How SellTo.us works
Last updated: October 10, 2026
Exactly how the AI-simulated participants are made, how their answers are collected and how every number in your report is calculated, for startup ideas, apps, services, products and ads. No black box.
The short version
- The participants are AI personas, not real people. Their answers show simulated reactions (purchase intent, interest, willingness to try or consider, or response to an ad, depending on the idea type), not real sales or sign-ups and not a forecast.
- An AI model plans the shopper groups for your audience. Our code then creates every shopper profile from those groups.
- Each shopper gets their own answer for every version: would buy, maybe or wouldn’t buy, plus a short reason.
- Percentages are simple counts: answers divided by the shoppers who answered. We don’t add statistical confidence or accuracy claims, because we haven’t measured the results against real sales.
- Use it as a fast first check, then confirm with real customers. Every report ends with practical steps to do that.
Why not just ask a chatbot?
A chat gives you one AI opinion. A SellTo.us test is structured: our code creates a fixed crowd of individual profiles for your audience, every profile answers each version, and the numbers are counts of those answers, checked by code before you see them. Every version and image goes to the same simulated people, results are split by audience group, and you can re-test an improved version on the very same simulated people (same test size). The report keeps evidence labels (“Simulated result” or “AI hypothesis”) and comes as a PDF. It is still AI imagining how people might react, so the limits below apply.
Idea types
When you set up a test you choose what kind of idea it is. The steps below are the same for every type; what changes is who the simulated people are, what the three answers mean and what the report concentrates on.
| Idea type | Simulated people | The three answers | Reported as |
|---|---|---|---|
| Startup idea | Potential customers | Interested · Maybe · Not interested | Simulated concept interest |
| App or software | Potential users | Would try · Maybe · Wouldn’t try | Simulated willingness to try |
| Service | Potential clients | Would consider · Maybe · Wouldn’t consider | Simulated willingness to consider |
| Physical product | Shoppers | Would buy · Maybe · Wouldn’t buy | Simulated purchase intent |
| Advertising idea | People in the audience | Would respond · Maybe · Would ignore | Simulated message response |
| Something else | People | Interested · Maybe · Not interested | Simulated interest |
If you don’t choose a type, “Something else” is used. Tests made before idea types existed are physical-product tests. In the rest of this page we use the physical-product words (shoppers, would buy); for other types, read them as the words in the table.
1. Planning the crowd
When you start a test, an AI model (Claude Sonnet 5.5, by Anthropic) reads your audience settings and your product, then splits that audience into shopper groups. Each group gets:
- a name and its share of the audience (for example “Busy parents 30–45, 12%”),
- an age range,
- 4 typical jobs or life roles,
- a budget level (tight, average, comfortable or high),
- a short note on how they shop.
The number of groups depends on the test size: 8, 25 or 100. The same step also checks the product against our content rules and stops tests for things like weapons or illegal drugs.
Photos: if you add photos, the planning model looks at each one once and writes a short description of what a shopper would see. The shoppers react to those descriptions, not to the image files themselves.
2. Creating the shoppers
The individual shoppers are created by our code, not by the AI. For a 10,000-participant test, the code makes exactly 10,000 profiles:
- each shopper is placed in a group in proportion to that group’s share,
- their age is picked at random inside the group’s age range,
- their job is picked at random from the group’s 4 jobs,
- in 100,000-participant tests, each shopper also gets a city or area from a list for your country (weighted by size where the list has that information).
The random choices are seeded from your test’s ID, so the same crowd can be rebuilt exactly when your report is made. That’s why we don’t need to store 100,000 profiles.
Be aware: shoppers in the same group share the same budget level and shopping note. Within a group they differ by age, job and, in the biggest test, city.
3. What “each shopper answers” means technically
The shoppers are sent to a second AI model, Claude Haiku 4.5 (also by Anthropic), in requests of 40 shoppers at a time. Each request contains:
- your audience and every version you’re testing (text, price and photo descriptions),
- a numbered list of 40 shopper profiles (age, job, budget, shopping note and, for the biggest test, city).
The model returns one answer per shopper, in order. Each answer has a verdict for every version (would buy, maybe or wouldn’t buy), a reason under 12 words and, when you add 2 or more photos, the photo that would most make that person want to buy.
So every shopper has their own profile and their own answer, but they are not 40 separate AI conversations: the model judges the 40 profiles in one go. Requests run through Anthropic’s Message Batches API, which processes them in the background.
| Test size | Shopper groups | Requests of 40 | Simulated interviews shown | Versions | Photos |
|---|---|---|---|---|---|
| 1,000 | 8 | 25 | 12 | 1 | 1 |
| 10,000 | 25 | 250 | 50 | up to 3 | up to 3 |
| 100,000 | 100 | 2,500 | 200 | up to 5 | up to 5 |
4. How the numbers are calculated
- Answered: answers that come back in the wrong format, or don’t come back at all, are skipped, never guessed. Your report shows how many shoppers actually answered (for example “9,880 of 10,000”).
- Would buy, maybe, no: for each version, the number of shoppers who gave that answer divided by the number who answered.
- Shopper groups and cities: the same count, inside each group or city.
- Photo vote: the share of answering shoppers who picked each photo.
- Crowd picture: up to 1,000 dots, an evenly spaced sample of the shoppers who answered.
- Simulated interviews: a random sample of answering shoppers and their reasons. The first names are made up by our code.
- If too few answer: if fewer than half of the shoppers answer, the test is marked as failed and your credit comes back automatically.
5. How the report is written
Claude Sonnet 5.5 receives the totals, the results for each group, the photo votes and a random sample of 160 shopper reasons. From those it writes the verdict, what shoppers love, why they don’t buy, suggested fixes, risks, possible alternatives and a SWOT. In 10,000 and 100,000-participant tests it also writes the ad targeting guide from the group results.
Every report is then checked by code before you see it. The key findings, all percentages and counts, the steps to check results with real customers and the example ad budget are produced from one validated set of numbers, so the same figure is identical on every page. Text written by the AI is checked sentence by sentence: numbers that don’t match the results, claims that one change caused a result when the versions differed in several ways, new product claims (certifications, “lab-tested”, health, safety or performance guarantees), competitor prices, market-size figures and unsourced citations are removed or replaced, and quotes whose words don’t match the shopper’s answer are left out. Each correction is counted in the report’s quality check.
Every page is labeled: Simulated result (counted from the simulated answers), AI hypothesis (ideas written from those answers, to test), Your input (what you entered, including the claims in it, which still need your own evidence) and Verified source (an outside fact with a named, dated source, shown only when one exists).
The best price is the AI’s judgment based on the answers. It is not calculated from a price curve. If you test the same product at different prices as separate versions, the comparison table shows the simulated answers for each price.
6. Comparing versions and photos
All versions and photos in a test go to the same shoppers, in the same requests, with the same audience settings. Every shopper answers for every version. That makes the comparison like-for-like inside the test.
It does not make it accurate in the real world. The AI may also be influenced by the order versions are listed in. Running the same test again gives somewhat different numbers, and we don’t calculate margins of error, so treat small gaps between versions as a tie.
Re-tests: before and after
Every test has a code like ST-7K4Q2X, shown on the report and the PDF. When you change your product and press Re-test (or link the old test by its code), the new test works differently from a fresh test:
- Your old version comes back as Version A, next to your updated version (or up to 4 updated versions). The old main photo goes with Version A, your new photo (if any) with the update.
- The crowd uses the same shopper groups as the first test, rebuilt from the same starting point. With the same crowd size you get the very same simulated shoppers again.
- The versions are shown in a rotating order (A then B, B then A, and so on), and shoppers are not told which one is old or new, so neither gets an advantage from its position or from being “the new one”.
- We compare shopper by shopper: how many switched to buying the new version, and how many switched away. The result says “higher” or “lower” only when the gap in would-buy answers is larger than a fixed threshold: at least 2 percentage points, raised when many shoppers switched in both directions. Otherwise it says “no clear difference”. This is a decision rule applied to one run, not a measured run-to-run variation or a statistical confidence level.
- The report then lists what you changed and, for each earlier suggestion, whether it was applied and whether it helped, based on the numbers and the shoppers’ reasons. It is allowed to say a change didn’t help or made things worse.
The side-by-side inside the re-test is the fair comparison. Don’t compare the re-test’s numbers directly with the first test’s: separate runs can give different numbers even for the same product, and we haven’t measured by how much. And like everything here, a better result is a better result among AI shoppers, not proof of real sales.
7. Limits you should know
- AI personas don’t spend real money, don’t see your actual website or competitors, and can be more or less generous than real people.
- Results depend on how you describe the product and the audience.
- Results are not validated against real sales, and no accuracy number is claimed.
- The AI can make mistakes, misread a product or repeat clichés.
- Runs aren’t perfectly repeatable: testing the same idea again can give somewhat different numbers, so treat small gaps as a tie.
- One AI request judges a batch of 40 profiles, and shoppers in the same group share a budget and shopping note, so answers within a group can be similar.
- The AI may be slightly swayed by the order in which versions are shown.
- Results can’t predict real sales, sign-ups, conversion rates, revenue or whether a business will succeed.
8. Timing and failures
Each test is scheduled to take about 10 to 12 minutes for 1,000 participants, 14 to 16 for 10,000 and 22 to 28 for 100,000. When the AI work finishes sooner, the finished report waits until then, so you know when to expect it; the progress screen never shows more answers than have really come back. At very busy times batch processing can take longer, up to 24 hours. Planning and the report are retried automatically if they fail. If a test can’t finish, it is marked as failed and your credit comes back automatically, so you’re never charged for a test that didn’t run.
9. The AI models and your data
We use Claude Haiku 4.5 for the shoppers and Claude Sonnet 5.5 for planning, reports and the ad guide, both made by Anthropic. If we change models, we’ll update this page. Anthropic receives the content of your test but not your email address, and under its commercial terms doesn’t use it to train its models. See our Privacy Policy.