AI Is a Microcosm of the Entire Market

Why Validating Demand Is So Hard

If you have ever launched a product, course, or collaboration, you have probably run into the same thing.

You ask someone "what do you think of this?" and you get back "it's great!" or "I'm interested." Then when it actually comes to paying, signing up, or scheduling time, they disappear.

This is not because people are deliberately misleading you. The question was wrong. You asked for opinions and predictions about future behavior, and people are actually quite poor at predicting their own future actions. What can serve as real evidence is the time, money, and effort someone has genuinely spent in the past trying to solve this problem.

The trouble is that real-person validation has practical barriers: it's slow, costly, and hard to schedule. To interview five to eight people or send out one or two hundred surveys, just finding the right people and coordinating schedules is enough to kill most ideas before validation ever happens.

And so I kept coming back to one thought.

AI Is a Microcosm of the Entire Market

Today's AI was trained on enormous amounts of human writing. Forum complaints, unboxing reviews, shopping comments, customer service conversations, social media discussions, all of it is in there.

In other words, it carries within itself the compressed tones, hesitations, reasons to buy, and reasons not to buy of countless people across an entire market. It is itself a kind of microcosm of the market.

What the Colgate Study Actually Did

In plain terms, this study asked one question: can AI actually simulate real human purchase intent? It was conducted by PyMC Labs and Colgate-Palmolive, and the full paper title is LLMs Reproduce Human Purchase Intent via Semantic Similarity Elicitation of Likert Ratings, published on arXiv (link at the bottom of this page).

What they did was straightforward:

• They took 57 real market surveys on personal care products, totaling 9,300 real human responses.
• They had AI (testing GPT-4o and Gemini-2.0-flash) play the role of consumers and simulate purchase intent for those surveys.
• They then compared the AI simulations against the real human responses.

They First Discovered a Trap

If you directly ask AI "on a scale of 1 to 5, would you buy this product?" the AI will give you a number that looks reasonable. The problem is that these numbers cluster heavily in the safe, inoffensive middle range, and the overall distribution looks nothing like real human responses.

Their Solution

Their method is called SSR. In plain terms, it means "have AI write its reaction first, then convert the reaction into a score." The order is exactly reversed:

1. First, let AI describe its reaction to the product in natural language, without giving it a score at this stage.
2. Then take that reaction and compare it against the descriptions for each score from 1 to 5, to see which score the meaning most closely resembles.
3. The result is a distribution, not a single number.

The key is in the order: speak naturally first, and the score is derived afterward. Genuine reaction comes first; judgment follows.

What That "90%" Actually Means

To state the conclusion upfront: the purchase intent simulated by AI is already close to the level of consistency a real person achieves when filling out the same survey twice, roughly 90%. The overall distribution of answers also closely resembles real human responses.

How to Run This Yourself

You do not need to replicate the algorithm from the study. What you can actually bring back and use is the order: let AI write genuine reactions first, then organize them into signals.

1. Prepare your inputs first. Organize four things: the product concept (one to two hundred words), the target customer (describe this person's situation, what they are trying to accomplish, and where they are stuck, not just "35-year-old female"), the purchase context (who sees it, who decides, where the budget comes from), and your core assumption (only validate one thing per round).
2. Build a diverse synthetic audience. Choose at least three types of people, and not all fans of yours: a core audience member, an impatient decision-maker who skims for 30 seconds, and someone with a tight budget who always looks for alternatives.
3. Let AI write reactions first, no scoring allowed. This is the core of the entire method. The prompt is below.
4. Organize reactions into four types of signals. Comprehension signals, cost signals, alternative signals, and next-step signals. If you want to convert to scores, put scores last and include reasoning and what was uncertain.
5. Return to real people. Use the output for one of three things: turn AI's doubts into follow-up interview questions, rewrite misunderstood sentences, or design a commitment test (waitlist, pre-order page, small information session).

A Prompt You Can Copy and Use Right Now

The Limits of This Method (Important)

This method works well, and precisely because it works well, its limits need to be stated clearly upfront to prevent overuse.

1. It cannot replace real people. Pre-orders, sign-ups, payments, referrals, retention: these real financial commitments cannot be simulated by AI. Its role is only to surface flaws before you spend time finding real people.
2. That 90% figure has prerequisites. The study data came from personal care product surveys, a category with low price points and fast decisions. If you are selling high-ticket consulting, enterprise services, or long-term courses, the decision logic is entirely different. You cannot apply the 90% figure directly.
3. Do not treat AI scores as validated market evidence. It is a filtering tool, not a conclusion. If no real-person action follows after reviewing the output, no validation has occurred.

I Turned It into a Skill Pack

I have already organized this method into a skill pack on GitHub. It is free and available to download and use directly.

Related Links and Further Reading