# Brand voice > What a voice profile is, how it changes the output, how to have one drafted from your own website, and how protected terms keep your names intact. > Canonical HTML: https://aixlifftranslate.com/docs/voice-profiles/ > Last updated: 2026-09-08 Generic machine translation produces text that is *correct* and sounds like nobody. It flattens a friendly second-person site into corporate third-person prose, drops your contractions, and turns your product names into descriptions of your product names. A **voice profile** is the instruction set that stops that happening. It travels with every unit we translate for your project. ## What is in a profile **Brand identity.** Who you are, who you are talking to, and what you sound like. This is prose, not settings - the more specific it is, the more useful. "Casual, like a knowledgeable local friend, for travellers aged 18 to 35" beats "friendly and professional", which describes almost every company that has ever existed. **Per-language guidance.** Tone does not translate one-to-one. German has a formal and an informal register and you must pick one. Italian marketing copy tolerates warmth that would read as unprofessional in Dutch. Each target language gets its own notes. **Rules for body content and for metadata separately.** A page title has 60 characters and an SEO job; a paragraph has room to breathe. They should not be written to the same brief. **Protected terms** - your glossary. See below. ## What you write, and what we handle You write the *content* of your voice: identity, tone, rules, the things only you know. The *structure* around it - the output contract, the formatting rules, the instructions that keep markup and shortcodes intact - is ours, and it is applied consistently to every profile. That split matters: when we improve how translations are constrained, every project gets the improvement without anyone editing their profile. You will not find placeholder tokens or fencing syntax in a profile, and they are rejected if you try to add them. Write prose. ## Generating a profile from your website Writing a good voice brief from a blank page is hard. You can instead give us your website URL and pick your target languages, and get a complete draft back: brand identity, per-language guidance, and a glossary of terms found on your site. How it works: 1. We discover and read a sample of your pages. 2. Those pages are analysed to produce a voice description and a candidate glossary. 3. You review the draft and apply it, in whole or in part. It is a **starting point, not an answer.** Read it before applying it. It will usually get your tone roughly right and will sometimes propose glossary terms you do not want frozen. Editing a good draft is far quicker than writing from nothing. There is a per-day limit on generations, and you are shown a cost estimate before one runs. ## Protected terms (the glossary) A glossary term is a word or phrase that must survive translation unchanged. Your company name. Your product names. Place names you use in a particular form. Without this, "Nest Pass" becomes "Nido Pase" in Spanish, and it does so *convincingly* - the translation is grammatically perfect and completely wrong. Protected terms are handled twice: declared to the model before translation, and verified afterwards. A translation that lost a protected term does not silently ship. Two details worth knowing: - **Protection applies to visible text.** A term appearing inside a CSS class name or a URL is left alone - rewriting it there would break your page rather than protect your brand. - **Import your existing list.** If you already maintain a glossary in WPML, you can import it as a CSV rather than retyping it. ## Getting a voice right Start from real copy. Take two paragraphs from your site that sound exactly like you, and describe *why* they sound like you - that description is your profile. Then translate five representative pages and read the output in a language you actually speak. Adjust, and only then run the rest. A voice profile improved after five pages is worth far more than one theorised before the first. Changing your voice does not retroactively re-translate existing content - see [Why the second run costs less](/docs/caching-and-dedup/) for why, and what to do when you want it applied anyway.