too much. "The honest answer is I probably overdid it."

6. If a colleague can't pick your copy out blind, go back to step two

Generate a piece with your context loaded. Put it beside copy from two competitors, cover the company names and ask a colleague which one is yours.

If they can't tell, the context is missing what makes you different, and more tone rules won't add it. The list from step two is where to look.

Helminen runs the first half of this check. "The test I would do is a blind read of the draft. If you can't tell that it sounds like the voice you're trying to make it sound like, then the system has sort of failed."

His read compares a draft with the voice he wants. Setting competitors' copy beside it is this guide's addition, and it's the half that shows whether the draft could belong to anyone in your market. A fuller version of this check, the interchangeability test, has its own post.

Readers liking AI copy doesn't show they'd know it was yours

Isn't default AI copy already good enough? There's a real study behind that view.

MIT Sloan reported on research by Yunhao Zhang and Renee Richardson Gosline in which people rated marketing copy for retail products and persuasive content for campaigns. Professional content creators and GPT-4 wrote the pieces, alone and together. Told nothing about who wrote what, people preferred the AI's work.

But the study measured how good each piece seemed. It didn't test whether readers could tell one company's copy from a rival's, which is the choice a buyer makes when two of you are on the shortlist.

On that question, the essay study from step one points the other way: writing with a feedback-tuned model made people's work more alike. Copy that wins a quality rating can still be mistaken for a rival's.

The tone rules are the easy part

Step two is where most of the thinking goes, because it asks you to say in writing where you part ways with your competitors. If you'd like to talk through what that list should hold for your brand, get in touch.

Keep the baseline draft from step one somewhere handy. When a new draft starts to read like that baseline, you'll know it's time to reopen your step-two list.

Frequently asked questions

How do I give AI my brand context?
Start by asking the tool for a homepage headline and an About paragraph with only your market named, and keep that draft as a baseline. Then give it a list of the positions you hold that competitors avoid, the proof only you can show and the promises you'll drop, followed by three to five labelled samples of your own writing.
Is uploading my brand guide enough to train AI on my brand voice?
A brand guide is a fair start for tone, but most hold a purpose statement, a values list and a few tone words that competitors' guides share. A model given only that tends to return your market's usual copy with your name on it. Pasting in three to five pieces of your best writing gives it far more to work with.
Where should I save brand instructions so ChatGPT or Claude uses them every time?
Save them as standing instructions: a Project in Claude, or a custom GPT or the custom instructions setting in ChatGPT. Put long reference documents first and your request at the end, which Anthropic's guide says can improve response quality by up to 30% in its own tests.
How can I tell whether AI copy sounds like my company?
Generate a piece with your context loaded, set it beside copy from two competitors with the names covered, and ask a colleague to pick yours. If they can't, the context is missing what makes you different, and the list of positions rivals avoid is where to look.

Sources