Pasting your style guide into the prompt is the most popular fix for generic AI drafts, and it is the fix least likely to work. The draft still reads like a press release with the logo removed, because the model was handed adjectives when it needed decisions: which claims you make and what proof sits behind each one. Give a model tone words and it returns the internet average; give it your claims with the proof attached, and it starts to sound like a company that exists.
The stakes are commercial. G2's buyer research found AI chatbots and review sites are now the two biggest influences on B2B shortlists, with 51% of software buyers starting research with an AI chatbot more often than with Google and one in three saying an AI chatbot led them to choose a vendor they hadn't considered. There, the description of your company is assembled from whatever the engine retrieves, and company websites, yours and your competitors', are the largest single source of AI citations. If it could belong to anyone, the engine has no reason to attach it to you.
The short answer has three parts. Replace tone adjectives with a written record of claims, proof and positions. Keep that record persistent, so every draft starts from it. Put a named reviewer between the draft and the publish button, with the approval visible in the record.
Why does AI-drafted content sound generic by default?
Because a language model's default output is the most probable continuation of what it has read, and what it has read is the public internet. OpenAI lists three primary source classes for its models, starting with information publicly available on the internet, and that internet is mostly marketing copy that already sounded like everyone else. Generic is the arithmetic, and no prompt instruction changes it. It also compounds: the model's own output joins its context as it writes, so the back half drifts even when the opening holds.
The convergence shows up in controlled studies of people. A field experiment with 758 BCG consultants found AI-assisted work was rated more than 40% higher in quality on tasks inside the model's competence, and materially less likely to be correct on tasks outside it. It also found ideas produced with AI were more similar to one another. A 2023 experiment with 453 professionals using ChatGPT found writing tasks took 40% less time, quality rose 18%, and the gap between the weakest and strongest writers narrowed.
The same averaging pulls copy toward sales language, and AI engines do not reward it. Scarcity and exclusivity framing measurably reduces how often an LLM recommends a product, in a test built on three sets of 10 fictitious products so brand memory could not contaminate it. The words a model reaches for when it has nothing specific to say are the words that make other models trust you less.
Why doesn't pasting the style guide into the prompt fix it?
Because a style guide describes tone, and tone was never the missing ingredient. "Confident but approachable" has a thousand valid renderings, and the model picks the most common one, the corporate default you were trying to escape. A guide that says what you sound like, without saying what you claim and what you refuse to say, leaves every consequential decision to the average.
Long pasted guides also fail mechanically. Liu et al. (TACL 2024) showed that models use information at the start and end of a long context far better than information buried in the middle, so a 2,000-word guide wedged between the instructions and the source material is partly ignored by design.
The re-prompting tax finishes the job. A prompt lives in one person's session, so every colleague rebuilds the context from memory. A 2025 NBER working paper on 1.1 million sampled ChatGPT conversations found roughly two-thirds of writing requests ask the model to modify text the user supplied rather than write from nothing. People already know the model needs their material; the fix is to make it persistent instead of pasted.
What does a brand voice consist of, if not adjectives?
Decisions a competitor could not copy without becoming you: the claims you are willing to make, the proof behind each one, the positions you hold, and the words you refuse. Everything else is formatting. Run the test on your last ten pieces: could a direct competitor publish this unchanged? Where the answer is yes, the piece contains no decisions, and tone editing will not rescue it.
Evidence is the most portable form of voice, and the engines reward it. The GEO study (Aggarwal et al., KDD 2024) measured up to 40% higher visibility for adding statistics, quotations and citations across a 10,000-query benchmark, where visibility means position-adjusted word count inside the answer, not a citation and not a recommendation; keyword stuffing did not help. Profound's citation analysis puts 57% of AI citations on company-operated sites, yours and your competitors', so company pages are what a buyer most often hears back. A number with a source travels; an adjective stays home.
Do not confuse this with phrasing tricks. C-SEO Bench (NeurIPS 2025) tested conversational-SEO rewrite methods across two tasks and six domains and found most of them largely ineffective and often harmful to citation rank, with gains shrinking as more sites adopted them. Rewriting for the engine is mostly noise, and the rewrites tested included added statistics, citations and quotations. Evidence the reader can check is still the better bet, though its citation lift is unproven.
Claims carry liability, which is why the record attaches proof. The FTC's advertising guidance requires objective claims to be substantiated, and its comparative advertising policy, in force since 1979, holds comparisons to the same standard as any other claim. A model will invent a comparison in a heartbeat, and a voice built from substantiated claims cannot be quietly wrong.
How should a human review AI drafts without rewriting everything?
In three short passes, with the reviewer's decisions written down so they stop recurring. Pass one is the voice check: read the opening and the closing section aloud and ask whether a competitor could run either unchanged, because the opening is where the prompt still holds and the back half is where it has let go. Pass two is the claims check: every number and comparison either has proof in the record or comes out. Pass three is the strategic edit, where the person adds first-hand detail the model could not know, and copyediting goes last.
Order matters: a voice failure means the draft came from the wrong context, and line edits teach the system nothing. Send it back with the missing decision instead. Each recurring edit becomes a rule, and edit time per piece tells you whether the voice is holding. Microsoft's 2025 Work Trend Index, drawn from 31,000 workers across 31 countries, describes the emerging job as directing agents rather than producing first drafts.
Approval has to be visible, not implied. A cleared draft should carry that the way a contract carries a signature: who reviewed it, and against which version of the record. NIST's AI Risk Management Framework makes "operate and monitor" a lifecycle stage in its own right, and the W3C's provenance model reduces to three things worth recording: the entity, the activity that produced it, and the agent responsible. A workflow that cannot answer those questions for every page is under human proximity, and proximity is not review.
How do you know the voice is holding once volume goes up?
You measure it inside the workflow and inside the answers your buyers see. Inside the workflow, track edit time per piece and drafts sent back at the voice check. If they are not falling month over month, the review is not feeding the record, and you have a manual process wearing an automation costume.
Outside the workflow, the test is portrayal. The IAB's 2026 measurement guidance sorts AI visibility into four groups for brands, and Portrayal, how a brand is described when it does appear, sits alongside Presence, Prominence and Persuasion. Ask the buyer questions your content answers and read the description that comes back. If the engine describes you in your claims and your terms, the voice survived the trip; if it describes you in category boilerplate, it never left the page.
Read portrayal as a base rate, never as a single answer. SparkToro and Gumshoe found AI engines are highly inconsistent when recommending brands, with recommendation lists that rarely repeat exactly in the same study, and a 2026 variance-components study found run-to-run noise large enough to swamp real differences in small samples. Repeat the questions across days and treat a shift in the pattern as evidence. One flattering answer is a sample, and work on quantifying uncertainty in AI visibility puts confidence intervals on it.
How does Trovance keep brand voice in AI drafts under human review?
Trovance starts from the record described above and calls it your Brand Core: the claims you are entitled to make, the proof attached to each one, the positions and vocabulary that are yours, and what you do not say. Drafts are produced from that record, so each starts from your approved claims instead of the internet average, and when the Brand Core changes, later drafts change with it.
Human review is a stage in the workflow. A person reviews and approves every draft before it publishes, and that approval is visible in the record. Recommended actions arrive with the reasoning attached: the tracked buyer question the piece answers and the answer evidence behind it. The reviewer's edits are how the record learns, and nothing publishes without them.
The loop closes on portrayal. Trovance observes the answers AI engines give to your tracked questions, preserves each answer snapshot with its sources, and compares how you are described over time. After a piece ships, the next analysis cycle reruns the same questions and shows you how the description changed. Reading that change against your Brand Core claim by claim is something Trovance is building toward.
What Trovance will not promise is that a model will cite you or describe you in your own claims on any schedule. Engines are probabilistic, and which sources they reach for on a given run is not something any vendor sets. It also will not remove the reviewer: the system produces drafts, a person decides what ships. What it can do is start every draft from your decisions and show you whether they reached the buyer.
What should you do this week?
Run the competitor test on your last ten published pieces and count the failures; that is your baseline. Then write the first version of the claims record: ten claims you make, the proof behind each, five positions a competitor would not take, and the words you refuse. Keep it short: a page of decisions beats twenty pages of tone.
Next, apply the three-pass review to the next draft and log every edit; anything you fix twice becomes a line in the record. Then ask three buyer questions to two AI engines on three days and write down how you are described. That is your portrayal baseline, the reading this work should move.
Be honest about the timeline. The workflow metrics are the ones you can actually move, because edit time and send-backs are yours to measure; portrayal moves on the engines' schedule, and every reading carries the variance documented above. Anyone promising a model will speak in your voice by a given date is selling against the evidence.
If you want the claims record, the review and the portrayal readings in one place, start a free Trovance trial and see how AI engines describe you today.
Keep the reviewer in the loop
Human-in-the-loop content review - where the three passes plug in.
Human oversight in an AI content engine - where approval sits in the engine.
Marketing agent autonomy is not the goal - why removing the reviewer is wrong.
Taste when AI does the execution - the judgment the third pass needs.
Write with proof instead of adjectives
AI agents need evidence, not more content - why checkable claims travel.
A relevant page can still miss the proof - the failure a claims record prevents.
Social proof is an attack surface, not a GEO tactic - what persuasion does to a recommender.
The most-quoted study in AI search, quoted correctly - what the GEO benchmark tested.
