I'm a former Amazon marketplace leader and current 8-figure seller. I write about advanced strategies and tactics for Amazon brands, that you won't read about anywhere else. Not for beginners.
βΌοΈ How To Use Synthetic Audiences To Split Test
This Issue's TLDR...
π Did someone forward you this newsletter? First of all, give them a crisp high five when you see them. Second, head over here to subscribe and read past issues. And, be sure to read last week's issue about why you should position instead of predict. Or, read my most popular issue ever: 15 Cool Hacks For Your Amazon Business. β
Look, I know that I talked last week about why you shouldn't try to predict the future, and instead you should position yourself to win, in multiple futures. I stand by that. But, the reality is that prediction is necessary in this Amazon and eCommerce world that we all live in. Fortunately, we have some great third-party tools (e.g., PickFu, Intellivy, ProductPinion) to predict, based on data. These market research tools all come with a cost though. What if you can't afford that cost? Or, what if you're just... Well, good news. Because a group of BIG BRAINED research scientists tested, and proved, that you can replicate the results of human polls/surveys using LLMs and "synthetic audiences." How exactly can you do this? You're going to need to run a series of prompts, which I'll lay out for you below. If you really want to scale this to large samples (e.g., n=1000), you're going to want to use the Gemini Batch API. Honestly, just ask Gemini to walk you through how to do that. Step 1: Upload Your Assets Start a new chat with Gemini 1.5 Flash. Upload the two product images you want to test. Label them clearly in the chat as "Image A" and "Image B" so the model doesn't get confused. Step 2: The "Mass Persona" PromptInstead of doing one persona at a time, you will ask Gemini to generate the audience and simulate their reactions in one go. Prompt 1: Audience Generation & Initial Reaction
Temperature: 0.8 (NOTE: Make this higher if you want more diverse responses)
β "I have uploaded two product images. I want to run a synthetic audience test with n=100 diverse consumers.
β
Step 1: Generate a list of 100 diverse consumer personas (Name, Age, Estimated Income, Primary Shopping Value). Ensure the distribution matches [Your Target Market, e.g., US fitness enthusiasts and nutritional supplement users].
β
Step 2: For each persona, simulate their internal thought process as they look at Image A and Image B.
β
Step 3: Have each persona pick a winner and write one sentence explaining why. Be highly critical. These consumers should only choose a product if it truly resonates with their values.
β
Format the output as a numbered list from 1 to 100."
Step 3: The "Aggregator" PromptOnce Gemini finishes the long list (it might pause; if so, just type "Continue"), you need to turn that wall of text into usable data. Prompt 2: Quantitative Summary "Now, acting as a lead market researcher, analyze the 100 persona responses above. Provide:
β
The Final Score: Total 'wins' for Image A vs. Image B.
β βSentiment Drivers: What were the top 3 visual elements that made people choose the winner? β βSegment Analysis: Did a specific age group or 'shopping value' (e.g., budget vs. luxury) lean more toward one image?β β The Recommendation: Based on this synthetic test, which image should I use for my main ad campaign?" β A Couple Refinements...You might want to dial in to your specific consumer persona with more specificity. Here's a prompt refinement to do that: The Specific Persona RefinementCopy and paste this prompt, filling in the bracketed info for your specific brand.
Prompt: "Generate a list of 100 highly specific, diverse consumer personas for a market research study. For each persona, provide a JSON object with the following fields:
β
Demographics: Name, Age (between [Range]), Primary Transportation. βLifestyle: Amazon Prime Member (Yes/No), Pet Owner (Yes/No), Diet (e.g., Keto, Vegan, Standard), Exercise Frequency. βConsumer Habits: Nutritional Supplement Use (Daily/Occasional/None), Beer Consumption (Frequency/Type), Interest in Cooking (1-10). βAesthetics: Cosmetics & Body Care Habits (e.g., Minimalist, Luxury, Medical-grade). βPsychographics: One core hobby and their #1 priority when buying [Product Category]. β
Ensure the group represents a realistic cross-section of [Specific Target Market, e.g., Suburban Professionals]." Or, you might want to make the consumer responses more raw and real. Use this refinement: The "Critical Persona" RefinementTo avoid the "Politeness Bias" (where LLMs tend to say both images look "nice"), add this System Instruction before you start the simulation:
The Critical Filter: "In this simulation, do not be an agreeable assistant. Act as a discerning, skeptical consumer with a limited budget. You are tired and bombarded by ads. You will only choose a product if the image immediately solves a problem or aligns perfectly with your lifestyle. If an image looks 'cheap,' 'cluttered,' or 'fake,' you must call it out. Use a 1-5 'Hard No to Hard Yes' scale." β BEST from LinkedInI don't think OpenAI will be the ultimate winner of the "AI wars." Yes, it was first. But, in business, it's often the case that: Best Beats First. There's an argument that Gemini comes out on top, as a consequent of its Google ecosystem moat, and just announced partnership with Apple to power Siri. There's also an argument that Copilot becomes the eventual winner, with Microsoft being the software suite of choice for most companies, and that leading people to use Copilot at home (since they use it so much at work). I don't know who will win. But I do think the winner will emerge sooner than most of us think. BEST from XIf you're trying to predict which of your creatives -- ad creatives or listing creatives -- will outperform this year, this is a good reminder why it's important to:
BEST from YouTubeSince I've been talking a lot about predicting in this issue, here's a fun exercise to test how well you understand the average consumer. (Spoiler: You live in an eComm bubble, and the average consumer thinks very differently) |
Best @ Amazon
I'm a former Amazon marketplace leader and current 8-figure seller. I write about advanced strategies and tactics for Amazon brands, that you won't read about anywhere else. Not for beginners.
This Issue's TLDR... Three 2026 Amazon predictions that you should prepare for Predicting what will happen at the intersection of AI and Commerce Predicting what will happen with Rufus π Did someone forward you this newsletter? First of all, give them a crisp high five when you see them. Second, head over here to subscribe and read past issues. And, be sure to read last week's issue with a long list of FREE stuff. Or, read my most popular issue ever: 15 Cool Hacks For Your Amazon Business....
This Issue's TLDR... A bunch of free stuff What LinkedIn *should* be Amazon history, immortalized at Harvard π Did someone forward you this newsletter? First of all, give them a crisp high five when you see them. Second, head over here to subscribe and read past issues. And, be sure to read last week's issue about how to elevate your packaging. Or, read my most popular issue ever: 15 Cool Hacks For Your Amazon Business. HIRE MY AGENCY ($$$) SPONSOR BEST@AMAZON ($$) GET AMAZON ADVICE ($)...
This Issue's TLDR... How to elevate your brand's packaging in 2026 How to trigger loss aversion in your customers The holiday shopping season isn't over yet π Did someone forward you this newsletter? First of all, give them a crisp high five when you see them. Second, head over here to subscribe and read past issues. And, be sure to read last week's issue about demand planning for 2026. Or, read my most popular issue ever: 15 Cool Hacks For Your Amazon Business. HIRE MY AGENCY ($$$) SPONSOR...