Extract a Brand Voice So AI Writes Like You

To extract a brand voice guide that actually makes AI write like you, stop describing your voice and start measuring it. Adjectives — "professional yet friendly," "bold but approachable" — describe nearly every brand and constrain none of them. Your published corpus, by contrast, contains falsifiable patterns: how long your sentences run, which words you reach for and which you ban, how often you hedge a claim versus assert it. Those patterns can be written down as a spec a model can follow and a reviewer can check.

The stakes are specific to the AI era. When generation is cheap, the constraint on content programs shifts from production capacity to consistency — and an underspecified voice degrades quietly, one plausible-but-generic draft at a time. Peer-reviewed work published in September 2025 found that even frontier models given real examples of an author's writing still measurably fail to reproduce informal personal style. The fix is not a better adjective list. It is a better specification.

This guide covers why adjective-based voice documents fail, what the research actually shows about example-driven style control, which stylometric measures are worth computing and which are numerology, a step-by-step extraction method demonstrated on our own public blog, a seven-row voice-spec template, and the blind discrimination test that serves as the only honest acceptance criterion.

The strongest indictment of adjective-based voice documents comes from an organization that publishes more copy than almost anyone: the UK government. GOV.UK's content-design guidance deliberately keeps its prose emotionless because, in its own words, "Adjectives can be subjective and make the text sound more emotive and like spin." Its tone guidance still names qualities — specific, informative, clear and concise, brisk but not terse, incisive, human rather than robotic, serious but not pompous — but every quality sits on top of mechanical rules: write as if talking one-to-one, prefer active voice, address the reader as "you," avoid "please note," never set large text blocks in capitals.

Mailchimp's content style guide — probably the most cited voice document in marketing — makes the same move. It first separates the two concepts most brand documents blur: "You have the same voice all the time, but your tone changes." Voice is the constant; tone flexes by context. It then operationalizes voice into four named values (Plainspoken, Genuine, Translator, Dry Humor), each with explicit rules underneath — active voice, plain English, positive framing — and ranks the priorities outright: it is always more important to be clear than entertaining.

Tone is also measurable from the reader's side. Nielsen Norman Group's tone-of-voice research — published in 2016 and reviewed as recently as January 2024 — scores copy on four dimensions, each a three-point scale with a neutral midpoint: funny versus serious, formal versus casual, respectful versus irreverent, enthusiastic versus matter-of-fact. In its quantitative phase — an online survey of more than 100 US adults rating organizations on five-point scales — trustworthiness, not friendliness, dominated: 52% of the variance in desirability ratings (willingness to recommend) was explained by trustworthiness alone, with friendliness adding a much smaller contribution of roughly eight percentage points on top.

Notice what the credible guides have in common. None of them stops at adjectives. All of them converge on a three-layer structure: a small set of named values, explicit do/don't rules under each value, and worked before/after examples. That is exactly the structure a corpus-derived voice guide should output — the difference is that extraction derives the values and rules from measured text rather than asserting them from taste, which is what the rest of this guide is about.

If adjectives underspecify, the obvious next move is examples — show the model your writing instead of describing it. Vendor guidance broadly supports this. Anthropic's prompt-engineering documentation recommends providing three to five relevant, diverse, well-labeled examples as one of the most reliable levers for steering format, structure, and tone — the underlying logic, paraphrasing its multishot-prompting guidance, being that examples teach the standard you expect better than written instruction alone. OpenAI's guidance takes a slightly different default — start zero-shot, add examples when zero-shot underperforms — but the two converge on the core claim: examples help tone control.

The important question is where example-driven imitation stops working — and here the peer-reviewed evidence is unusually clear. A September 2025 study from researchers at Stony Brook University, Penn State, and the Bosch Center for AI, "Catch Me If You Can? Not Yet" (arXiv:2509.14543, EMNLP 2025 Findings), evaluated six frontier LLMs imitating more than 400 authors across four datasets — Enron emails, blogs, Reuters news, and Reddit forums — using four independent measures: authorship attribution, authorship verification, stylistic distance, and AI-detection.

The domain split is the finding that matters for brands. Where text follows a template — news copy, formal email — models imitate style convincingly. Where voice is genuinely personal and informal, imitation collapses. The study also ran human-likeness detection and found generated text remained frequently flaggable as AI-written even with in-context examples of the target author. The authors' conclusion is the load-bearing sentence for this entire method:

Read the two halves of the evidence together and the practical conclusion writes itself. Exemplars beat adjectives — that much vendor guidance and the structured-format numbers agree on. But exemplars alone leave the informal, implicit layer of voice on the table. That layer — sentence rhythm, lexicon habits, certainty posture — is exactly what an extraction step converts from implicit to explicit. A written spec distilled from your corpus, stacked on top of few-shot examples, is doing real work that neither component does alone.

Stylometry — the statistical analysis of writing style — offers a menu of measures, and an honest voice guide has to admit that they are not equally trustworthy at marketing-corpus sizes. A brand blog is a few dozen to a few hundred posts, not the thousands-of-tokens-per-author corpora academic stylometry was built on. Some measures survive that downsizing; some quietly turn into decoration.

The canonical cautionary tale is the type-token ratio — unique words divided by total words, the most commonly cited lexical-diversity measure. As Sketch Engine's glossary entry notes, TTR drops dramatically as sample length increases, because vocabulary growth is sublinear to token count. A 500-word post and a 3,000-word guide will produce incomparable TTRs from the same author on the same week. The standard fixes are windowed or mean TTR — compute the ratio over fixed-size chunks and average — or a switch to MTLD (Measure of Textual Lexical Diversity), which is designed to stay stable regardless of text length and is the more defensible single number to track over time.

Readability formulas earn a similar asterisk. Flesch-Kincaid and its relatives measure exactly two things — sentence length and syllable count — as proxies for difficulty. They measure neither tone nor audience fit nor actual comprehension, and because the two inputs collapse into one number, a low score cannot tell you which lever to pull: shorter sentences, or simpler words? Useful as a diagnostic to notice drift; useless as a target to optimize.

The hedge-to-booster ratio deserves a note, because it is both the most brand-relevant measure on the list and the easiest to misuse. The framework comes from Ken Hyland's metadiscourse research on academic writing: hedges mitigate a claim (might, suggest, possibly, tends to); boosters assert certainty (clearly, undoubtedly, always, never). In the academic-writing corpora that line of research analyzed, hedges were consistently the most frequent metadiscourse category — representative figures from one discipline-specific corpus study put hedges at roughly 23% of metadiscourse markers versus roughly 13.5% for boosters, about 1.7 times as many. Treat those numbers strictly as an illustration that even confident expert writing hedges more than it boosts. They describe academic prose, not marketing copy, and they are emphatically not a ratio to tune your brand copy toward. What you can borrow is the mechanism: counting hedges and boosters per 100 words in your own corpus gives you a reproducible confidence-posture score that both a writer and an AI checker can be held to.

The method has four steps: assemble the corpus, run the computable measures, distill rules, and pressure-test. The first step decides whether the rest means anything, because extraction measures whatever you feed it. Feed it everything you have ever published and you extract your average; feed it your edited best and you extract your standard.

For a worked example we ran the method on the most convenient corpus we have: Digital Applied's own public blog — our content, no confidentiality questions, and roughly 1,900 published posts in one format. We did not feed all of them in. Applying our own inclusion rules above, we measured only the editorially reviewed subset — pieces that went through full human review before publication — and excluded anything that had not, which is the same filter we would ask a client to apply to their own archive. Step two is mechanical: compute sentence-length distribution per post, paragraph rhythm, a length-normalized lexical-diversity number, and hedge and booster counts per 100 words with a word list tuned to marketing prose. Any scripting language handles this; none of it needs specialist tooling.

Step three is where measurement becomes specification, and it is the step most teams skip. The useful output was not a sheet of statistics — it was the rules the statistics forced us to make explicit. Patterns we had never written down anywhere surfaced immediately: ledes that open with the sharpest concrete fact rather than a scene-setting question; short declarative sentences deployed at turns in the argument; a hedge-forward posture on any vendor-published number paired with unhedged statements about our own workflow; an out-list of constructions that simply never appear in our edited work. None of that lives in an adjective. All of it is checkable against text — which is what makes it a spec. It is the same move as extracting design tokens from a mature interface instead of describing the aesthetic in adjectives — a method we cover for the visual side in extracting a design system for AI app builders.

The deliverable of extraction is deliberately small: a seven-row worksheet that fits in a system prompt and survives contact with a busy editorial team. Each row pairs a measurement with the rule it produced. The example rules below show the shape an on-voice rule takes — yours come from your corpus, not from this table.

Structurally this is the same three-layer document GOV.UK and Mailchimp converge on — values, rules, examples — with one difference that changes everything downstream: every rule carries its measurement with it. When a stakeholder asks why the guide bans a construction, the answer is not "taste" — it is "zero occurrences across the corpus we extracted from." That makes the guide defensible in review, checkable in production, and re-derivable when the voice legitimately evolves.

A voice guide that lives in a PDF is decoration. Deployed, it becomes three layers of the same prompt stack — and the evidence from Section 02 explains why all three are necessary rather than redundant. The exemplars carry surface style, which is what in-context examples demonstrably do well on formatted content. The written spec carries the implicit layer — rhythm, lexicon, posture — which is what examples alone measurably fail to transfer on informal voice. The checks close the loop.

This is also where voice work joins a larger pattern we keep arriving at across disciplines: output quality tracks specification quality more closely than it tracks model choice — the argument we make in full in our brief-method guide to specification quality. A voice spec is simply the specification layer for tone. And like any prompt asset, it decays without review: version it, date it, and audit it on a schedule the same way we recommend auditing a prompt library. When the spec changes, the back catalogue needs triage too, which is the same call we work through in prioritizing which pages to refresh first. If you want template-level starting points for the exemplar layer, our marketing copy prompt templates pair naturally with an extracted spec. Teams that want the whole pipeline built — extraction, spec, exemplar set, and the production gates — can see how we assemble it inside our content engine service.

Every measure in Section 03 is diagnostic. None of them is proof. The only honest acceptance test for a voice guide is blind discrimination: mix AI-generated pieces built from your extracted guide into a batch of your real published copy, and see whether reviewers who know your brand can sort them. The best available template for that test comes from the Turing-test literature — and its most useful lesson is about variance, not pass rates.

In a controlled, preregistered interactive Turing test by Jones and Bergen (arXiv:2405.08007, 2024), GPT-4 was judged human in 54% of five-minute text conversations, against 22% for the 1966 chatbot ELIZA and 67% for actual human participants. The expanded, peer-reviewed version in the ACM FAccT 2025 proceedings — Jones, Rathi, Taylor and Bergen — adds a second study run on an undergraduate population using GPT-4o, in which the model was judged human 77% of the time, above that study's own human baseline of 71%. Neither the 77% nor the 71% appears in the preprint. Same test design and the same prompt, different population and model: the result moved by more than 20 points.

Never quote one of those numbers alone as "the" pass rate — the spread is the finding. Discrimination results are exquisitely sensitive to who is judging, what they were primed to look for, and which model version generated the text. That sensitivity cuts both ways for your own testing: a single flattering result from three colleagues on a Friday afternoon proves nothing, and the posture the data recommends — state the range, name the conditions — is precisely the hedge-forward posture this guide argues your copy should carry too.

For the protocol itself, borrow the four-part evaluation design from the style-imitation study: mix real and generated pieces; have reviewers attempt attribution (which pieces are ours?) and verification (is this specific piece ours?); compute stylistic distance between the generated set and the corpus using the Section 03 measures; and run an AI-detector pass as a fourth, independent signal. If reviewers hover near chance and the stylistic-distance gap is small, the guide is doing its job. If they sort the piles comfortably, interrogate the misses — the failures name the next rule your spec is missing, and the cycle repeats until the piles blur.