Brand Voice for SaaS: Why Your AI Copy Sounds Generic

You asked an AI assistant to write your landing page. What came back was fluent, grammatical, and completely interchangeable. Swap your product name for a competitor's and not one sentence would need editing.

That is not a prompting failure. Your brand voice was never written down, so the model produced the statistical average of every SaaS landing page it has ever read — which is exactly what you asked for when you gave it nothing else to go on.

Personas tell a model who is reading. Positioning tells it what you claim. Nothing in either one tells it what you sound like.

Who has this problem

You have shipped something real. You have a product page, maybe a few blog posts, and you generate marketing copy with Claude, ChatGPT, or Cursor because writing it by hand is slow and you are not sure what good looks like anyway.

The output is never wrong. It is just anonymous. You read it, feel vaguely dissatisfied, tweak two words, and ship it — because you cannot articulate what is missing, and "make it sound more like us" produces a different flavour of the same thing.

If you have ever rewritten the same AI paragraph three times without being able to say what you were fixing, this is the gap.

Why every AI draft drifts to the same voice

A language model predicts the most probable next token given its context. With no constraint on style, the most probable marketing sentence is the one closest to the centre of all marketing sentences. That centre is where "seamlessly integrate," "empower your team," and "take your growth to the next level" live.

The fix is not a better prompt. It is a constraint the model can apply mechanically, and there are three reasons founders never write one:

Brand voice is taught as a feeling, not a field. Most guidance tells you to pick adjectives. "We're friendly, innovative, and authentic." Every one of those is unfalsifiable. No sentence has ever been rejected for insufficient authenticity, which means the adjective changes no output.

The document lives in the wrong place. A founder who does write a tone-of-voice doc puts it in Notion. It is not in the prompt when the draft is generated, so it governs nothing.

Nobody records the reasoning. A rule with no rationale gets rounded off the first time it is inconvenient. "Don't say unlock" survives one draft. "Don't say unlock, because it names a feeling instead of a mechanism and our audience reads it as the register of the people who failed them" survives indefinitely.

How to define a brand voice a model can obey

The goal is a set of fields that decide sentences. Work in this order — it is ordered by how much each step changes the output, not by how much thinking it requires.

Step 1: Pick an archetype instead of inventing a voice

Choosing from twelve established archetypes — Sage, Creator, Hero, Outlaw, Explorer, Magician, Everyman, Caregiver, Ruler, Innocent, Jester, Lover — is far easier than describing a voice from nothing, and it decides more than it looks like it does.

A Sage teaches and cites evidence, and never uses urgency tactics. A Hero challenges and rallies. Pick the one closest to how you already talk to customers, not the one you find most flattering.

The Mark Platform lists all twelve with what each promises, how it changes your writing, and the failure mode when overplayed. A Sage overplayed lectures and never asks for the sale. Knowing your own failure mode is half the value.

Step 2: Write contrast pairs — this is the highest-value field

A contrast pair is "we are X, not Y." It is the single most useful thing you can record, because it draws a boundary an adjective cannot.

• Direct, not blunt
• Methodical, not rigid
• Confident, not boastful

"Direct" alone permits a sentence that is contemptuous. "Direct, not blunt" rejects it. Write three to six of these and you have done most of the work — about two minutes of thinking that changes every draft afterwards.

Step 3: Record rules with a severity and a reason

Split what you record into three kinds: words to avoid, claims you may never make, and things to always do. Give each one a rationale and mark it hard or soft.

Hard rules are compliance or brand-critical and block a publish. Soft rules are preferences that warn. Treating "never promise a revenue figure" and "we prefer 'system' to 'framework'" as equally weighted is how the important one gets ignored.

Rules must be added, never re-submitted as a list. Sending the whole list back to change one item is how twenty-nine banned words silently disappear when you add the thirtieth.

Step 4: Save on-brand and off-brand pairs

This outranks everything above it. Take one message and write it twice: the way you write it, and the way you do not.

On-brand: "You shipped your product. Nobody bought it. You're not bad at this. You were never taught it."

Off-brand: "Let's face it — most developers are terrible at marketing. Here are 7 growth hacks that will change everything."

Every other field describes the voice. A pair shows it. A model given one message written well and badly infers the boundary far more reliably than one handed a list of adjectives — and the contrast is where the rule actually lives, so never save the good version alone.

Step 5: Set the mechanical switches

Formality, contractions, emoji policy, reading level, and grammatical person. These are dull and they are checkable, which is the point — a machine can enforce them without judgement, leaving your attention for the parts that need taste.

Step 6: Put the voice in the prompt, not in a document

This is the step that makes the other five matter. The voice has to be assembled into instruction text and injected into every generation — copy, outreach, ad drafts, image prompts, video scripts — from one place.

The Mark Platform builds that block from your saved fields and attaches it to every draft it generates, with your worked examples placed last, closest to the task, because that is where few-shot examples work hardest.

Why this framework holds up

Voice work usually fails because it produces a PDF. This version produces fields, and fields can be applied by a machine and checked afterwards.

That checking matters more than it sounds. The platform scores a draft against your rules and reports which one it broke and why — banned words, forbidden claims, emoji policy, contractions, filler. It deliberately does not score whether copy "feels right," because an invented taste score would get trusted as though it were measured. Taste is what your contrast pairs are for, applied by you.

Where this will not help: a brand voice cannot rescue a product nobody needs, and it cannot substitute for a positioning statement that names a specific customer. If your copy is anonymous because you do not yet know who it is for, voice is the wrong layer to fix — start with your ideal customer profile instead. Voice decides how you say it. It cannot decide what you have to say.

Key takeaway

Adjectives do not constrain a model; contrasts and worked examples do. Write three "X, not Y" pairs and two on-brand/off-brand samples, and your generated copy stops being interchangeable.

Frequently asked questions

How long does this take? The archetype and three contrast pairs take about two minutes and deliver most of the benefit. Rules and samples accumulate naturally — every time you correct a draft, you have found one.

Should I let AI write my brand voice for me? No. An invented voice that gets saved is worse than no voice, because it is wrong and it looks decided. A model can ask you the questions and structure your answers, but the answers must be yours.

What if I run several products? Each product gets its own voice. Two products sharing one brand record means every draft sounds like whichever one you filled in last — which is why the constraint is one brand per product.

Does this apply to images and video too? Yes, and it is where the effect is most visible. Typefaces, logo placement, palette, and imagery rules feed the same generation path, which is what stops every generated creative looking like a template.

---

You have the framework. The Mark Platform stores your archetype, contrast pairs, rules and worked examples, then attaches them to every draft it generates and scores the result against them. See what else it does, or start your marketing journey →