Treat AI as a research assistant for four cheap tests, not as a judge: scan competitors, script and summarise ten customer conversations, put up a one-page offer with a sign-up or deposit button, and check your price against what rivals charge. Spend nothing on build, stock or branding until real people have acted on the offer.
The common mistake is asking a chatbot "is this a good idea?" and getting a cheerful yes. A language model has no view of your buyers, so the answer is a polite echo of your own prompt. What it does well is the grind around the evidence: finding rivals, drafting neutral questions, sorting transcripts and spotting what you assumed without checking.
What counts as proof and what does not
Before opening any tool, decide what would change your mind. An idea is validated when strangers in a defined group take a costly action: they give an email address, book a call that takes time, pay a deposit or sign a letter of intent. Compliments, likes and "sounds interesting" are free, so they tell you almost nothing.
| Signal | Cost to the person | How much weight to give it |
|---|---|---|
| "Great idea" in conversation | None | Almost none |
| Joined a waiting list from an advert | An email address | Low, but countable |
| Booked a 30-minute interview | Half an hour | Medium |
| Described the problem unprompted, in detail | Honest effort | Medium to high |
| Paid a refundable deposit or pre-ordered | Money | High |
| AI summary says "strong demand" | None | None; it is a model's guess |
Write your pass mark before you start. For instance: "Of 200 people who see the page, at least 8 leave an email, and of ten interviewees at least four describe the problem without my prompting." Numbers like these stop you moving the goalposts after you have grown fond of the idea.
A running example: a mobile bike-repair service
To keep the steps concrete, imagine a founder who wants to start a repair service that visits flats and offices to fix commuter bikes. Nothing here is a real business; the figures are illustrative. The founder has a van-sized budget in mind (several thousand dollars) and wants to know whether to spend it before a single customer has said yes.
The pass mark they set: ten interviews with commuters who own a bike, at least four of whom have skipped a repair in the last year because of the hassle; a landing page that collects 8 enquiries from about 200 visitors; and a price that sits inside what rivals already charge for a basic service.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowStep 1: scan the competitors without trusting the summary
Ask an AI tool with live web search to list who already serves your group, then check every name yourself. The task is finding the rivals you did not know about, including indirect ones: the customer's current workaround is a competitor too.
I'm testing a mobile bike-repair service for commuters who live in flats. List direct competitors (other mobile repairers), indirect ones (bike shops that offer collection, DIY kits, repair cafes) and "do nothing" alternatives. For each, give the name, what they charge for a basic service if published, and the URL where you found it. If you cannot find a source, write "no source" instead of guessing.A sample reply (illustrative) might return six names, four with a source and two marked "no source". The two unsourced ones are the ones to investigate by hand or drop. Models invent plausible companies, so the rule is simple: if there is no URL that opens to a real business page, it does not go in the table.
Next, paste the real pages you collected (pricing pages, review excerpts) into one chat and ask for a comparison table: what each offers, what it charges, what customers complain about. Asking the model to work only from pasted text keeps it from filling gaps with guesses. What you are hunting for is a gap that customers repeatedly mention, such as "they only come on weekdays" or "no quote until they arrive".
A quick sum on market size
Resist the large, round "market" figures that AI tools offer. Build a bottom-up sum you can defend instead. If the founder serves one commuter district with 6,000 bike owners, and 5% might pay for a visit in a year, that is 300 customers. At an illustrative $60 average visit and two visits a year, revenue is 300 × $60 × 2 = $36,000. Whether that supports a van, tools and a wage is a sums question the founder can answer in ten minutes, and far more useful than a headline figure with no source.
Step 2: use search-interest data for direction only
Google Trends shows how interest in a search term moves over time as a relative score from 0 to 100, not a count of searches: Google divides each data point by all searches in that place and period, then scales the result. That makes it good for seasonality and for comparing two phrasings, and poor for sizing a market. Its help pages say the current explore view compares up to eight groups of terms at once, and terms with very little search volume show as zero, which is not the same as nobody wanting the thing.
A founder might compare "bike repair at home", "mobile bike mechanic" and "bike service at work" (the phrasing people use varies, so test several). If one phrase dwarfs the rest, use that wording on your landing page. If all three are flat and tiny, that is a prompt to ask why, not a reason to quit: some buyers never search and come through word of mouth or a landlord.
Paid keyword tools and AI market-sizing tools produce volume estimates that vary widely between vendors, because each models search data differently. Use them to rank phrases against each other, and say "estimate" when you write the number down.
Step 3: write interview scripts that do not lead the witness
This is where AI earns its keep. A good script asks about the past, not the future: what happened the last time a bike broke, what it cost, what they did instead. People are poor at predicting whether they would buy a product, but decent at describing what they actually did.
Draft a 25-minute interview guide for commuters who own a bike. Rules: ask only about the last time something went wrong with the bike; no questions that mention my service or ask "would you use"; include follow-ups for cost, time lost and who they asked for help; end with "who else should I talk to?". Mark any question that could lead the interviewee.Compare a leading question with a clean one. Leading: "Would you pay $40 for someone to fix your bike at your door?" Nearly everyone says yes. Clean: "Tell me about the last time your bike stopped working. What did you do next?" The second gets the real story, including the detail that the repair took three weeks because the shop was across town.
Finding people to talk to
Ask AI to draft three versions of a short, honest invitation, then send them to people in the group through the places they already gather: a residents' group, a cycling club, a workplace bike-share channel. Say plainly that you are researching and not selling, and offer a small thank-you such as a coffee voucher. Here is an invitation worth sending (illustrative):
Hi [first name], I'm researching how people keep commuter bikes running and I'm not selling anything. Could I ask you 20 minutes of questions about the last time yours needed fixing? Happy to send a $10 coffee voucher as thanks. [your name]Summarising the calls
After each interview, drop the transcript or your notes into your chosen assistant and ask for structured extraction, not opinions:
From these notes, list: (1) the problem in the person's own words, quoted exactly; (2) what they did about it; (3) what it cost in money and time; (4) anything they said they would not pay for. Do not infer anything that is not in the notes. If a field is missing, write "not mentioned".A realistic output for one interview (illustrative): "Problem: 'I couldn't face carrying it on the train to the shop.' Did: left it chained for six weeks. Cost: about two lost commutes a week, no money spent. Would not pay for: collection if it costs more than the repair." Put all ten into one sheet, one row per person, then count how many described the same problem. That count, not the model's mood, is your result.
Before you upload anything, check what the tool does with your content. Interview notes can contain names and workplace details, so remove them or read the tool's data-use settings first; this tutorial on customer data in chatbots covers what to check.
Step 4: a one-page test with a real call to action
A landing page is a small, honest advertisement for something that may not exist yet. An AI assistant can draft the headline, the three benefit lines and the form, and a page builder or a plain form tool can publish it in an afternoon. The test is whether strangers act.
The page needs one promise, one price signal and one action. For the bike founder: "We fix your commuter bike at your door in one visit. Basic service from $55. Join the first-visit list." The action is the email field, or, for a stronger test, a $10 refundable deposit that holds a slot.
Honesty matters here. Say that the service is launching soon, and never invent testimonials, customer counts or press mentions to make the page look established. If you collect deposits, make the refund terms plain and honour them quickly. A test that tricks people also poisons the list you wanted to build.
Sending the right traffic
Two hundred visitors from the wrong group prove nothing. Send people from the group you interviewed: a small paid advert aimed narrowly, a post in the club you joined, a flyer in the building lobby with a short link. Add a tag to each source's link so you can see which channel brought the sign-ups. If the page collects 8 of 200, you hit the mark set earlier; 1 of 200 is a result too, and a cheap one.
Comparing two headlines
Ask the assistant for three headline options aimed at different problems the interviews surfaced, for example "No more carrying it to the shop", "Back on your bike by Friday" and "Fixed while you work". Run each for the same budget on a separate version of the page. With 200 visitors the differences are rough, so only act on a gap that is large, and run it again before you rebuild the business around one line.
Step 5: check the price before you commit to it
Price is the part most founders guess. Combine three sources: what rivals publish, what interviewees said they paid last time, and what your costs plus a wage require. Then test the number on the page.
A simple cost sum (illustrative): a basic service takes 45 minutes on site plus 30 minutes travel, so 75 minutes of labour. At a target wage of $28 an hour that is $35. Add $8 of consumables and parts, $6 for van and tool wear and $5 for booking and card fees, and the minimum viable price is about $54. If rivals charge $45 to $70 and your interviewees remember paying around $50, a launch price of $55 to $60 is defensible and the margin is thin, so the next question is whether you can do two visits in one trip.
Ask the AI to stress-test the sum, not to produce it: paste your assumptions and ask "which of these numbers is most likely to be wrong, and what would the price be if it is 20% worse?" You still own the arithmetic.
| Price test | What to do | What it tells you |
|---|---|---|
| Two prices, two page versions | Show $55 to half the visitors and $65 to the rest | Whether a higher price cuts sign-ups sharply |
| Ask at the end of each interview | "What did you pay last time and was it fair?" | Anchors in real spending |
| Deposit at different amounts | $5 versus $20 refundable | How serious the sign-ups are |
| Rival price table | Pasted pages, then a comparison | Where you sit against the field |
Where AI market data misleads
Four failure patterns turn up again and again, and each has a quick check.
- Invented sources. A model may cite a report that does not exist or attach a number to the wrong publisher. Open every link. If you cannot find the figure on the page, delete it from your notes.
- Old data presented as current. Prices and rivals change quickly. Check the date on any page the tool relied on, and prefer a vendor's own pricing page to a roundup.
- Survivorship bias. Public articles feature businesses that worked. The model rarely sees the ones that quietly closed, so every "success story" overstates your odds.
- Averages that hide your group. A national figure for "bike owners" says nothing about people who live in flats and commute by train. Narrow the group before you trust a number.
The same caution applies to AI-written customer personas and synthetic "survey" tools that simulate respondents. They are acceptable for brainstorming questions, but their answers are the model's patterns, not your market's. If a tool offers an instant survey of "500 respondents" with no real recruitment, it is not research. Our tutorial on checking AI quality, not just speed has a short routine for spot-checking any figure before it goes in a plan.
A mistake that cost a founder two weeks
Here is an illustrative failure. A founder asked a chatbot for "the top five competitors and their prices" and pasted the result straight into a pitch table. Two of the five had closed, one price was from a different product line, and one competitor never existed. The founder found out when an adviser asked for the sources. Rebuilding the table from real pages took longer than doing it properly would have. The fix was the rule above: every row needs a URL that opens, a date seen, and the price as written on the page.
An eight-day plan for the whole test
- Day 1: write the pass mark and the one-sentence customer group. Run the competitor prompt and open every link.
- Day 2: draft the interview guide and invitations; send twenty invitations to get ten replies.
- Days 3 to 5: run the interviews, extract each into the sheet the same evening while it is fresh.
- Day 5, evening: build the landing page with the wording customers used themselves. Check the form works on a phone.
- Days 6 to 7: send targeted traffic, watch sign-ups and add the price test.
- Day 8: compare results with the pass mark. Decide one of three things: go ahead, change the group or promise, or stop.
Total cost in this example: a few hours a day, perhaps $50 to $100 of advertising and thank-you vouchers, and whatever you already pay for an AI assistant. That is a small price for knowing whether the van is worth buying. For broader help choosing which ideas deserve a test, the AI use-case scoring template works well on business ideas as well as AI projects.
Reading the result honestly
Three outcomes are common. First, the sign-up rate clears your mark and interviewees describe the same pain: move to a paid pilot with a handful of customers before building anything big. Second, interviews are strong but the page converts badly: the problem is real and your promise or price is wrong, so rewrite the headline using their words and test again. Third, nobody describes the problem unprompted: the pain may be too small to pay for, which is the cheapest possible lesson.
Write down what you learned, even if you stop. A one-page note of the customer group, the problem in their words, what rivals charge and what the test showed saves you from repeating the work if you return to the idea next year. If a service business is closer to what you have in mind than a product, this guide to testing a new service idea with AI covers the differences, such as selling the first job by hand.
Keeping your research honest and private
Two habits keep the process clean. Keep a source log: every fact in your notes carries a URL and the date you looked. And keep personal data out of tools that train on inputs; check the tool's data setting and anonymise names before pasting. If you record calls, ask first and say how the recording will be used. If you would rather map the whole test with someone alongside you, that is what my AI implementation consultation covers; you can get in touch here. Otherwise everything above can be done alone, with the tools you already have.
Questions about validating an idea with AI
Can AI tell me whether my idea will succeed?
No. AI can organise what is already published and sharpen your questions, but it cannot observe your future customers. The only evidence that counts is behaviour: people booking a call, joining a waiting list, paying a deposit or pre-ordering. Treat AI as a research assistant that speeds up preparation, never as a verdict.
Are AI-generated customer personas worth using?
Use them only to draft interview questions or to spot assumptions you have not examined. A synthetic persona repeats patterns from its training data, so it cannot tell you what real buyers in your niche want. Never quote a persona's answers as if they were research, and never use them in place of five real conversations.
How many customer interviews are enough?
Aim for at least eight to ten in one narrow customer group before judging an idea. You are listening for the same problem, in similar words, coming up repeatedly. If every conversation describes a different problem, your group is too broad. There is no magic number; stop when new interviews stop surprising you.
Should I tell interviewees I use AI to summarise calls?
Yes. Ask permission before recording or transcribing, say which tool you use and who sees the transcript, and offer to skip the recording if they prefer. Consent rules for recording differ by place, so check the rules that apply to you and your interviewee before you press record.
Further reads
- How to Set Up Human Review for AI Work Without Slowing Down — Set up a light check on AI output before you rely on it.
- AI Startup Credits and Discounts a Small Firm Can Actually Claim — Cut tool costs while you test the idea.
- How to Pilot AI in Shadow Mode Before Customers See It — Run a trial quietly before anything reaches customers.
- How to Calculate AI ROI for Your Business (Worked Example) — Put numbers on whether the tools you picked pay back.
- How to Measure Customer Reaction After Introducing AI — Read customer response once something is live.
- How to Start a Business Using AI: Tools and Costs for a Small Shop — A costed one-shop starter plan: idea test, name, store, first customers and admin, with monthly tool budgets at three levels and the break-even orders.
- Startup AI Tools for Growth: 8 That Earn Their Subscription — Eight growth-focused AI tools judged by the job they do, ranked by payback, with a 14-day trial rule and a cancel list.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Google Trends help pages on how search interest is normalised and on comparing terms in Explore, checked November 2026.