The problem isn't using AI. Most people who've tried ChatGPT or Claude for marketing copy have discovered that within about 30 seconds. The problem is that the output sounds like it was written by someone who has read a lot of marketing copy but never actually tried to sell anything.
There's a specific workflow that fixes this. It's not complicated, but it requires you to understand why AI produces the output it does — and what you need to give it to get something usable.
What you'll learn in this post
- Why AI marketing copy sounds the way it does
- The three inputs that change everything
- How to structure a prompt that produces usable output
- The editing pass that makes AI copy sound human
- What AI should and shouldn't write for you
Why AI copy sounds the way it does
AI language models are trained on enormous amounts of text. A huge portion of that text is marketing copy — and a huge portion of that marketing copy is mediocre. So when you ask AI to "write a caption for Instagram," it produces the statistical average of every Instagram caption it's ever seen. That average is enthusiastic, vague, and sounds like nobody in particular.
The fix is specificity. The more specific your input, the further the output gets from the statistical average.
This is the core insight. Everything else follows from it.
The three inputs that change everything
Before you type a single word of a prompt, you need three things ready:
1. A real example of your own voice
Paste in a piece of copy you've written yourself — an email, a caption, a paragraph from your website — and tell the AI to match it. Not "write in a friendly tone." Give it actual words you've used and tell it to write like that. AI is very good at style imitation when given a concrete sample. It's useless when given adjectives.
2. Something specific about your customer
Not "small business owners" or "mums aged 25-45." One specific thing your customer believes, worries about, or says out loud. "My customers always say they've tried everything and nothing works." "My customers are embarrassed to admit they don't understand their Google Ads account." That's the specific observation AI needs to write something that actually resonates.
3. One job for the copy to do
Not "write a good caption." What should happen after someone reads it? Click a link? Feel understood? Book a call? Name the single outcome. AI will optimise for whatever you describe — if you don't describe anything, it optimises for sounding helpful, which is useless.
How to structure a prompt that produces usable output
Here's a before and after.
Generic prompt: Write an Instagram caption for a digital marketing agency.
What you get: "Ready to take your digital marketing to the next level? Our expert team is here to help you grow your business and reach your goals. DM us today! #marketing #digitalmarketing #smallbusiness"
Specific prompt: Write an Instagram caption for Hedgehog Marketing, a digital marketing consultancy based in Australia. Here's an example of how I write: "Your Google Ads account isn't failing. It's running blind. The fix isn't a bigger budget — it's knowing what the data is actually telling you." Write a caption in this voice about why most small business owners are running paid ads without proper tracking set up. The goal is to make someone who's doing this feel understood, not criticised. Single paragraph, no hashtags.
What you get: Something that sounds like a person wrote it, because you've given the AI enough to work with.
The structure of that prompt:
- Who you are and where you're based
- A concrete voice sample
- A specific topic
- The emotional job it needs to do
- Format instructions
That's it. Those five elements will transform your output.
The editing pass that makes AI copy sound human
Even with a good prompt, AI output needs one editing pass before it's usable. This pass takes two minutes and catches the tells.
Remove the transitions. AI loves "Furthermore," "In addition," "It's worth noting that," and "This is why." Delete every one of these. Replace with a new sentence or nothing.
Break up the meter. AI writes in metronomic rhythm — sentences that are all roughly the same length. Read your output out loud. Where it sounds sing-songy, break one long sentence into two short ones, or combine two short sentences into one long one. Vary the pattern.
Kill the adjective clusters. "Proven, effective, results-driven strategies" means nothing. Pick the one word that matters and cut the rest.
Add one specific detail AI couldn't have known. A real client result, a thing that happened in a meeting, a specific number. One grounding detail is enough to make the whole piece feel real.
Check the ending. AI almost always ends with a call to action that sounds like a brochure. Rewrite the last sentence yourself.
That editing pass, applied consistently, is the difference between AI copy and copy that sounds like you used AI as a research assistant.
What AI should and shouldn't write for you
Good uses:
- First drafts of anything structural (email templates, post frameworks, FAQ sections)
- Headline and subject line variations to test
- Repurposing long-form content into shorter formats
- Research and summary of topics you're going to write about yourself
- SEO meta descriptions (these are functional, not expressive)
Bad uses:
- Anything where your specific experience and opinion is the value (strategy advice, case study analysis, client recommendations)
- Brand voice copy where sounding like a specific person matters
- Anything you haven't checked against reality (AI will fabricate statistics, misquote platforms, and describe features that don't exist)
The pattern: AI is excellent at structure and first-pass speed. It's poor at originality, voice, and accuracy. Use it for the former and stay hands-on for the latter.
Frequently Asked Questions
Will people know I used AI to write my content? If you use the workflow above, no — and this is the wrong question anyway. The question is whether the output is good. Lots of human-written copy is mediocre. The goal is output that serves your audience, not output that passes a test.
Which AI tool is best for marketing copy? For most marketing applications, Claude and ChatGPT are comparable. Claude handles longer documents and structured prompts slightly better in our experience. The tool matters less than the quality of your inputs.
How much editing should good AI copy need? If a prompt is working well, the output should need about 20-30% editing — mostly the tells listed above plus personalisation. If you're rewriting 70%+, the prompt isn't specific enough. Go back and add more voice sample, more specific customer observation, or more precise formatting instructions.
Can AI write a full blog post I can publish without editing? Technically yes. But you'll get a generic post that sounds like a generic post. If SEO is the goal and you don't care about voice, that might be fine. If the post is supposed to represent your expertise and perspective, you'll need to be in it somewhere.
The short version
AI produces the average of everything it's seen. Your job is to give it enough specific inputs — your voice, your customer insight, the job the copy needs to do — to drag the output away from average. Then edit out the tells. That's the whole workflow.
