Hey there,
Isn't it strange that with marketing, we put our message (content, ads, videos, etc.) out first and course correct based on results?
You would think in the age of AI we could at least have an educated guess about what works before spending hundreds, sometimes thousands, on marketing.
Or sinking hours into content with no idea how it will land.
Back in January I had a 2 hour plane ride to kill and spent the entire flight mulling this problem over in my head.
I was headed to an AI/media mastermind down in Florida with 25 other founders.
The very first person I met at that event was a guy named Justin Brooke. He's a 20 year advertising vet with millions in online sales to his name.
And he showed me one of the coolest AI marketing strategies I have ever seen.
What's cool about Justin's strategy is that he "stole" it from Big Media.
He watched companies like Harvard, the New York Times, and other big media outfits use this exact approach, then adapted it to his own small business.
So he set to work building an entire focus group of AI agents that tell him (with a scary level of accuracy) how well his ads and content will perform before he spends a dime promoting them.
βI recorded a whole video interview breaking down the process with him. But here is the short version, plus how to build your own.
Justin calls it predictive wear. The academics call it a synthetic audience. Most people call it a virtual focus group. At the end of the day it's all the same thing.
Here is the core idea:
We already tell AI to pretend to be a copywriter, an SEO, a blog writer. Justin just tells it to pretend to be his customer. Then he asks that customer what it thinks of his ads/content.
The old way of marketing is reactive. You write the copy, push it out with ad spend, and find out if it converts after the money is gone. You learn by spending.
The new way flips it. You get the feedback virtually, before anything goes out the door.
Here's how he built the machine. He pastes his ad copy into a simple form. That copy gets passed to 13 different personas, each one a variation of his ideal customer.
They all answer the same set of questions.
- Does it relate to me?
- What turns me off?
- Would I buy this?
All that feedback goes to a copywriter agent, which rewrites three new versions based on what the panel said.
Then a prediction engine scores each version and tells him which one to run, like "this one has an 86% likelihood to convert."
The naysayers are loud with this one. He's constantly dealing with people who claim "this stuff can't work".
Two answers. First, the studies.
Justin points to Harvard and Stanford research on this, and says the New York Times found it 92% accurate against their real human focus groups, which is why they use it to pick which headlines and articles to run.
Second, the people who say it does not work are the ones who skip the hard part. They write a lazy prompt that says "pretend to be a struggling business owner" and get garbage back.
The sauce is the persona behind the prompt, not the prompt itself.
Each of Justin's personas is a 1,400-word dossier.
- Demographics
- Background
- Goals
- Pain points
- Emotional frustrations
- An empathy map of what that person thinks, feels, and fears
That depth is what gets you to 85 to 92% accuracy instead of fake-sounding nonsense.
And the economics are absurd. A top 1% copywriter charges 100 to 500 dollars to write three ad variations. This costs Justin about 13 cents.
He has personally made around $260,000 off this system, and clients with none of his experience are seeing 4 to 6X returns on their ad spend.
Most marketing agents pay off months down the road. This one pays off the morning after you run it.
Here's how to build your own version this weekend:
Step 1: Pick a no-code agent builder.
Justin uses MindStudio because it gives you access to every major model without wiring up your own APIs. n8n or Make work too. Set it to trigger on demand so you just click run.
Step 2: Build your personas. (This is the part you CAN'T skip).
Create 8 to 13 personas, each a variation of your ideal customer. Mix it up: different ages, genders, and business types, and crucially, mix beginners with successful buyers so the panel also tells you who is NOT a fit.
Each persona needs a 1,400-word dossier: demographics, background, goals, pain points, emotional frustrations, and an empathy map. Use Claude in deep research mode to help you write them. Save each one as its own data source.
Step 3: Add your own ICP as a separate data source.
This is the doc about your company and your products, so the copywriter knows what it is actually writing about.
Step 4: Write the panel prompt.
Tell each persona: "You are part of a panel of prospects reviewing this ad as a focus group." Then ask the questions that matter. Does it relate to you? What appeals to you? What turns you off? What would make you buy right now?
For sales pages, add a straight yes or no: would you buy this? Tell it to answer as the persona, not as a generic copywriting expert.
Step 5: Run the personas in parallel.
Select them, right click, and group them as parallel instead of sequential. Otherwise it runs one at a time and takes forever. If 13 bogs down, split them into two parallel groups.
Step 6: Add the copywriter agent.
Start the prompt with "embody a world-class copywriter with 40 years of experience." Justin found that "embody" beats "pretend you are," because the model treats it as become this, not fake this.
Feed it the original ad plus all the panel feedback, and have it write three new versions in your voice. Ask it to format the response like an internal team email that quotes the key feedback, shares a few insights, then gives the rewrites.
Step 7: Build a simple prediction engine.
This is a scoring matrix. Weigh copywriting best practices, the panel feedback, and the variations, then have it score each version's likelihood to convert. Ask Claude to build the matrix, or pay a math professor 500 dollars to build a real one (worth it).
Run the version that scores 85% or higher.
For sales pages it's the same flow. Export the page as a PDF, upload it as step one, and watch the yes count. Get a low score, make the changes it suggests, rerun, and rinse until it is dialed. Then ship.
That is the whole thing.
A weekend to build it, about 13 cents to run it, and you never send another ad or sales page out blind again. You stop shooting your shot and hoping. You test it against your customers first, then ship the winner.
Build it this week. Reply and tell me what piece of copy you are running through it first. I read every response.
- Corey
A few notes/tweets/cool things sourced from the AI community.
1)
This is amazing.
2)
Now that anyone can build anything, this is the only way to win.
3)
|
11:44 PM β’ May 31, 2026
|
|
β
A business in a (agent) box.
Thanks for reading the Corey's Notes newsletter. I'd appreciate it if you sent it to a friend that might find it interesting.
Be sure to check out t he Build With AI podcast on Apple, Spotify, and YouTube. If you're a non-technical entrepreneur who wants to learn how to integrate AI into your business, you'll love it.
Be back next week.
-Corey
β