Your Brand Voice Guide Is Invisible to AI. Fix It With Machine-Usable Rules.

By the end of this post you will have converted your existing brand voice guide into enforceable rules that AI writing tools actually follow, and you will know where those rules belong in a content pipeline. You need your current voice guide (even if it is only bullet points), your last three published blog posts, and thirty minutes.

1. Run the blank-doc test before you rewrite anything

Paste your last three published blog posts into one document. Strip the company name. Read what remains and ask whether any competitor in your space could have published it. If the answer is yes, your voice guide is not reaching the page, no matter how well it reads to a human editor.

This is the diagnostic Bartlett calls the blank-doc test [1]. It exposes whether your voice is actually present or merely assumed. A voice guide that lives in a PDF your team admires but your AI never sees produces exactly this generic output.

The pitfall: treating the test as a style exercise rather than a diagnostic. The point is not to make the stripped text sound better. It is to discover that your current guidelines are invisible to the machine generating the draft.

2. Why adjectives fail and rules work

A voice guide that says "professional, warm" tells a human what to aim for and tells a machine nothing. The gap between documented guidelines and machine-usable rules explains why, per research cited by Averi, 64% of the most successful content marketers have voice guidelines but only 23% use them to train AI tools. Machines need do/don't instructions, not mood boards.

Concrete constraints work: ban hedging words like might, could, and somewhat; prefer verbs like decide, build, and ship [1]. These are enforceable. An AI can check whether "might" appears in paragraph one. It cannot check whether the paragraph feels warm.

The pitfall: rewriting adjectives into longer adjectives. "Professional, warm" becomes "confidently approachable" and the machine still has no instruction. Replace each tone word with a sentence pattern that demonstrates it and one that violates it.

3. Convert your existing guide in four steps

Step 1: Extract every adjective and replace it with a sentence pattern. For each tone word, write one sentence structure that demonstrates it and one that violates it. "Professional" becomes "State the outcome in the first clause, then the method" versus "Open with how you feel about the topic."

Step 2: Build a three-column vocabulary reference. Glean recommends preferred terms in the first column, banned phrases in the second, and a use-sparingly list in the third. The third column matters most at scale. Words like "leverage" and "empower" creep back in unless they are explicitly rationed.

Step 3: Add annotated on-brand/off-brand example pairs, Atom Writer recommends at least 4, each with a note on why the off-brand version fails. Four pairs beat a hundred adjectives because they show the boundary. The annotation is the teaching signal; without it, the AI sees only preference, not principle.

Step 4: Assign numeric tone positions by scenario if your voice shifts by channel. A product announcement might run at 7/10 energy and 4/10 formality; a technical deep dive at 3/10 energy and 8/10 formality. Numbers prevent drift where adjectives collide.

The result should fit in 2–3 pages. Longer documents get ignored by humans and truncated by machines.

The pitfall: preserving the original document's structure because it took effort to produce. Start fresh. The machine-readable version is a different artifact with a different job.

4. Embed the rules where generation happens

A 40-page PDF in your brand folder is invisible to the tool writing the draft. Store the 2–3 page rule set as persistent context, custom instructions, or a system prompt, wherever the AI reads before it generates.

In ContentRails, the Project DNA document carries audience and voice summary, and standing custom instructions carry the sentence-level rules [2]. Both are read at generation time, not pulled in after the draft exists. The method in this post is how a user builds those instructions. Where other guides end with "paste this into ChatGPT," this is where the rules land in an actual content pipeline.

The pitfall: treating the rules as a post-generation checklist. If the AI does not see them before it writes, they are editing notes, not instructions. Editing notes cost human time. Instructions cost machine time, which is the point.

5. Score drafts against human examples

Glean recommends a five-dimension rubric for AI drafts: voice match, factual accuracy, clarity, structure, and audience fit. Run it side by side with a strong human-written post, not against the voice guide itself. The test is whether the draft sounds like your best work, not whether it avoids banned words.

A draft that passes every rule but reads like a competitor's content still fails. The blank-doc test from step one is the final filter.

The pitfall: over-optimizing for rule compliance. A post with zero banned words and perfect numeric tone scores can still be generic. The rubric is necessary, not sufficient.

6. What instructions cannot do

Custom instructions shape how something is said; they cannot make a tool invent expertise or unsourced figures [2]. Voice rules improve distinctiveness, but they do not replace topical authority or structured data [1]. If your team lacks subject-matter depth on a topic, the best voice guide in the world will produce polished but empty copy.

This is a hard limit, not a temporary one. Instructions are a form factor, not a substitute for knowledge. Budget for human expertise on topics where your authority matters, and for human editing on everything else.

The pitfall: believing voice consistency fixes content quality. It fixes content recognizability. These are related but not identical.

7. Check it monthly

Run the blank-doc test again once the rules have been in force for a while [1]. Track AI share of voice by running target queries in ChatGPT, Perplexity, and Google AI Overviews and noting how AI describes your brand [1]. If the description is generic, your rules are still too abstract. Tighten one sentence pattern and retest.

Monthly checks catch drift before it compounds. Voice rules decay like any other code: the environment changes, new writers add exceptions, and the original intent erodes.

The pitfall: running the test only when output feels wrong. By then the decay is visible to readers too.

Where this lands in a content pipeline

ContentRails builds this into its pipeline in two places. The Project DNA document stores your audience, voice, and already-covered topics, and every post follows it [2]. Standing custom instructions (up to 2,000 characters for topic research, writing, and rewrites) let you embed the concrete rules this post describes directly where the tool reads them [2]. If you want to see how the pipeline works, start at contentrails.ai.

Frequently asked questions

How many content marketers actually use their brand voice guidelines to train AI tools?

Per research cited by Averi, 64% of the most successful content marketers have documented brand voice guidelines, but only 23% actively use them to train their AI tools.

Can custom instructions in ContentRails override the Project DNA voice profile?

No. Custom instructions shape sentence-level behavior; the Project DNA sets audience, voice summary, and topic coverage. They operate at different layers and are read together at generation time [2][3].

What happens if I ask ContentRails to invent a statistic in my custom instructions?

The instruction will not override the sourcing constraint. ContentRails discards any figure it cannot match to a source, regardless of what custom instructions request [2].

Does the free plan include standing custom instructions and Project DNA editing?

Yes. The free plan at $0/month covers 1 project and 3 posts a month, and includes Project DNA editing and standing custom instructions up to 2,000 characters per step [4].

Does consistent brand voice improve AI search citations?

Yes. In one analysis of B2B SaaS sites, pillar-organized content with consistent topical voice achieved a 41% AI citation rate versus 12% for standalone pages [1]. Content with statistics and primary sources gets 30–40% more visibility in AI responses [1]. Voice alone does not earn citations, but voice consistency combined with sourced authority and structured data does.

Run the blank-doc test on your last three posts this week. If the voice is invisible, open your voice guide and rewrite the first adjective into a sentence-level rule you can paste into custom instructions today. Then schedule a monthly check: strip the company name from the next draft before it goes live and see if it still sounds like you.

Sources

1. Brand Voice for AI Search: Build One That Gets Cited (bradleebartlett.com)
2. contentrails.ai
3. contentrails.ai
4. contentrails.ai