The brand-voice problem in AI marketing (and how to fix it)

"Why does my AI marketing sound off?" is the question we get most after teams have used AI marketing automation for 60 days. The answer is rarely the model. It is almost always the voice scaffolding around the model. Below is what works.

Why AI marketing sounds generic by default

Foundation models are trained on the entire internet. The internet is full of marketing copy. Most marketing copy is generic. The model's prior is "average marketing voice" — exclamation marks, "elevate," "transform," "unlock," "synergy." Without strong counter-signals, the output averages toward generic.

The fix is to give the model strong counter-signals about what YOUR voice is. Four things do this.

1. Training examples (the foundation)

Drop in 50-100 examples of your past content. Real artefacts:

• Your best 20 social posts
• Your last 10 newsletter intros
• Your About page
• 3-5 blog posts you wrote (not ghost-written)
• Your last 10 WhatsApp broadcasts
• Customer-facing emails you sent yourself

Less than 50 → the AI averages into generic. More than 100 → diminishing returns. Quality over quantity: avoid examples where you wrote in a register you no longer use.

Update quarterly. Add new strong examples; remove anything stale.

2. Voice rubric (the spec)

A one-page document the AI references on every generation. Five sections:

Tone — 5 we are, 5 we are not

Specific adjectives. "Warm but technical" is better than "professional." "Direct, not curt" beats "honest." Examples we use ourselves: warm, plain, specific, dry-humored, technically honest. Not: enthusiastic, salesy, jargon-heavy, exclamation-marky, hyperbolic.

Vocabulary — words we use, words we don't

Concrete list. "Use" not "leverage." "We" not "the team." "Real" not "world-class." This is the part most teams skip and it's where the most visible damage shows up.

Sentence rhythm

Preferred sentence length (12-18 words average for us). Paragraph cadence (3-5 sentences, occasional one-line paragraphs for emphasis). Comma policy (we like Oxford). Em-dash policy (sparingly; never stacked).

Reference examples

Five pieces of your own content marked "good." The AI looks at these on every generation as anchor points.

Anti-examples

Two or three pieces marked "off-brand." Generic LinkedIn motivational. Sales-spammy email. The AI uses these as negative examples.

See how to write good prompts for the broader technique.

3. Drift detection (the eval loop)

Voice drifts. Model updates, prompt tweaks, new training examples, edge cases that ship through approval — each can nudge the voice off-axis.

Catch drift with a weekly sample:

1. Pull 5 random AI-generated pieces from the week.
2. Score each against the voice rubric: 1-5 on each section.
3. Log scores in a running spreadsheet.
4. If average drops more than 0.5 over 4 weeks, voice training needs attention.

Done weekly, this takes 20 minutes. Done quarterly, you discover drift after the damage has shipped.

4. Quarterly voice audit (the fresh-eyes pass)

Every 90 days, someone outside the day-to-day operation reviews 20 pieces:

• 5 most-engaged posts
• 5 least-engaged posts
• 5 random posts from the middle
• 5 most-recent posts

Fresh eyes catch drift the daily reviewer becomes blind to. Score each piece against the rubric; write a one-paragraph summary of what's working, what's drifting, what to adjust.

Update the rubric or the training examples based on findings.

What to do when voice drifts

Three diagnostics in order:

1. Check the training examples. Did you add stale or off-brand pieces recently? Pull them.
2. Check the rubric. Has your actual voice shifted (e.g., you became more technical) faster than the rubric? Update it.
3. Check the prompts. Has someone tweaked the campaign prompts in a way that drowns out the voice signals? Compare against the previous month's prompts.

Almost never: change the model. That's the panic move; it rarely fixes voice.

The hardest cases

Brands with split voice

A B2B SaaS that sells to engineers (technical voice) and CFOs (formal voice). Solution: split into two voice profiles, two rubrics. Tag content by audience; AI uses the right rubric.

Brands where the voice IS the founder

Personal-brand businesses, creator economy. The voice is hard to capture from artefacts because so much is delivery (timing, intonation). Solution: founder writes the headline + first line. AI writes the body. Founder approves.

Brands with regional voice variation

Same brand in India vs the US — the voice rubric is similar but the cadence, references, and vocabulary differ. Solution: regional rubrics. The trunk is shared (brand promise, anti-examples); the leaves are regional.

What to measure

• Voice rubric average score. Weekly. Target: 3.8+ on a 5-point scale.
• Approval pass rate. % of AI output shipped as-is. Target: 70-85%. Below 50% = voice training problem.
• Edit distance. When a human edits, how much do they change? Track average edit %. Should decline over time.
• Customer feedback on tone. When customers describe your content, are they using words from your "tone we are" or your "tone we are not" list?

What this means for you

• Four-part discipline: training examples, voice rubric, drift detection, quarterly audit.
• 50-100 real examples is the training floor. Less averages into generic.
• Weekly sample-and-score. 20 minutes. Catches drift before it ships.
• Voice problems are almost never solved by changing the model.
• Marketing Autopilot bakes the rubric + drift detection into the platform.
• Read our AI marketing automation pillar for the broader frame.

Want a second pair of eyes on your brand voice rubric? Book a 30-minute call. We will pressure-test it.