Brand voice
The Brand Voice page is the deep voice editor for the content Zubby AI writes on your behalf. Its settings flow into recovery emails, content studio and social drafts, catalog improver output, and the widget’s product FAQ answers, so generated content reads consistently in your voice. Find it under Settings → Brand → Brand Voice, or go to /app/<storeId>/settings/brand-voice.
Where this is applied
The fields
Tone
A short free-text description of how the brand speaks - for example “Warm, expert, plain-spoken.” Keep it to a handful of adjectives; this is the headline instruction the model leans on most.
Reading grade (3–16)
A number from 3 to 16 that sets the target reading level. Lower values (e.g. 6–8) keep copy simple and broadly accessible; higher values allow more technical, dense writing. Leave it blank to let the model use its default.
Persona
A few sentences describing who the brand is speaking as - their background, expertise, and what they care about. This grounds the tone in a believable character rather than a generic assistant.
Forbidden words (one per line)
A blocklist of words and phrases the AI must avoid. Enter one per line. Use it to stop the agent making claims you can’t stand behind (“guaranteed”, “cure”), or to ban competitor names and off-brand slang.
Do examples / Don’t examples (one per line)
Two side-by-side lists of example phrasings. Do examples show the model the kind of wording you love; Don’t examples show wording to avoid. A few concrete examples in each list steer style far more reliably than abstract adjectives alone.
Channel variants
An optional block with four text areas - Chat, Email, Blog, and Ads - for per-channel overrides. Use these when a channel needs to sound different: snappier in ads, warmer in email, more conversational in chat. The Chat variant is used by widget chat and non-email messaging agents; Email is used by the email agent. Blog and Ads fields are stored by the editor, but this code does not wire those overrides into the content generators. Do not rely on an unused override to control a campaign.
Save commits the whole form
Brand Voice vs. the chat voice controls
There are two saved sources that can contribute to the same reply:
- This page (Brand Voice) - the deep editor for the store-wide profile: tone, persona, reading grade, forbidden words, do/don’t examples, and supported channel variants. Content generators use relevant fields; live chat and messaging agents also load the profile.
- Chat brand voice - the tone profile, greeting style, emoji usage, and custom-instructions knobs. They live on the AI Setup tab and in the Widget Designer’s Behavior section (both edit the same widget configuration). They add more specific conversational instructions after the store-wide profile in the live agent prompt.
Tone rules from the Rules page layer after the chat brand voice in the system prompt, so when they conflict, the rule wins. Set the durable content rules here once, then tune the conversational feel on AI Setup or in the Widget Designer as you learn what converts.
Set it up in four steps
- Write the tone and persona first - they anchor every generated draft.
- Add forbidden words for any claim you can’t stand behind. This is a prompt instruction, not a compliance guarantee or hard stream filter; pair consequential restrictions with reviewed sources and Block Claim rules.
- Paste two or three real do/don’t examples from past emails - concrete phrasing steers style better than adjectives alone.
- Click Save brand voice, then generate a recovery email or run the catalog improver on one product to confirm the output sounds right before it goes to shoppers.
Field-to-output scope
| Output | Profile use and verification |
|---|---|
| Widget and messaging replies | Store-wide voice, examples, reading level and forbidden words, plus Chat or Email override. Check a fresh conversation after saving and resolve conflicting AI Setup or Rules instructions. |
| Recovery emails | Base tone/persona, examples, reading level and forbidden words. Inspect the generated copy; the Email override is not loaded by this generator. |
| Catalog improver and product Q&A | Relevant base voice fields guide drafts. Verify the actual output and factual source before applying or publishing. |
| Content Studio and social drafts | Base profile guides supported generation when available. Stored Blog/Ads overrides do not establish per-output enforcement. |
Worked example and troubleshooting
Illustrative outdoor shop: use “plain-spoken and practical,” ask for short sentences, and forbid “guaranteed waterproof” unless a reviewed product specification supports the claim. Save the full form, open a fresh chat, and generate a reviewed product draft. Check tone, unsupported claims and missing facts separately. A pleasant tone does not establish accuracy.
If a chat seems unchanged, inspect its more specific AI Setup voice and tone rules. If only an Ads override changed, test the base voice instead: that override has no connected content consumer here. Save reusable corrections explicitly; editing one reply is not an immediate global voice update. Team members need settings access to edit this profile.
Tips for a strong brand voice
- Be specific in the persona. “A 15-year baker who explains things simply” gives more direction than “friendly and helpful.” Use a role appropriate to your business; do not invent credentials, experience, or a real person.
- Treat forbidden words as a safety rail for compliance-sensitive categories (supplements, finance, beauty) - pair it with your system prompt rules.
- Add do/don’t examples whenever you spot an off-brand recovery email or generated draft; it’s the fastest way to correct style drift.
Related
Tune everyday tone and emoji on AI setup & behavior. Combine voice with hard rules in System prompt, and use flagged replies in Training to keep refining your examples.