Ask Claude “who is the best musician right now?” and it will tell me Bruno Mars. Not because Bruno is objectively the best (though I’ll fight you on that), but because the model knows I’ve spent weeks asking it what to bring to his concert in Amsterdam on the 4th of July, whether the Silk Sonic tracks make the setlist, and how early to queue. It has watched me care. So when I ask a vague question, it hands me a personal answer.
My wife asks the exact same question and gets Remy van Kesteren. The harpist. Because that’s who she’s been retrieving.
I like this example because it gets funnier when you narrow it. Ask “best band ever” and I get *NSYNC, she gets Coldplay. Same model, same day, two answers that would embarrass us at different dinner parties. That gap isn’t the model being inconsistent. It’s the model being relevant. And that gap is the whole point of this article.
Personal context is just SEO wearing a different jacket
Traditional SEO always rewarded history. An account with behaviour behind it, a browsing pattern, a click trail: that’s what let Google hand you results that fit you instead of the average of everyone. Personal, but relevant.
LLMs do the same thing, they just do it faster and they do it from what you tell them directly. The history is the memory. The account is the context you feed it. Feed a model nothing and it answers the median internet user, which is nobody. Feed it who you are and it starts answering you.
This is Search Everywhere Optimisation in miniature. The surface changed from a results page to a chat window, but the mechanism didn’t: the more a system knows about the searcher, the more useful it gets. We spent fifteen years optimising for that on Google. Almost nobody is doing it on purpose inside an LLM.
Context already moves the needle, even when you don’t touch it
If you think this is speculation, Anthropic did the homework. In July 2026 they published research on how Claude’s values shift across models and languages, analysing more than 300,000 real conversations. The finding: the same request gets a different mix of values depending on context. Ask for feedback on a business plan in Hindi and you get more warmth. Ask in Russian and you get more rigour. Same model, same question, different lean.
My favourite detail, and I promise I’m not making this up: Claude leans hardest toward candor when it speaks Dutch. So my language is quietly making the model own its mistakes more often than yours does. You’re welcome. But the point isn’t the trivia. The point is that if something as passive as the language you happen to type in already shifts the values in the output, then context you supply on purpose is a far bigger lever. A Brand Profile is you grabbing that lever instead of leaving it to chance.
What a Brand Profile actually is
A Brand Profile is the document you hand a model so it stops guessing. Not a logo and a tagline. The real stuff:
- Tone of voice. How you sound. Dry, formal, blunt, warm. Mine is apparently “sarcastic Dutchman who won’t shut up about sneakers.”
- Background. Where the business comes from, what it’s done, who it serves. And, if you’re brave, who you are as a person behind it.
- Core values. What you’ll say yes to and, more usefully, what you’ll refuse.
- Reference points. Real examples, real clients, real before-and-after. The things a stranger couldn’t invent.
That last one is where the profile earns its keep. Because it does two jobs at once: it makes output sound like you, and it stops the model making things up.
The sneaker test
Here’s what I mean, and yes, we’re doing sneakers.
Say I run a small sneaker shop and I ask an LLM to write a product description for a pair of Air Jordan 1s. With no profile, I get the median result: “Step up your style with these iconic kicks that blend comfort and heritage.” I have read that sentence four thousand times. It is beige. It could be selling shoes, mattresses, or life insurance.
Now I give it my Brand Profile. It knows my shop only deals deadstock, that I write like a collector and not a catalogue, that I care about the story of the 1985 ban far more than the outsole tech, and that my customers already know their sizing. Same request, different answer: it opens on the ban, name-drops the exact colourway, skips the comfort waffle because my audience finds it insulting, and lands on why this pair is worth holding rather than wearing.
One of those descriptions gets scrolled past. The other sounds like a human who actually knows shoes wrote it. The only variable was context. Here is an overview of what I mean and where you can see the differences:
| No profile | With Brand Profile | |
|---|---|---|
| Relevance | Median, generic | Fits the audience |
| Creativity | Safe, recycled | Angle you’d actually use |
| Hallucination risk | High, invents specs | Low, grounded in real detail |
| Sounds like you | Never | Consistently |
Fewer hallucinations, and this is the part people underrate
When a model doesn’t know the specifics, it fills the gap. Confidently. That’s where the invented product features, the wrong founding year, and the fake client quote come from. A Brand Profile shrinks the gap. You’ve handed over the facts, so the model retrieves instead of inventing. Grounded input, grounded output. It’s not magic, it’s just not leaving blanks for the model to paper over.
Automation isn’t limited by how clever the model is anymore. It’s limited by data: how well you turn raw, messy source material into input a system can actually use. A good Brand Profile is that translation done deliberately. It’s the difference between automation that sounds like your brand and automation that sounds like every other brand that skipped this step.
So build one, before the noise decides for you
Most audiences are drowning in AI output right now, and I’m not convinced they can still tell what’s good from what’s merely fluent. The brands that win the next stretch won’t be the ones shouting loudest about tools. They’ll be the ones whose content is unmistakably theirs, because they took an afternoon to tell the machine who they are.
Your LLM doesn’t know you yet. That’s not the model’s failure. It’s an empty field you haven’t filled in, and empty fields get filled with the average of everyone else. Fill it yourself, deliberately, and average stops being your default answer. That’s the whole job.
