Product research AI prompts cannot tell you whether a product will sell, because demand is not in the text a model reads. What they can do is turn a vague hunch into a structured set of questions, pressure-test your reasoning, and tell you exactly which checks to run yourself with real data. The crowded "find winning products with AI" format usually produces a confident list instead.
The honest workflow has three stages. Use a model to expand and stress the idea, use platform tools and a supplier to verify it, then use the model again to write the order and the test. Below are prompts for each stage, plus an explicit list of what no model can answer for you.
What AI can and cannot verify about demand
Models read text. They do not read your marketplace's live order volume, your category's return rates, or the ad costs in your specific market. That means anything a prompt returns about "demand" is either something you told it or something it inferred from general patterns, and general patterns do not know that the niche you are considering is saturated in your country.
Use models for the parts that are genuinely text-shaped:
- Naming and framing: what a buyer calls this problem, what they compare it against.
- Structured doubt: the reasons this product fails, which a model can list faster than you can dismiss them.
- Research plans: which sources to check, in what order, and what answer would change your mind.
- Output drafts: supplier emails, listing copy, test structures.
Use tools for the parts that are numeric:
- Actual sales and rank data from marketplace tools you subscribe to.
- Search result counts and competitor review counts.
- Ad cost and click estimates from the ad platform.
- Return rates, sizing or compatibility issues, and supplier stock levels.
Any prompt that blurs this line will produce a number with no source. Ask for the source or delete the number.
Stage one: turn a hunch into a testable claim
Most niche ideas fail as a sentence. "A product for dog owners" is not testable. "A slow feeder for large breeds that solves mess at feeding time" is.
I am considering [product idea] for [market and marketplace]. The buyer is [specific person in a specific situation]. Do not tell me whether this is a good product. Do the following instead: 1. Rewrite my idea as one specific claim about a problem the buyer already has, in their words, under 25 words. 2. List the eight most plausible reasons this product fails commercially. For each, give the earliest signal that would reveal it, and roughly how long that signal takes to appear. 3. List what I would need to believe for this idea to fail, and mark which of those beliefs I could actually verify this week with free data. 4. Name three adjacent products a buyer might buy instead, and say in one line why they would choose that instead. Do not invent statistics, prices, sales figures or market sizes. If you need a number, write [CONFIRM: source needed] instead. Separate what you are reasoning from what you are assuming.
The eight failure reasons are the valuable output. A list of reasons your idea could fail, each paired with a signal, turns optimism into something you can watch.
Stage two: verify with the numbers you can actually see
This stage is not a prompt, it is the discipline of pasting real data into one.
Here is what I found by hand. Marketplace best-seller rank in my category: [data]. Number of results for my main keyword: [data]. Review count on the three closest competitors: [data]. Supplier cost per unit at quantity [n]: [data]. My estimated ad cost per click: [data]. Work only from these numbers. Do not add any figure I have not given you. 1. Compute, showing every step, the maximum landed cost per unit I can afford at my target selling price and margin, including shipping, packaging, fees and ad cost per sale. 2. Tell me what unit economics look like at three selling prices: my target, 10 percent lower, and 10 percent higher. 3. From the review counts you were given, describe what a low review count on a close competitor would and would not tell me, and what it cannot tell me. 4. List the two numbers I have not supplied that most change the answer, and say where I would look for each. If a calculation needs an input I have not given, write [CONFIRM: input needed] and stop rather than assuming a value.
The review-count question is the one that quietly matters most. A handful of reviews on a close competitor can mean a new entrant, a listing that failed, or a product that sells and has no reviews. The data cannot distinguish them, and a model asked directly will invent a conclusion.
Stage three: order small, then write the test
Most product failures are discovered after purchase, not before. The cheapest test is one unit or a small batch with a listing live long enough to see whether anyone clicks.
Write a pre-order test plan for [product] at a landed cost of [cost] and a target price of [price]. 1. State the maximum units I should order before any demand signal, and the reasoning from the numbers I gave you. 2. List the signals that would justify a second order, and the ones that would justify stopping. 3. Draft a short supplier message asking for the smallest available quantity, the sample cost, and the current restock lead time. 4. Draft a listing title and five bullet points that lead with the buyer problem, not the product features. No unverifiable claims about quality or durability. Mark anything I must confirm with [CONFIRM]. Do not promise delivery dates, sales, or that the product will sell. Do not invent supplier names, prices or lead times.
Run the order small, review the physical sample yourself, and write the listing after. Reversing that order, listing before the sample arrives, is how sellers end up describing a product they have never held.
What AI cannot tell you, stated plainly
Keep this list visible, because it is what separates research from guessing dressed as research:
- Whether people are actually searching for this right now in your market.
- Whether a supplier's quoted cost holds at your quantity, or after shipping.
- Real return rates and the reasons behind them.
- Whether your ad cost assumption survives contact with bidding.
- Whether your closest competitor is profitable or just visible.
- Anything about trademark, patent or product liability.
For the tooling that covers those numeric gaps, use a proper research tool rather than a prompt. The prompts above handle the reasoning around the numbers. They do not replace them.
Common product research prompt mistakes
- Asking "will this sell?" The model answers yes, because your idea as written is always coherent.
- Accepting market size figures. If the model did not give a source, the figure is decoration.
- Treating a competitor review count as proof. Low reviews can mean anything, including a failed listing.
- Researching before naming the buyer. Product research without a named buyer produces a list, not a decision.
- Ordering before the sample. One sample order costs less than the first hundred units of a wrong product.
- Never revisiting. A niche that failed last month is the same niche in a different season, and the second look is cheaper.
Frequently asked questions
Can ChatGPT tell me which products will sell?
No. It can structure your thinking, list failure modes, and help you plan research, but it cannot see live sales, ad costs or return rates in your market. Anyone presenting a model output as a winning product list is describing a guess.
What is the cheapest way to test a product idea?
Order the smallest quantity a supplier will sell, or a single sample, with a listing live before you commit. The signal that matters is whether anyone buys at your target price, not whether you can raise the margin in a spreadsheet.
Should I use a tool or a prompt for product research?
Both, in sequence. A prompt to frame the question and list what could go wrong, a tool to get real numbers, then a prompt to work the numbers and write the test plan.
How many products should I research at once?
One or two. Researching twenty ideas at once produces twenty lists. A single idea taken through the three stages above is worth more than twenty that stop at stage one.
Take one idea through all three stages
Pick the idea you are most tempted by, write the claim, and gather the five real numbers by hand before you ask a model to work them. Then decide on the smallest order you can justify. That sequence takes a day and it replaces most of the content you would otherwise watch. For the tools that supply the numbers, browse finding products with AI tools.