Why generic AI content doesn't convert and how to train AI on your actual voice

Series: AI for the Home Services Owner · Post 2 of 6

Most AI-generated content reads like every other AI-generated content. That's why it doesn't work. Here's the specific process for teaching AI tools to write in your voice so the output actually sounds like your business.

By Ben Blair · 8 min read

You've seen the content. "In today's fast-paced world, HVAC systems play a crucial role in ensuring your home's comfort. At [Company Name], we are committed to providing exceptional service." Every home services site has this page. It was written by ChatGPT with a three-sentence prompt. It reads like it. Google knows. Your customers know. The content does nothing, and the owner concludes that "AI content doesn't work."

The problem isn't AI. The problem is that the owner treated a tool that can sound like anyone as if it could only sound like one generic thing. Here's how to actually get AI to write in your voice — and why it's worth the hour it takes to set up.

What "your voice" actually means

Voice is a vague word. Most people can tell when content is "in their voice" and when it isn't, but they can't tell you why. Here's the breakdown.

Voice is made up of specific, observable patterns:

1. Sentence length and rhythm. Do you write short, punchy sentences? Long, winding ones? A mix? Where do the pauses go?
2. Vocabulary choices. Do you say "furnace" or "heating system"? "Fixed" or "resolved"? "I think" or "in our opinion"? Every word signals brand.
3. Opinion density. Do you state what you think, or do you hedge? Most home services content hedges too much. Owners who are actually experts have opinions.
4. Stories and specifics. Do you reference actual jobs, neighborhoods, customer names? Generic content doesn't. Your voice does.
5. What you refuse to say. Voice is partly what you don't include. "Best-in-class service." "Seamless experience." "Leveraging our expertise." The words you avoid define your voice as much as the ones you use.

Generic AI content fails on every one of these. Uniform sentence length. Corporate vocabulary. Zero opinions. No specifics. And it uses every phrase on the "words you refuse to say" list.

The three-part voice training system

Teaching AI your voice takes about two hours of setup and maybe 15 minutes a month to maintain. After that, every piece of content you generate sounds like you.

Part 1: Build a voice corpus

A voice corpus is a collection of writing samples that demonstrate your voice. You need 3,000–5,000 words. This is less than it sounds. Sources:

• Emails you've written to customers (the ones you're proud of)
• Social media posts you wrote yourself
• Estimates or proposal language you've hand-written
• Text message follow-ups to customers
• A 30-minute recording of you talking about your business, transcribed

Paste it all into a single document. Call it "voice_samples.md" or whatever works. This is your reference material going forward.

If you don't have enough written material, spend 30 minutes with a transcription app (Otter, Rev, even your phone's voice memo) and answer these questions out loud:

• How do I describe what my company does to a neighbor at a cookout?
• What's the most common question customers ask, and how do I answer it?
• What do I say when a customer asks "why are you more expensive than the other guy?"
• Tell the story of the hardest job we ever did and how we handled it.
• What would I want a potential customer to know about us that we never say in marketing?

Transcribe the recording. Now you have 3,000 words of you, talking like you. Edit lightly for punctuation and coherence but don't over-polish — the rough edges are the voice.

Part 2: Write the style guide

A style guide for AI is a one-page document that tells the tool exactly what to do and what not to do. Here's the structure:

About the business. Two sentences. Who you serve, what you do, where you work. Example: "We're a family-owned plumbing company serving Denver and the surrounding suburbs. We do residential service calls, drain cleaning, and water heater installations — no remodels, no commercial."

Voice characteristics. Five to ten bullets describing how you write. Examples:

• Short, direct sentences. Cut anything unnecessary.
• Use first person ("I think," "we handle") rather than third person ("our company believes").
• State opinions when they're grounded in experience. Don't hedge.
• Use specific numbers and examples, not generalities.
• Warm but professional. Conversational, not corporate.

Words and phrases to always use. Your actual vocabulary. "Water heater" not "water heating appliance." "Fix" not "remediate." "Customer" not "valued client."

Words and phrases to never use. Your kill list. Examples for most home services:

• "Leverage" (in any form)
• "Seamless"
• "Best-in-class"
• "In today's fast-paced world"
• "Valued customer"
• "Committed to excellence"
• "At [Company], we pride ourselves on..."
• "Robust solutions"
• "Turnkey"
• "Synergy"

Topics you have opinions on. List the hills you die on professionally. Example: "Tankless water heaters aren't right for most homes under 3,000 square feet, despite what the industry says. We'll explain why in plain terms."

Examples of good and bad. Paste two paragraphs — one that sounds like you, one that doesn't. This single move does more work than any other instruction.

Part 3: The prompt pattern

Every time you use AI to draft content, your prompt should include the style guide and a few voice samples. It looks like this:

"I need you to write a blog post for my plumbing business. Below is my style guide and voice samples. After you read them, write a 1,000-word post on [topic]. Follow the style guide exactly. Match the rhythm and vocabulary of the voice samples. Do not use any phrase from my kill list. Include at least one specific example or number. End with a natural next step, not a generic 'contact us today.'

[Paste style guide]

[Paste 500–1,000 words of voice samples]

Topic: How to tell if you need to replace your water heater."

That's the full prompt. Every time. The AI gets re-trained on your voice at the start of every conversation, because models don't remember between sessions (unless you set up a custom GPT or equivalent, which is the next step).

Upgrade: custom GPTs and projects

ChatGPT Plus ($20/month) lets you create a "custom GPT" that has your style guide and voice samples permanently attached. Claude's "Projects" feature (included in Claude Pro, $20/month) does the same thing. You set it up once.

After that: you open the custom assistant, type "write me a blog post about X" in one sentence, and the output is already in your voice. No re-pasting the style guide. No re-attaching samples. It just works.

This is where the setup pays off. Two hours of prep, then every piece of content forever takes 30 seconds to prompt.

What the output still needs from you

Even with perfect voice training, AI content needs editing. The standard workflow:

1. Generate the draft. 30 seconds.
2. Edit for truth. Does every fact in the draft match what you know about your business? AI invents things. Fix them.
3. Edit for specifics. Did AI write "many customers" where you could say "three customers last month"? Replace generalities with specifics.
4. Edit for voice drift. Even trained AI occasionally slips into corporate mode. Read the piece out loud. If a sentence feels off, rewrite it.
5. Add a human touch. One personal anecdote, one photo from a real job, one opinion that only you would have. This is what makes AI-assisted content read as yours.

The whole cycle takes 15–30 minutes per piece after the first draft. Compare to writing from scratch (2–4 hours) and you've just 8x'd your content output without sacrificing quality.

The difference in results

A well-trained AI workflow for a home services owner can produce:

• One good blog post per week (vs. one per month hand-written)
• Three social posts per week (vs. one)
• All review responses same-day in your voice (vs. skipped or delayed)
• Customer email follow-ups in 5 minutes instead of 30
• Estimate cover letters personalized per customer, not template copy-paste

Each of these individually moves the needle a little. Together, they compound into a marketing machine that outputs 5–10x more than a manual process, all in a voice that still sounds like a real human who owns the business.

The mistake to avoid

Don't skip the corpus. Don't skip the style guide. Don't just "use AI" with no training and wonder why the output is mediocre.

The owners getting huge ROI from AI content have all done the setup. The ones who haven't are the ones posting about how AI content doesn't work. The gap between those two camps is about two hours of work and a willingness to think clearly about what your voice actually sounds like.

What to do this week

1. Record 30 minutes of yourself answering the five voice-corpus questions above. Transcribe it.
2. Gather 2,000–3,000 words of your existing writing (emails, socials, proposals).
3. Combine into a single voice_samples document.
4. Write a one-page style guide: bullets of characteristics, kill list, example good/bad.
5. Set up a custom GPT or Claude Project with these attached. Test it with one blog post draft.
6. Compare the output to a zero-setup prompt ("write me a blog post about X"). Feel the difference.

Next week: AI for estimates and proposals. The two-hour workflow that saves twenty.

Next in this series

AI for estimates and proposals: the two-hour workflow that saves twenty →

Voice training is the foundation of my Content Engine service. I build the corpus, write the style guide, and set up the custom AI workflow in the first two weeks of any engagement. After that, you're producing content at 10x speed that still sounds like you.

Book a free audit