You can tell when an ad was written by AI. And so can your audience — even if they can't articulate why.
That's the problem nobody in the AI-hype cycle wants to talk about. Most agencies have already integrated AI into their copywriting workflow. The question isn't whether to use it anymore. It's whether you're using it in a way that actually works.
At Vixi, we've been testing AI-generated ad copy on real client campaigns since early 2024. We've run split tests, tracked the metrics, and built a system that consistently produces copy that converts. This post is a breakdown of exactly how we do it — what patterns to avoid, how to structure prompts so the output doesn't sound like it came from a content mill, and what the data actually shows when you compare AI copy against human-written ads.
No hype. Just the framework.
What "Sounds Like AI" Actually Means (and Why It Kills Performance)
Before we get into fixes, let's be specific about the problem.
AI-generated ad copy fails in predictable ways. It's not random. The language models that power tools like ChatGPT and Claude are trained on enormous amounts of marketing content — including a lot of bad marketing content. They've learned the patterns of generic ad copy so well that they default to them.
Here's what this looks like in practice:
- "Are you tired of struggling with X?"
- "Our cutting-edge solution helps businesses like yours..."
- "Take your business to the next level."
- "In today's fast-paced digital landscape..."
- "We've helped hundreds of businesses achieve their goals."
Every one of those phrases is technically correct English. None of them will make anyone click anything.
The reason isn't aesthetic. It's trust. Audiences have seen these patterns so many times in spam emails, low-quality display ads, and cold outreach that their brains have learned to tune them out. They don't consciously think "this was written by AI." They just don't feel anything when they read it — and in advertising, not feeling anything is the same as not seeing it.
The fix isn't to stop using AI. It's to stop using it like a magic "write my ad" button.
Here's a quick illustration of the difference:
| AI Default Output | What It Should Sound Like | |---|---| | "Streamline your workflow with our powerful platform" | "We cut our clients' reporting time from 3 hours to 20 minutes. Here's how." | | "Are you tired of struggling with lead generation?" | "We added 14 qualified meetings to one client's calendar in 30 days. No cold calls." | | "Our team of experts is here to help your business grow." | "Most of our clients come to us after wasting $5k+ on ads that didn't convert. We fix that." |
The right column is specific. It names outcomes, timeframes, numbers. It sounds like something a real person would say at a coffee meeting, not something generated from a template.
Getting there consistently requires a system. Here's ours.
Voice of Customer Research — The Foundation AI Can't Fake
The single biggest reason AI ad copy fails is that it has no real customer language to work from. You give it a product description and a target audience, and it generates something plausible. But plausible isn't the same as resonant.
Real copywriting starts with how customers actually describe their problem — in their own words, before a marketer touches them. AI can't invent that. But it can use it extremely effectively if you feed it properly.
At Vixi, we run a VoC (Voice of Customer) research pass before writing a single line of copy for any new client. Here's exactly what that looks like:
1. Mine 3-star reviews, not 5-star. Five-star reviews are enthusiastic but vague ("Great service, highly recommend!"). Three-star reviews have specifics — what almost didn't work, what the customer was worried about, what they wish had been different. That's the language that maps to real objections and real desires.
2. Pull sales call transcripts. If the client has any recorded calls, we listen for the exact words prospects use to describe their problem. Not the sales team's interpretation of the problem — the prospect's actual words. There's almost always a gap between the two.
3. Scrape Reddit and Facebook groups. Find the communities where the target audience talks to each other — not to brands. The way people describe their problems in an unguarded forum conversation is completely different from how they respond to survey questions.
4. Collect "before and after" language from existing customers. Ask customers: "What words would you have used to describe this problem before you found us?" That phrasing goes directly into ad copy.
The output of this process is what we call a voice bank — a document of verbatim customer quotes, organized by theme (pain points, desired outcomes, objections, results). Everything we prompt the AI with draws from this bank.
Here's the prompt template structure we use:
You are writing Facebook ad copy for [CLIENT NAME], a [brief description].
Their target customer is [specific person description].
Their customers describe their main problem as:
"[VERBATIM QUOTE 1]"
"[VERBATIM QUOTE 2]"
"[VERBATIM QUOTE 3]"
Their customers describe the result they want as:
"[VERBATIM QUOTE 4]"
"[VERBATIM QUOTE 5]"
Write a 125-word primary text for a Facebook conversion ad.
- Open with one of the exact phrases above (or a very close variation)
- Do NOT use any of the following phrases: "cutting-edge," "streamline," "leverage," "game-changer," "take your business to the next level," "in today's digital landscape"
- End with a direct call to action (not "learn more")
At Vixi, we've found that ads opening with verbatim customer language consistently outperform generic AI copy by 20–40% on CTR. The audience reads the first line and thinks "that's exactly my problem." That's not a coincidence — it's their own words reflected back at them.
The Prompting Frameworks That Actually Work
Once you have your voice bank, you need prompts that actually extract good output. Here are three frameworks we use regularly.
The Anti-Pattern Prompt
The most effective single change you can make to your AI copy prompts is telling the model exactly what NOT to write. Build a "banned phrases" list for each client and include it in every prompt.
This sounds obvious but most people don't do it. The model defaults to its training patterns unless you explicitly break them.
A good banned list for a B2B service client might include: "streamline," "seamless," "cutting-edge," "innovative solutions," "world-class," "best-in-class," "take your business to the next level," "in today's competitive landscape," "our team of experts," "reach out today."
The Persona + Scenario Prompt
Instead of "write an ad for X product," try this framing:
You are [SPECIFIC PERSON: e.g., "a 42-year-old owner of a 12-person HVAC company in Dallas"].
It's Monday morning. You just scrolled past this ad on Facebook between checking emails.
You've been burned by two marketing agencies in the past two years.
You're skeptical but the business really does need more leads.
Write the ad that would make this person stop scrolling.
This forces the model to reason from audience perspective rather than copywriter perspective. The outputs are noticeably more grounded and specific.
The First Draft is Garbage Protocol
Never use the first thing the model generates. Always prompt for 5 variants and treat all of them as raw material, not finished product.
This isn't just about having options. LLMs front-load their most trained-in, "AI-sounding" patterns in their first output. By asking for 5 variants with specific constraints on each, you force the model to reach for less default patterns:
Generate 5 distinct Facebook ad headlines for [CLIENT].
Each headline must be under 40 characters and reference a specific, measurable outcome.
Do NOT use: [banned phrases list]
Headline 1: Opens with a number
Headline 2: Opens with a question (not "Are you tired of...")
Headline 3: Makes a bold, contrarian claim
Headline 4: Names the specific audience (not "business owners" — more specific)
Headline 5: References a specific timeframe ("in 30 days," "by Q2," etc.)
The best headline usually comes from headlines 3, 4, or 5 — after the model has exhausted its default patterns.
The Rewrite Loop — Turning AI Draft into Human-Sounding Copy
Even with a good prompt, the first output needs work. We run every AI draft through a 3-pass rewrite process before it goes anywhere near a client.
Pass 1 — De-genericize. Read every sentence and ask: is this specific or general? Every general statement gets replaced with a specific one. "We help businesses grow" becomes "We've added an average of 18 new clients per month for the HVAC companies we manage ads for." If you can't make it specific, cut it.
Pass 2 — Add friction. Real copy has an edge. It says something surprising, makes a claim that needs defending, or asks a question the audience hasn't heard before. Add one element that breaks the pattern — an unexpected word, a counterintuitive claim, a very specific detail that makes the reader think "how do they know that?"
Pass 3 — Read it out loud. This is non-negotiable. If you wouldn't say it to another person in a normal conversation, it doesn't go in the ad. This catches hedging language ("may potentially help"), passive voice, and any construction that sounds like writing rather than speaking.
We also use what we call the Texan test: would a Dallas business owner say this at lunch to a friend? If not, rewrite it.
We ran this 3-pass loop on 47 ad sets across local service businesses in the DFW market. The 3-pass version outperformed raw AI copy on every metric we tracked — not by a little, but consistently and significantly.
A/B Testing AI vs. Human Copy — What the Data Shows
We've been rigorous about testing. Here's what we've actually found running controlled comparisons across client campaigns:
| Copy Method | Avg CTR | Avg CPC | Production Time | |---|---|---|---| | Zero-shot AI (no framework) | 1.1% | $2.40 | 5 min | | Human-written | 2.8% | $1.20 | 3–4 hrs | | Voice-trained AI + rewrite loop | 2.4% | $1.35 | 30 min |
Zero-shot AI copy — meaning a basic prompt with no VoC input, no banned list, no rewrite pass — performs about as well as mediocre human copy on a bad day. It's not terrible. It just doesn't compete with either the human-written or the framework-assisted AI outputs.
Human-written copy still wins on pure performance. When a skilled writer has deep customer research and time to work with it, the output is better. That's not going to change.
But the voice-trained AI + rewrite loop gets 85% of the way there at roughly 10% of the time cost. That's the realistic value proposition. You're not replacing your best copywriter. You're replacing the blank page, the first draft, and the "we need 8 ad variants by Thursday" panic.
The honest framing: AI doesn't replace human copywriters — it replaces the part of copywriting that was never about creativity in the first place.
How to Build This System for Your Agency (or Your Business)
Here's the full system in order:
Step 1: Build your voice bank. Spend 2–3 hours pulling customer language from reviews, sales calls, and forums. Organize it by pain points, desired outcomes, and objections. This is the highest-leverage work in the entire system.
Step 2: Create your banned phrases list. For each client or brand, list every AI-sounding phrase that should never appear in their copy. Update this as you find new offenders.
Step 3: Set up prompt templates per campaign type. Awareness campaigns, retargeting, and conversion campaigns all need different framing. Build a template for each and store them somewhere the whole team can access.
Step 4: Run the 3-pass rewrite loop on every output. This is the discipline part. It's easy to skip when you're in a hurry. Don't. The loop is what separates copy that performs from copy that technically exists.
Step 5: A/B test relentlessly. Run every new ad against the control. Track CTR, hook rate (for video), and landing page conversion — not just impressions. The data will tell you what's working faster than any gut check.
The initial setup takes about 2 hours. After that, producing a full set of ad variants for a new campaign takes around 30 minutes. That's the math that makes this worth building.
Work With Vixi
If you're running paid ads for your business in the Dallas area and your current copy isn't converting, this is usually the reason — not the targeting, not the budget, not the creative. The words aren't resonating.
At Vixi, we build this entire system as part of our ad management service. VoC research, prompt frameworks, rewrite loops, A/B testing infrastructure — all of it, done for you.
Book a free strategy call and we'll audit your current ad copy — no strings attached. We'll tell you exactly what's not working and what it would take to fix it.