How do you get AI to actually sound like your brand?

A brand voice document is the instruction artifact you hand an AI tool so that everything it writes comes out in your register rather than its default one. This page builds one backwards from evidence you already own — your published writing — and then puts it through a test that can fail: extract rules by contrast, cut them to a budget the model can hold, install them where the tool applies them to every generation, and blind-test the output against human writing you held back.

When should you use this — and when should you not?

Use it when you generate the same kind of content repeatedly, have an archive of approved work, and can borrow two or three colleagues for an hour of judging. Budget three to four hours for the document and two more for a test round. Stop and fix something else first if:

• You have no corpus. Below roughly 15-20 genuinely on-voice published pieces there is nothing to extract. That floor is practical judgement — no study we found sets one.
• Your voice is not distinctive. Ten-minute pre-check: strip names, products and URLs from three of your pieces and three of a competitor's, and ask a colleague to sort them by company. If they cannot, the document will faithfully encode blandness. In a survey of 2,003 B2B marketing leaders commissioned by Optimizely — a content-software vendor with a direct interest in the finding — only 30% called their brand voice "genuinely unmistakable" and 15% said their content "wouldn't survive" a competitor logo swap (Optimizely 2026 study; the fieldwork was "conducted by Savanta between May and June 2026" per the methodology note, and the confidence figures are independently reported by MarTech).
• Your target voice fights the model's defaults. In a 2026 preprint across 11 models and 16 tasks, constraints opposing default behaviour failed at rates from 10% to 100%, while conventional ones hit 99%+ regardless of formatting (arXiv 2604.07192 — measured in code generation, not prose). Fragments, profanity and eight-word sentences are an editing problem, not an instruction-layer one.
• Your register is informal and personal. Authorship verification of LLM imitations ran 95-97% on news and email but 17-66% on blogs and forums (arXiv 2509.14543). A founder's newsletter is the hard case.
• You cannot hold back 20 human pieces or run 20 briefs. Run the test anyway, but as a smoke check for gross failure — and label it that way.

What do you need before you start?

The corpus is what people get wrong. Take 15-30 published, approved pieces from one channel family — all long-form or all email, since mixing channels averages incompatible registers — preferably from the last 18 months and from more than one author, so you capture the brand rather than one writer's habits. Exclude anything ghostwritten, committee-edited or AI-assisted (an exclusion rule set out by MindStudio): AI-assisted pieces make you extract the model's defaults and label them your voice. Annotate each with one line on why it works — Glean makes the useful point that vocabulary appearing consistently across teams "without anyone coordinating it" is "often the truest signal of your voice".

Then gather 5-10 off-voice pieces for contrast, and separately hold back 20 on-voice human pieces that never enter the corpus and are never shown to the model. Those are the control arm; without them the result cannot be read at all. For judges, take two or three people who read your output often and did not write the document — the evidenced configuration is a five-person expert panel (below), so this is a practical compromise, not a finding.

How do you do it, step by step?

How do you pull the rules out of your own writing?

Do not ask a model "what is my brand voice?" — you get four adjectives that fit every brand alive. Ask for observations that discriminate your corpus from off-voice text:

You are analysing writing style. Do not summarise content or evaluate quality.

CORPUS A (our published brand writing):
[PASTE 15–30 PIECES, SEPARATED BY ---]

CORPUS B (off-voice: competitor copy / generic AI content / old brand):
[PASTE 5–10 PIECES, SEPARATED BY ---]

Identify style features that DISTINGUISH A from B. Ignore any feature that is
equally true of both. Ignore subject matter entirely.

Cover: sentence length and variation; paragraph length; opening moves; person
and address (I/we/you); contractions; hedging vs assertion; jargon density;
humour and its type; metaphor and its source domains; punctuation habits;
list vs prose; how claims are evidenced; how pieces close; recurring phrases;
words that appear in A but never in B, and vice versa.

Output a markdown table, one row per feature, exactly these columns:
| Feature | Rule as an instruction | Evidence quote from A | Counter-quote from B | Checkable? (yes/no) |

"Rule as an instruction" must be a positive directive, not a prohibition, and
specific enough that two people reading it would edit the same sentence the
same way. "Checkable?" = yes only if a reader could verify compliance in one
pass without knowing our brand. Produce 20–30 rows. Never include a rule you
cannot quote evidence for.

Expect a third of the rows to be junk; this is a generation step. While you are there, count two or three things by hand: mean sentence length, share of sentences under ten words, contraction rate, em-dash frequency. Classical stylometry identifies authors from countable surface features rather than from adjectives about tone — its standard feature space is the most frequent words and character n-grams (Eder et al., The R Journal); the four habits above are a spreadsheet-grade stand-in for that idea, not features that paper measures. A number in a voice document beats an adjective because it is checkable.

How many rules should the document hold?

Sort the candidates by checkable, then by discriminating (is there a real counter-quote), then by how often generic AI content violates them. Keep the top five to seven; delete the rest.

Compliance degrades as constraints stack: on the FollowBench benchmark, GPT-4's hard satisfaction rate fell from 84.7% at one constraint to 61.9% at five, and the authors conclude that "even the most sophisticated models are limited to following instructions with about three constraints" (Jiang et al., ACL 2024). Those are 2023-vintage models and no 2026 replication was found — the shape is durable, the figures are dated.

What fails is narrower than the folklore. Style constraints in isolation are among the easier categories in that same benchmark (GPT-4: 96.7% at one constraint, 90.0% at five). Models can do style; stacking is what breaks — a 40-rule voice bible competing with the brief, the format spec, the SEO requirement and the word count for one compliance budget.

How do you order and phrase them?

# Brand voice — [BRAND]

## Non-negotiables (max 3)
1. [MOST IMPORTANT RULE]
2. [SECOND]
3. [THIRD]

## Voice rules
4–7. [REMAINING RULES — each: directive + one line on why]

## Measurable targets
- Mean sentence length: [N] words; under [X]% of sentences over [M] words.
- Contractions: [always / never / in these contexts]

## On-voice examples
[2–3 EXCERPTS FROM CORPUS A, 60–120 words each]

## Off-voice examples, with the reason
[2–3 EXCERPTS FROM CORPUS B]
Why this is off-voice: [ONE LINE NAMING THE VIOLATED RULE]

## Avoid
- [3–5 SPECIFIC BANNED MOVES, each paired with its replacement]

Three choices there are load-bearing. Most important rules first, repeated at the end: model performance is highest when relevant information sits at the beginning or end of the context and "significantly degrades" mid-context (Liu et al. — a retrieval finding used here as a layout heuristic, not a measured style result). Positive phrasing, every prohibition paired with its replacement: "don't be corporate" fights a default and offers nowhere to go, while "open with a concrete customer situation" gives a target. Two independent sources agree — the constraint study above, and OpenAI's own advice to "prefer positive, concrete instructions ('Do X') over long lists of prohibitions" (Creating a GPT) — though neither measures prose. Two or three examples, not twenty: varying in-context writing examples across 2, 4, 6, 8 and 10 "affects the four metrics very little, suggesting limited gains in stylistic alignment" (arXiv 2509.14543).

Where do you install it?

Every mainstream assistant has a persistent instruction slot and a separate knowledge or file area. The document goes in the instruction slot. Uploading it as a PDF into project files makes the rules retrievable instead of always applied; OpenAI states the split directly — "use knowledge for reference material, not rules or behavior. Put rules, tone, and workflow guidance in instructions" (Creating a GPT). Interface details as of 2026-07-27:

Tool	Instruction slot	Documented length cap	Worth knowing
Claude	Project instructions	none published	Apply only to chats inside that project
ChatGPT (account)	Custom Instructions	1,500 characters on Free/Go; 5,000 on Plus, Pro, Enterprise, Business, Edu	The cap is tiered by plan, not a change over time
ChatGPT Projects	Project instructions	none published	They "override your global custom instructions" — useful for per-client voices
Custom GPTs	GPT Builder → Instructions	none published	GPTs "do not use saved memory, custom instructions, or previous conversations" (FAQ) — paste the document in again
Gemini Gems	New Gem → instructions	none published	Official structure guidance: persona, task, context, format

No vendor publishes a cap for project-scoped instructions, and the "8,000 characters for custom GPTs" figure circulating in forums appears in no documentation. Where a tool has no persistent slot, paste the document at the top of the conversation and restate the three non-negotiables immediately after the brief — that puts them at the end of the context.

How do you run the blind test?

Run 20 briefs through the configured tool, one output each, no cherry-picking and no editing. Mix them with the 20 held-back human pieces, trim everything to comparable length, strip product names, dates and links, shuffle, label the items 01-40 and keep the key separate. Judges must not know which items are which, or the ratio.

Each judge answers two questions per item, independently and without discussion: would you publish this under our brand as it stands? and human-written or AI-generated? Who judges changes the answer. Among 300 non-fiction articles, a majority vote of five annotators who use LLMs heavily for writing misclassified only 1 of 300, outperforming commercial detectors and holding up against paraphrased output (Russell et al., ACL 2025). General populations sit near chance — one 2025 study of thesis excerpts reports about 57% recognition of AI texts and 64% of human texts (study, a weaker source cited here only for direction). Naive judges hand you a comfortable, meaningless result.

Record five numbers: the pass rate on each arm, the share of AI items correctly identified, the agreement between judges, and N.

What does a good result look like?

A finished rule carries a directive, a reason, evidence, counter-evidence and a one-pass check:

3. Open with a situation, not a claim. [NON-NEGOTIABLE]
   Why: our openings earn attention with a specific moment; claims read as ads.
   From our writing: "The invoice arrived on a Friday, which is how these
     things always start."
   Off-voice: "In a fast-moving landscape, invoicing is more critical
     than ever."
   Check: the first sentence names a concrete person, moment or object.

The test collapses into one table. Fill it with your own numbers; the point is that no cell is missing.

Measure	AI arm	Human control arm
Items judged (N)	[N]	[N]
Would publish as-is	[%]	[%]
Correctly identified as AI	[%]	—
Inter-judge agreement	[kappa or alpha, with judge count]

A result without the control column is not a weaker result; it is an uninterpretable one.

How do you know the output is good?

• Every rule in the finished document carries a verbatim quote from your own published work and a counter-quote from off-voice text; count the rules and count the ones with both — the two numbers match.
• The document holds 7 rules or fewer, at most 3 of them marked non-negotiable, and those 3 appear in its first 200 words.
• No step in your workflow uses the model's own score of its own output as evidence that the voice landed.
• The blind test runs at least 20 AI items against at least 20 held-back human items, judged independently by at least 2 blinded judges, and inter-judge agreement reaches Cohen's kappa 0.61 or Krippendorff's alpha 0.67.
• In that test the AI arm's pass rate lands within 15 percentage points of the human control arm, and judges correctly label fewer than 15 of the 20 AI items — at 20 items, 15 is the point where guessing stops being a plausible explanation.
• The recorded result states N; if you ran fewer than 20 items per arm it is labelled a smoke test rather than evidence that the arms are equivalent.

Three of the criteria above are settled by counting. The ones about the test each exist because of a specific way of fooling yourself.

The control arm. A 65% pass rate on AI content means nothing alone. If the held-back human pieces also pass at 68%, the AI is doing fine and your judges are strict; if they pass at 95%, you have a real gap.

Chance correction. When a judge picks between two options, guessing alone scores about 50%. A threshold like "below 60% identification means your voice lacks distinctiveness" — the bar set in one widely circulated version of this test (Atom Writer) — never subtracts those free hits. At 20 items, judges need 15 correct (75%) before guessing stops being a plausible explanation; at 40 items, 26; at 100, 59. That is our own calculation against the binomial baseline, run 2026-07-27, and it is why the bar above is a count rather than a percentage. The arithmetic cuts both ways: "they couldn't tell" at ten items per arm is nearly empty, because the 95% interval around an observed 50% then runs from roughly 24% to 76%. Failing to find a difference is not evidence of sameness.

Sample size. By our own power calculation (Fisher exact, two-sided, 2026-07-27), 20 items per arm catches a catastrophic voice failure — a 0.5 versus 0.9 gap — about three times in four; ten items per arm catches it less than one time in three. Separating a good result from a very good one (0.7 versus 0.9) costs far more than it looks: 40 items per arm catch that gap only about half the time, 50 just under two times in three, and holding it to the same three-in-four standard takes just over 60. The test catches "the voice is broken", not "the voice is 92% right".

Agreement between judges. If judges disagree, you are measuring their taste rather than your voice. Two judges and a yes/no rating: Cohen's kappa. Three or more: Krippendorff's alpha. Convention puts kappa 0.61-0.80 at "substantial"; for alpha, the K-Alpha authors call ≥0.80 "a satisfactory level of agreement", read 0.67-0.79 as "the lower bound for tentative conclusions", and treat anything below 0.67 as "unreliable for drawing triangulated conclusions" (Marzi et al., MethodsX). Below that, rewrite the rules until they are checkable and re-run.

No self-grading. The sweeping version of this warning is false: strong LLM judges match human preferences on general answer quality at "over 80% agreement, the same level of agreement between humans" (Zheng et al., NeurIPS 2023). Two narrower things break. A model grading its own output overestimates its compliance relative to rule-based scoring (arXiv 2604.07192), and judges reward text familiar to them — they "assign significantly higher evaluations to outputs with lower perplexity than human evaluators" (Wataoka et al.), which is precisely the generic register you are escaping. Voice-match judgement is where measurement falls apart outright: a 2026 preprint evaluating 50 authors and 1,000 generations across three metric traditions found near-zero correlation between them (absolute r below 0.07), with "an LLM judge declar[ing] a clear winner while LUAR finds no meaningful differentiation" (Sawant, arXiv 2604.26460 — a single non-peer-reviewed source; the direction is corroborated by the judge-bias literature, the specific figure is not). Let a model do countable checks — sentence length, banned-phrase scans — and never let it certify voice.

Where does this usually break?

Failure	What you see	What to do
The adjective trap	Rules like "confident, human, bold", with no counter-quote	Re-run extraction against the off-voice corpus; delete any rule without a counter-quote
Voice-bible bloat	A 3,000-word guide performs worse than a 400-word one	Cut to 5-7 rules: compliance falls as constraints stack (FollowBench) and mid-document material is used least reliably (Liu et al.)
A rule that fights defaults	One rule violated in nearly every output regardless of phrasing	Stop rephrasing; move it to the editing stage (arXiv 2604.07192)
Persona theatre	"You are a world-class award-winning copywriter" in the prompt	Drop it. Across 162 personas, 4 model families and 2,410 questions, personas in system prompts "does not improve model performance", with largely random per-persona effects (Zheng et al., EMNLP 2024) — measured on factual accuracy, so the honest claim is "no evidence it helps, and it costs context"
Corpus contamination	Extraction returns the generic register you were escaping	Exclude AI-assisted, ghostwritten and committee-edited pieces (MindStudio); a corpus that already contains model output hands you the model's own defaults back, labelled as your voice
Selection bias	The corpus is "our best-performing pieces"	Select for on-voice, not for engagement — "do not just grab your most popular blog posts. Grab the posts that sound most like you" (The Brand Algorithm) — or you encode whatever the algorithm rewarded
Expecting indistinguishability	Detection stays high although the pass rate is fine	Accept the ceiling: across four domains the share of LLM text classified as human-written was "always below 55% and often less than 20%" (arXiv 2509.14543). Aim for "publishable after a light edit" — see edit-ai-draft
Testing once	A passing result quietly stops holding	Re-run after any model or tool-default change, one change at a time

If several rounds fail, the voice may be out of reach at the instruction layer. That is a legitimate finding: move the effort into editing and review rather than into a longer document.

What else do people ask?

Can you skip the document and paste writing samples every time?

For a one-off, yes. The document earns its cost when many people or many generations need the same register and you want a result someone else can reproduce. Samples are not a substitute for rules either: past two or three, extra examples changed style metrics very little (arXiv 2509.14543).

Does on-voice AI content perform better?

Unknown. No credible study links brand-voice consistency in AI content to conversion, recall or preference, and the survey figures above describe how marketers feel about the risk, not what it costs them. Treat voice as a quality standard, not a performance lever — and keep the separate question of whether the content is factually sound on its own track (spot-fabricated-stats, ai-draft-pre-publish-check).

How does this fit the rest of the workflow?

The voice document is one input to production. Scoring outputs consistently over time is a separate discipline (score-ai-output), as is the pre-publication check that catches everything voice does not cover (ai-content-pre-publish-compliance). The full path for this track is in ai-for-marketers.