Motionhooks
How to Train an AI Brand Voice From a Client’s Website, Reviews, and Past Captions

How to Train AI Brand Voice From Client Content

Written by 9 min read
Agency strategist mapping client website, reviews, and captions into an AI brand voice profile

A client sends you three job site photos and asks for captions “that sound like us.” If you have to rediscover their tone every week, AI becomes another approval bottleneck instead of a faster publishing system.

The practical answer is to train AI brand voice from client content once, then reuse that voice across future captions, rewrites, short-form videos, and scheduled posts. The best source material is usually already available: the client’s website, customer reviews, past captions, and project notes.

How to train AI brand voice from client content without starting from a blank prompt

A reusable AI brand voice is not just “friendly and professional.” That describes almost every small business and gives the model too little to work with.

You want a compact voice profile that tells the AI what the client sells, how they describe value, what customers praise, which words they repeat, and what they avoid. For trades and local service businesses, this matters because a painter, roofer, electrician, and renovation contractor can all post similar project photos, but their tone should not feel interchangeable.

If you manage multiple clients, build this profile before you scale scheduling. Motionhooks is built for that workflow: agencies can create AI-generated videos, captions, approvals, scheduling, and analytics from one workspace, with brand voice memory for reusable caption generation. You can see the agency workflow on Motionhooks for agencies.

Use 5 source types, but weight them differently

Not every piece of client content deserves the same influence. A homepage usually tells you positioning, while reviews tell you what customers actually believe after the job is done.

Source materialWhat to extractHow much to trust it
Website service pagesOffers, locations, credentials, process, guarantees, calls to actionHigh for facts, medium for tone
Customer reviewsTrust signals, objections solved, plain-language benefits, emotional payoffHigh for proof and customer language
Past captionsSentence length, emoji use, recurring phrases, CTA style, platform habitsHigh if posts performed or were approved quickly
Project briefs and job notesBefore-and-after details, materials, timelines, crew process, client concernsHigh for specificity
Existing videos or reelsPacing, hooks, spoken phrases, text overlay style, visual emphasisMedium to high if they match the desired future style

Website pages give the AI facts it must not guess

Start with the homepage, core service pages, about page, and contact page. Pull the exact services, service area, typical customer, certifications, warranty language, and preferred calls to action.

For a kitchen installer, that might include “custom cabinetry,” “measure and fit service,” “serving Leeds and surrounding areas,” and “book a design consultation.” These facts stop AI captions from drifting into generic claims like “transform your home today” with no useful detail.

If the client uses Motionhooks, the same facts can inform captions and short-form video briefs generated from a single photo and brief. The process is shown on how Motionhooks turns a brief into posts and videos.

Reviews reveal what customers trust enough to repeat

Collect 10 to 30 recent reviews from Google Business Profile, Facebook, Trustpilot, or trade directories. Do not just paste them in raw; extract patterns.

Look for repeated phrases such as “turned up when they said they would,” “left the place spotless,” “explained every step,” or “finished before the weekend.” These are stronger than agency-written claims because they reflect buyer priorities.

For example, if 12 out of 25 reviews mention tidiness, the voice profile should include “clean, respectful, tidy finish” as a recurring proof theme. Future AI captions can then say, “A sharper finish without leaving the house looking like a worksite,” which sounds more credible than “quality workmanship.”

Past captions show what the client has already approved

Use 20 to 50 historical captions if the client has them. Sort them into three groups: strong, acceptable, and off-brand.

Strong captions are the fastest training material because they show rhythm and taste. Off-brand captions are also useful because they help you define what the AI should avoid, such as heavy sales language, overuse of emojis, too many hashtags, or exaggerated promises.

If you publish across Instagram, TikTok, YouTube, Facebook, LinkedIn, X, Threads, Bluesky, and Pinterest, consistency becomes harder as formats change. This is where a scheduler helps; this guide on scheduling posts across platforms without losing brand voice explains the platform side of that problem.

Write posts like this in Motionhooks

Brand voice + AI. Your tone, not the internet’s average.

Try free

Project briefs add the detail that makes AI captions feel current

A brand voice profile should not replace a project brief. The voice tells the AI how to speak; the brief tells it what happened on this job.

For a roofing post, the brief might say: “Replaced cracked ridge tiles, repaired lead flashing, completed in one day, customer had a leak above bedroom.” With that input, the AI can write a caption that mentions the real repair instead of producing a vague “another great roofing project completed.”

If your team struggles to collect the right job details, use a short repeatable prompt. This related guide shows how to write an AI video brief for a renovation social post in under 5 minutes.

Build a reusable voice profile in 30 minutes

You do not need a 12-page brand document. A useful AI voice profile for captions and short-form video can fit on one page if it is specific.

  1. Collect the source material: Save the homepage, two service pages, 10 to 30 reviews, 20 to 50 past captions, and 3 to 5 recent project briefs.
  2. Extract factual rules: List services, locations, target customers, claims that are allowed, and claims that need approval.
  3. Extract voice traits: Choose 3 to 5 traits backed by evidence, such as “practical,” “reassuring,” “straight-talking,” or “detail-led.”
  4. Write phrase banks: Add approved phrases, customer language from reviews, preferred CTAs, and banned phrases.
  5. Create platform rules: Define caption length, hashtag style, emoji use, and CTA intensity for each platform.
  6. Test with 5 real posts: Generate captions from recent photos or briefs, then mark what sounds right and what needs tightening.
  7. Save the final version: Store it as the client’s reusable persona so future posts start from the same voice.

Motionhooks uses GPT-4o for caption generation, rewrites, brand voice training, and content suggestions. Its brand voice memory means a trained persona can be reused across future captions instead of rebuilt for each post.

A simple brand voice card agencies can reuse for every trade client

Use this structure when you turn raw client material into a usable AI voice. Keep it short enough that an account manager can review it quickly.

Client: Northside Painting Co.Audience: Homeowners planning interior repainting, landlords preparing rentals, small commercial property owners.Voice: Calm, tidy, practical, reassuring. Short sentences. No hype.Proof themes: Clean edges, protected floors, punctual arrival, clear quotes, rooms usable quickly.Approved phrases: “sharp finish,” “proper prep,” “kept disruption low,” “ready for the next tenant.”Avoid: Luxury language, hard-sell CTAs, more than two emojis, claims about being the cheapest.CTA style: “Send us a photo for a quick estimate” or “Message us to check available dates.”

That card gives the AI enough direction to write a caption for a bedroom repaint, a rental refresh, or a shopfront touch-up without changing personality each time.

Test the first 20 outputs before you publish at scale

The first test should use real posts, not invented examples. Pick five recent jobs and ask the AI for four variations per job: Instagram, Facebook, LinkedIn, and TikTok text.

Score each output from 1 to 5 on four criteria: factual accuracy, client tone, usefulness, and platform fit. If a caption scores below 4 on tone, add the reason to the voice profile rather than correcting only that one caption.

Here is a realistic agency example. Suppose you manage six trade clients, each posting four times per week across three platforms. If brand voice setup reduces rewriting by 7 minutes per post, that saves 168 minutes per week: 6 clients × 4 posts × 7 minutes.

At an internal blended cost of £45 per hour, that is about £126 saved each week, or roughly £6,552 per year before counting fewer approval rounds. The larger win is consistency: your junior coordinator, account lead, and AI assistant all work from the same voice rules.

Connect the voice profile to AI videos, captions, approvals, and scheduling

A brand voice profile becomes more valuable when it travels through the whole publishing workflow. If it only sits in a document, people will forget to use it when deadlines get tight.

Motionhooks turns a single photo and brief into platform-ready short-form videos, captions, and scheduled posts. It uses HeyGen for short-form AI video rendering and supports approvals, cross-platform scheduling, and per-post analytics from one workspace.

For trade clients, the source photo still matters. A strong voice cannot rescue a confusing image, so use clear before-and-after shots, uncluttered job site photos, and visible workmanship details. This checklist on job site photo mistakes that weaken AI videos is useful before you train the caption style.

If you serve multilingual clients, store language rules too. Motionhooks supports 8 languages, so a reusable persona can include whether the client wants direct translation, local phrasing, or separate caption styles by market.

Protect client data while training the voice

Only use content you have permission to use. Public website copy, public reviews, approved captions, and client-supplied briefs are usually appropriate, but remove private customer details that are not needed for tone.

Do not include phone numbers, addresses, private complaints, access codes, or full customer names unless the client has explicitly approved that use. You can preserve the useful pattern by rewriting “Mrs. Taylor at 14 Green Lane was worried about a leak above the nursery” as “a homeowner was worried about a leak above a bedroom.”

If the client is cautious about access and accounts, ask how the platform handles security before connecting social channels. Motionhooks includes AES-256-GCM encrypted OAuth tokens, optional TOTP two-factor authentication, and data export or deletion.

Decide what goes into the handoff when the client approves the voice

Approval should not be a vague “looks good.” Ask the client to approve a specific voice card and a small batch of example captions.

Your handoff should include the final voice profile, approved phrases, banned phrases, CTA rules, platform rules, and three caption examples the client would be happy to publish. If you use Motionhooks, store the persona in brand voice memory so future rewrites and content suggestions can use it directly.

For agencies testing the workflow, the free tier includes 50 AI messages, 5 posts, and 3 images per month with no credit card required. You can review the current plans on Motionhooks pricing.

The payoff is simple: better first drafts, fewer approval loops, and a client voice that stays consistent when you publish more often across more platforms.

Frequently asked questions

What is the best source material for training an AI brand voice?
The strongest mix is the client’s website, customer reviews, approved past captions, and recent project briefs. Website copy gives factual boundaries, reviews reveal trust signals, and captions show what the client has already accepted as their tone.
How many past captions do I need to train an AI brand voice?
Use 20 to 50 captions if the client has posted consistently. If they have fewer, combine the available captions with 10 to 30 reviews and two or three service pages so the AI has enough language patterns to learn from.
Can agencies reuse one AI brand voice across multiple platforms?
Yes, but the profile should include platform rules. The core voice can stay consistent while caption length, CTA style, hashtags, and opening hooks change for Instagram, TikTok, LinkedIn, Facebook, and other platforms.
How do I know if an AI-generated caption matches the client’s voice?
Score it against factual accuracy, tone, usefulness, and platform fit. If a draft misses the mark, update the reusable voice profile with the reason instead of only fixing that one caption.
Should customer reviews be copied directly into AI prompts?
Extract patterns rather than relying on raw copying. Use repeated themes like punctuality, tidiness, or clear communication, and remove private customer details that are not needed to define the brand voice.

More from How to Train an AI Brand Voice From a Client’s Website, Reviews, and Past Captions