You've felt it. You ask ChatGPT for a LinkedIn post about your product. It produces something competent, grammatical, and completely soulless. You read it back and think: "this could be from anyone."
That feeling has a name in the AI industry. It's called generic AI voice — and it's not a glitch. It's a feature.
Why AI defaults to generic
Large language models like GPT-4, Claude, and Gemini are trained on enormous amounts of text from the internet. When you give them a prompt, they predict the most statistically likely next word, then the next, then the next.
Here's the problem: the most statistically likely next word is almost never the most interesting one. It's the safe one. The one that appears most often in training data.
When millions of LinkedIn posts, blog articles, and marketing emails feed into a model, the average is what you get back. And the average reads like: "In today's fast-paced world, businesses must innovate to stay ahead."
This isn't a writing style. It's the absence of one.
What "brand voice" actually means
Real brand voice is a set of constraints — choices about what NOT to say.
When Linear writes documentation, they don't use words like revolutionary, unleash, or empower. When Patagonia writes copy, they don't talk about unprecedented value propositions. When MailChimp writes onboarding flows, they're warm without being precious.
These aren't stylistic flourishes. They're rules, often unwritten, that the brand follows consistently.
The problem with generic AI prompting is you're asking the model to write like you without giving it any of those rules. You're asking it to predict what your specific brand would say, but the training data doesn't include "what your specific brand would say." It includes "what most brands say." So you get the most.
What actually fixes this
There are three approaches that work, in order of effectiveness:
1. Few-shot prompting with real samples
Take 3-5 paragraphs your brand has actually written. Paste them into the prompt. Tell the model: "Write in this exact style."
This works moderately well for one-off pieces. The model matches surface-level patterns — sentence length, vocabulary, rhythm.
The limitation: you have to manually do this every time. And the model still doesn't understand the brand — it just mimics patterns.
2. Forbidden words and required vocabulary
Give the model an explicit list:
- "Never use these words: revolutionary, leverage, unlock, game-changer..."
- "Always prefer: ship, build, fix, sharp..."
Combined with samples, this dramatically reduces generic output. The model can't fall back on its default vocabulary because you've listed those words as forbidden.
This is closer to how brand voice actually works in practice — as a set of constraints.
3. Structured brand profile extraction
The most effective approach: extract a brand voice profile from existing content (a website, past blog posts, marketing copy), then use that profile as system context for every generation.
A proper brand profile includes:
- Tone characteristics (3-5 specific tags)
- Audience description
- Sample voice (a real paragraph in the brand's voice)
- Preferred vocabulary
- Forbidden phrases
- Customer personas
- Pain points the brand solves
- Topics to always mention / never mention
When this entire profile feeds into the prompt as structured context — not just "write like Linear" but a full breakdown of what makes Linear sound like Linear — the output dramatically improves.
The "voice infrastructure" problem
This is exactly what most AI marketing tools miss. They treat brand voice as a prompt-tweaking exercise — something you do each time you generate content.
It should be infrastructure. Set up once, applied automatically to every piece of content you create.
Tools like Hey Molly handle this automatically: when you connect your domain, the system reads your homepage, key pages, and existing content. Then it builds a complete brand voice profile that gets injected into every generation — blog posts, social media, emails, repurposing — so the AI never starts from scratch.
The result isn't "AI content that sounds less generic." It's content that sounds like you.
How to test if your AI output is actually on-brand
Before publishing AI-generated content, run this 30-second test:
- Hide the byline. Could a colleague guess this is from your brand without seeing your name?
- Search for forbidden phrases. Does it contain any of these? "In today's fast-paced world", "Unlock unprecedented", "Game-changer", "In an ever-evolving landscape"
- Read it out loud. Does it sound like something a person from your team would actually say?
- Compare to your last 3 published pieces. Same vocabulary? Same rhythm? Same opinions?
If any of these fail, the content isn't ready. Either tighten the prompt, paste more examples, or use a tool that handles brand voice systematically.
The bottom line
AI content sounds like AI for a technical reason: models default to averages. The fix isn't a magic prompt. It's giving the model enough constraints, samples, and structured context that it can't fall back on those averages.
Set this up properly — as infrastructure, not per-prompt — and the difference is dramatic. AI content stops feeling like AI and starts feeling like another writer on your team. One who happens to never sleep.
If you want to try this with your own brand, Hey Molly extracts your brand voice from your website automatically — full profile in 60 seconds, applied to every piece of content you generate. See the pricing.