A subtle but crucial detail in this study is that the researchers didn’t just generate lots of simulated responses—they specifically leveraged the LLM’s next-token probabilities (the internal “weights” the model assigns to each possible answer) to estimate how a large population would respond, rather than simply sampling random outputs over and over. This approach mathematically mirrors how real-world market demand is formed (through the “law of large numbers”), smoothing out individual irrationality and randomness without needing thousands of explicit persona runs.

Why it matters: Most experts might assume you need to run the LLM hundreds or thousands of times with different random seeds or persona variations to simulate a crowd, but this study shows that you can extract a near-instantaneous ""aggregate"" market response by tapping into the model’s probability distribution for each answer. This not only speeds up the simulation dramatically, but gives a more stable and representative result than brute-force sampling or naive persona generation.

Example of use: Suppose a pricing manager wants to know if a $99 or $149 price point will be seen as “too expensive” by the majority of potential buyers. Instead of running 1,000 separate AI personas and counting votes, they can prompt the LLM with the key scenario and directly read the probability assigned to “Too Expensive” versus “Good Value” for each price. The manager instantly sees, for each price, what percent of the simulated market would reject or accept the offer—mirroring how a real brand might estimate market acceptance curves.

Example of misapplication: A team uses traditional random sampling—having the LLM generate hundreds of long-form persona answers for each scenario—then averages the results, thinking this will give them a realistic market prediction. In reality, this wastes time and compute, and risks being skewed by outlier responses or prompt drift, especially if prompts are inconsistently worded. By overlooking the LLM’s internal probability outputs, they miss out on both efficiency and the more statistically robust aggregate that this study demonstrates is possible.

• Use many simulated responses, not just one-off answers: The research shows that asking the AI to generate lots of responses and then averaging them gives a much more realistic picture of what a market would do, because it cancels out individual quirks or random errors that a single AI ""persona"" might make.

• Vary the scenarios or prices you test, not just the question wording: Running your experiment across a range of possible prices or offer conditions and looking at the aggregate pattern is more reliable than just asking about one specific scenario. This is especially useful for pricing studies or testing market reactions to new offers.

• Carefully design your survey choices to encourage clear decisions: The study found that using “extreme” answer choices (e.g., “Too Cheap” vs. “Too Expensive”) forced the AI to take a stand, making simulated demand curves and results much more interpretable than using vague or neutral options.

• Don’t worry too much about which AI model you use—as long as you aggregate and design prompts well: The difference between various AI models (like GPT-4, Llama, Mistral) was much less important than how you structured the experiment and the use of multiple samples. This means you don’t need the latest or most expensive model to get good results.

• Expect better accuracy for spotting trends than for predicting exact numbers: The AI method was best at capturing what direction things are moving (e.g., which product will get more interest), but not as strong at guessing the exact level of demand or price. Use it for “which option wins” or “which way the market is shifting,” not for pinpoint forecasts.

• Simulated audiences are robust for comparing different options, not for absolute forecasting: The method worked well for testing which idea or message performs better, which is ideal for A/B testing, concept ranking, or pricing tiers—less so for predicting exactly how many units will sell.

• Run larger, more varied virtual experiments for more stable results: The more price points or conditions you test and the more AI responses you aggregate, the less random noise in the output, so bigger synthetic studies (more “virtual buyers/sellers” and more test cases) give more trustworthy answers.

Prompt Explanation: The AI was prompted to assume the role of a market participant (buyer or seller) and respond to whether a given financial asset price is ""Too Expensive"" or ""Too Cheap"" (for buyers) or ""Too Low"" or ""Too Good"" (for sellers), based on provided contextual information and price, in order to simulate market demand and supply responses using next-token probabilities.

You are an investor buying [ASSET] stock. Given a set of news and the current market price, consider if it is too expensive or too cheap.
News: [contextual information]
Price: [relative change from previous Close]
Options: Too Expensive, Too Cheap

(For seller agents)
You are an investor selling [ASSET] stock. Given a set of news and the current market price, consider if the bid price is too low or too good.
News: [contextual information]
Price: [relative change from previous Close]
Options: Too Low, Too Good

1. Define audience personas: Identify 3–5 realistic customer profiles relevant to your brand, such as ""Tech-savvy urban professional,"" ""Budget-focused college student,"" and ""Home-focused parent."" For each, write a short description including age, lifestyle, shopping priorities, and attitudes toward technology.

2. Prepare the pricing concept prompt: Draft a short, consistent description of the product and each price point. Example:
Product: ""The EchoMax is a wireless smart speaker with high-fidelity sound, voice assistant, and multi-room support.""
Price Points: $49, $99, $149.

3. Create the persona prompt template: Use the following structure to ensure clear, consistent responses:

You are simulating a [persona description].
The product you are considering is: ""[Product description]""
The price is: $[insert price point]
Would you buy this product at this price? Why or why not? Respond as your persona would, in 2–4 sentences.
Choose only one: ""Too Expensive"" or ""Good Value"".

4. Run the simulation for each price point and persona: For each persona, input the prompt at each price point (e.g., $49, $99, $149). For more robust results, generate 10–20 responses per persona/price by rewording slightly or varying the AI’s randomness.

5. Aggregate the responses: For each persona and price, count how many chose ""Too Expensive"" vs. ""Good Value,"" and note any recurring reasons or objections from the open-ended answers.

6. Chart and compare results: Create a simple table or chart showing, for each persona, at which price point ""Good Value"" peaked and what reasons were given. Highlight patterns, such as ""Tech-savvy professionals prefer $99, citing features; budget students only choose $49.""

7. Summarize pricing insights: Write a brief summary for stakeholders, noting the most attractive price points for each segment, key motivators (e.g., sound quality, brand, price), and most common objections (e.g., too expensive for students, missing feature for parents).