A company spun out of Stanford raised $300 million this year to sell AI copies of real people to CVS and Deloitte.
Synthetic customers, basically. A virtual focus group you can ask what it thinks of your ad before a dollar goes behind it. One of the few vendors that even publishes a price, Synthetic Users, starts at $12,500 a year.
I've been running my own version in MindStudio (the no-code agent builder I use for everything) for over a year, and it costs pennies per run. The first sales page I ever put through it, 2 out of 13 synthetic buyers said they'd buy. I read their reasons, rewrote it, ran it again, and got 4 out of 13, with the ones that mattered most to my business (the CMO, the digital advertiser, the agency owner) fired up. Then I mailed it. $5,000 the first night, $5,000 the second night, still converting.
Fam, this is the secret weapon issue. And the secret is that it isn't magic.
A synthetic audience is your customer research, written down, with a grading system bolted on. The AI doesn't know your customer. YOU do. What the panel has that you don't is a perfect memory. It never forgets what your buyers care about, at 11pm, on draft six, when you've fallen in love with a headline they'd scroll right past.
Here's the order: what this thing is, how my virtual focus group runs (with the prompt that powers it), why a machine can even do this, and whether it works.
What is this thing?
You know what a focus group is. Eight or ten people who look like your buyers, sitting around a table reading your ad and telling you what they think. Now make each one an AI persona built from a detailed profile of one real customer segment, run them all at once, and get the transcript in three minutes instead of three weeks.
That's a synthetic audience. A virtual focus group. Mine has 13 members, they've never met, and everything I ship goes through them first. Ads, sales pages, emails, social posts.
How my virtual focus group runs
My build is called the Ad Prediction Agent. There's a Sales Prediction sibling that takes a PDF of a sales page. Both run in MindStudio, both are three steps. Here's the whole thing on one canvas:
Step 1: The panel. Thirteen persona nodes inside a Parallel block (a node is one AI worker with one job, and the Parallel block makes all 13 read the ad at the same time). Sarah Chen the ecommerce owner, Carmen Delgado who runs a family Mexican restaurant in Chicago and spends $2,800 a month on Facebook, Instagram and Google without knowing if any of it puts a family in a booth, Jennifer Walsh the life coach in Denver running $800 a month on Facebook and Instagram with too many tire-kickers in her leads, David Kim the B2B consultant spending $3,500 a month, mostly on LinkedIn... and nine more. Every one of them is a real segment of MY market.
Every node runs the same prompt. Only the first line and the persona file change:
Pretend you are a struggling female ecommerce store owner. You are now part of a panel of prospects reviewing ads as a focus group. Your job is to give your opinion about how the ads make you feel and what they could do better.
For additional context on the role you are pretending to be use {{dataSource "Sarah Chen"}} as your persona.
When I give you ads to look at I want you to give me your raw personal feelings about that ad.
- Does it relate to you?
- Does it address your needs?
- What about it appeals to you?
- What about it turns you off?
- What do you wish the ad said that would make you buy now?
- How could this ad get your attention better?
Do not use generic information about copywriting or advertising, stick to your persona and critique the ads based on your own desires, challenges, fears, frustrations, and goals.
Here is the ad:
{{new_ad}}
(The double-curly-bracket bits are MindStudio placeholders that get swapped for the real text when it runs. {{dataSource "Sarah Chen"}} pastes in Sarah's persona file. {{new_ad}} is whatever you uploaded.)
Read that "do not" line twice. Without it you get ChatGPT's opinion. With it you get Sarah's.
For sales pages I add two questions at the end: "Would you buy this product? Just a simple yes or no." And "What was your reason for saying either yes or no?" That's where the 2 out of 13 came from.
The persona file is 90% of the value. Each one runs 1,200 to 1,500 words. Demographics, professional background, what they've already tried with ads, goals, pain points, an empathy map (what they think, feel, see, say, and do), the tools they use, how they decide to buy, their budget, their deal breakers.
The skeleton can come from AI. Ask your model for "a detailed customer persona and empathy map for a female life coach doing $75K a year" and you'll get a solid start. Mine started that way too, and you can tell. Then you layer in the muscle, which only comes from your customers: support tickets, DMs, email replies, your ad account (what did they click?), and if you can swing it, three to five real customer calls, twenty questions, thirty minutes each.
If the "what she SAYS" section of your persona is stuff you made up, you built a mirror.
One more block goes in the prompt:
Core Traits:
1. You want to topple the giants, beat the gurus, you feel like David vs Goliath.
2. You resonate more with towards pleasure angles more than you do with away from pain angles. For example, you'd rather learn how to make millions than save thousands.
That block is the result of thousands of campaigns. I KNOW which way my buyers lean because I've paid to find out, over and over, since 2007. Writing it into the persona is how you get what you know out of your head and into the machine, where it can't be forgotten.
And here's the trick with it. Say you build 12. If your data says 65% of your buyers lean toward the pleasure angle, tell 8 of them that and let the other 4 lean away from pain. Your panel should look like your customer file, not like your favorite theory about it.
Step 2: The copywriter. All 13 reactions plus the original ad go to one node playing a 25-year direct response copywriter. It lists who was into it, who wasn't, and why (in the sales version, who'd buy), writes a critique, then writes improved variations. And it coaches. It'll tell you something like "you slipped here, this paragraph is talking to a beginner and Sarah isn't one." That coaching is the part that makes YOU better, not just the ad.
Step 3: The ranker. The variations go to one last node playing a data analyst. It reads all 13 reactions and every rewrite, gives each version a Success Probability Score from 0 to 100%, and hands me the winner in copy-paste format.
One thing about that score so your own tool never fools you. It's a RANK. A "72%" won't tell you 72 of 100 people will click. All it tells you is this variation beat the others against the panel.
Upload, wait three minutes, get a report. Who'd buy, who wouldn't, why, a rewrite, a winner.
(The full copywriter and ranker prompts, plus both prediction agents pre-built and ready to install, the ad version and the sales page version, are inside Build Club. Swap in your personas and you're running by tonight.)
Ok so why does this even work?
A language model (an LLM, the engine under ChatGPT and Claude) was trained on the internet. Every product review, every "is this worth it?" forum thread, every one-star rant and five-star gush ever typed. The biggest focus group ever assembled, and it read every word.
And the ONE thing an LLM does is predict what comes next. So when you ask "what would a struggling ecommerce owner say about this headline?" you're asking it to do the thing it's best at.
Left alone, it plays a generic customer. Hand it 1,500 words of real detail about Carmen and it plays a much better Carmen. Not the real one, which is why the winner still gets tested with real money.
A Stanford team showed the real-detail part with 1,052 real people (it's still a preprint, but the sample's big). Two-hour interview each, an AI clone built from every transcript, then the clones and the humans took the same tests. The clones matched their humans 83% as well as the humans matched THEMSELVES when they retook the test two weeks later, and clones built from real interviews beat clones built from demographics alone by 9 points. QUALITY IN, QUALITY OUT.
Harvard Business School and Microsoft, in a working paper, pointed a plain ChatGPT-era model at pricing surveys. It got the SHAPE right (raise the price, fewer people buy, same as humans) and the number wrong. It said fluoride in toothpaste was worth $8.20 to people. Real people said $2.60.
So ignore every accuracy badge you see on a vendor's homepage. Those numbers measure how well the AI agrees with a survey, and nobody, me included, has published a study showing a synthetic panel's pick matched what a real ad did in market. Here's the truer sentence, and the one I want you to keep:
A synthetic audience is a compass, not a ruler. It'll tell you which ad is pointed the right way. Don't ask it how far.
Ok so does it work?
You think that page did $5,000 two nights running because I know how to write a headline?? I've known how to write a headline since 2007. It did $5,000 because thirteen synthetic buyers told me what was wrong with it before I mailed it.
Now everything goes through the panel. I'm picking better winners than my gut picked, and my gut has 20 years and tens of millions in ad spend behind it. The panel doesn't have better taste than me, it has a better memory. It's my customer file with a pulse. It's the version of me that read every support ticket and never got tired, and it never forgets that Carmen can't track a Facebook ad to a booth, even on the nights I do.
And I'm not the only one anymore:
- The Times of London built a synthetic panel of its own readers. Their director of data operations put it in one line: "The same team, ten times the output." They liked it enough to start selling it to advertisers.
- Nestlé Purina and Ipsos tested cat food with 50 real people plus 150 synthetic ones against a 225-person all-human test. The two tests moved almost in lockstep (a 0.92 correlation, where 1.0 is identical), and they landed on the same go or no-go decision 95% of the time.
- Colgate-Palmolive and a stats shop called PyMC Labs, in a preprint, reproduced 57 surveys asking how likely you'd be to buy (9,300 real answers) at 90% of the humans' own consistency.
- US Bank's CMO says his synthetic audiences overlapped 90 to 95% with real people on the themes.
- Omnicom's Australia and New Zealand shop has run 60-plus tests and projects at about a tenth of a normal research budget.
Same idea, different fuel. They feed in survey answers from thousands of real people. You feed in support tickets and five customer calls. Which is why the persona file is the whole game.
So here's the recap
A synthetic audience is a virtual focus group built from your real customer research. The persona file is 90% of it, the prompt makes it talk, the copywriter makes it useful, and the ranker picks the winner. It isn't magic. It's your best data with a perfect memory.
Twenty years ago I would've paid a research firm five figures and waited three weeks to hear what thirteen customers thought of a sales page. Now it costs pennies to run and it's done before my coffee gets cold.
And it doesn't have to work miracles to pay for itself. If it talks you out of ONE losing ad this year, it paid for itself a hundred times over. If it finds you one winner, that's thousands. For some of you reading this, that's millions.
This is your secret weapon. The big brands are paying $12,500 a year and up for theirs. You're going to build yours for the cost of your MindStudio plan and a weekend spent writing down what you already know about your customers.
Go build your focus group. Then let it read everything before your customers do.
Sources cited in this issue
- Simile (the Stanford spinout), $100M Series A (Feb 2026) and $200M Series B at a $2B valuation: TechCrunch, Julie Bort, 2026-07-30, https://techcrunch.com/2026/07/30/synthetic-user-startup-simile-raises-200m-at-2b-valuation-5-months-after-100m-series-a/ . Customers CVS Health and Deloitte per https://www.simile.com/
- Synthetic Users pricing page, $12,500/yr entry tier: https://www.syntheticusers.com/pricing
- Park, Zou, Shaw, Hill, Cai, Morris, Willer, Liang and Bernstein (Stanford, Google DeepMind), "LLM Agents Grounded in Self-Reports Enable General-Purpose Simulation of Individuals," arXiv:2411.10109 v3, Jun 28, 2026, preprint (1,052 participants; interview-based agents 0.83 normalized accuracy vs 0.74 for demographic-only agents, against each person's two-week retest): https://arxiv.org/abs/2411.10109
- Brand, Israeli and Ngwe (Harvard Business School, Microsoft), "Using LLMs for Market Research," HBS Working Paper 23-062, rev. Apr 30, 2026 (GPT-3.5 Turbo; fluoride willingness to pay $8.20 vs $2.60 human): https://www.hbs.edu/ris/Publication%20Files/23-062_1f58623a-ee21-44b9-a262-276047bc5543.pdf
- The Times / Electric Twin synthetic reader panel: Digiday, Tim Peterson, 2025-11-04, and Jessica Davies, Apr 2026; Chris Courtney-Smith (Director of Data Operations, The Times) quote from https://www.electrictwin.com/case-studies/times
- Nestlé Purina and Ipsos, hybrid concept test, 50 human + 150 synthetic vs a 225-person all-human test (r = 0.92; same decision 95% of the time): Ipsos deck, Quirk's Event New York, 2025-07-24.
- Maier et al. (PyMC Labs, Colgate-Palmolive), "LLMs Reproduce Human Purchase Intent via Semantic Similarity Elicitation of Likert Ratings," arXiv:2510.08338, preprint (57 surveys, 9,300 responses, 90% of human test-retest reliability): https://arxiv.org/abs/2510.08338
- US Bank CMO Michael Lacorazza on synthetic audiences ("90% to 95% overlap"): AdExchanger, Allison Schiff, 2025-06-25.
- Omnicom Media Oceania and Ideally, "over 60 tests and projects" at "about one-tenth of a traditional research budget": Digiday, Michael Bürgi, 2026-08-07.
- Build Notes #001 (April 2026), the 2-of-13 / 4-of-13 / $5,000 sales page story.
Continue with Build Notes+
This Monday deep dive is free in full. For complete Thursday builds and supporting files, join Build Notes+.
