How it works
What actually runs when you press Run
What the machine does between the moment you click Run and the moment you get a shortlist. No jargon, and no step you'd have to look up.
1What you get
At most 5 ideas, each with the documented problem behind it and a link you can open, a 0–100 score for what one person with your skills, time, budget and reach can actually sell and support, one first move you can do this week, and the point where you should walk away.
In the run we publish, 69 candidates came back as 5. A thin market comes back as zero — we would rather hand you nothing than manufacture a fifth idea to fill the page.
2What happens after you click Run
We plan where to look, splitting your market into the sub-markets worth searching separately. We search live public sources. We read the pages we're allowed to read — our reader honours each site's rules for automated visitors, and never follows a link the model invented rather than found. We drop every candidate with no documented problem behind it. We score what's left against your profile on a fixed checklist. Then we rank the top few and write each one up.
One market takes 10 to 40 minutes; several markets run one after another, so a wide search takes longer. You don't wait on the page — it updates live with what has been found so far, and we email you when it's done. A sweep of several markets sends one email at the end, not one per market.
3Where the evidence comes from
Every problem we cite carries a label saying how we know it. These are the five, and they are the same words you'll see on your own results:
- from a real page
- We opened the page ourselves and read all of it. The strongest thing we can offer.
- from search
- We saw the search result but not the whole page — usually because the site blocks automated readers.
- you provided this
- You gave us this link and we read it.
- you pasted this
- You pasted this text and we read it.
- not verified
- We could not find a source. This is our own inference, and it is labelled as one.
The last one is the important one. When we can't find a source, we say so rather than write something plausible — and an idea resting on unverified problems is scored accordingly.
4How the score is computed
The 0–100 is not the AI's opinion of the idea. The AI rates each area separately and supplies the evidence for its rating; the weights, the total and the build / test / drop decision are computed in code from a fixed checklist. It never sets its own total, so it cannot talk itself into a passing number.
Before any of that, a set of deal-breaker checks runs. An idea that fails one is dropped with the name of the check that dropped it, regardless of how well it scores elsewhere.These are the 10:
- 1Is there a real, documented problem behind it? — shown on a dropped idea as “No real problem behind it”.
- 2Is there a path to money that repeats? — shown on a dropped idea as “No clear way to make money”.
- 3Is there a repeatable way to reach the buyer? — shown on a dropped idea as “No repeatable way to reach buyers”.
- 4Can it charge enough to be worth one person's time? — shown on a dropped idea as “Can't charge enough”.
- 5Can it run without hand-holding every customer? — shown on a dropped idea as “Too much hands-on support”.
- 6Does it survive if one platform changes its rules? — shown on a dropped idea as “Leans too hard on one platform”.
- 7Is there a product left if an AI vendor ships this feature? — shown on a dropped idea as “Just an AI wrapper — no real product”.
- 8Can you, specifically, sell and operate it? — shown on a dropped idea as “Hard for a solo founder to sell or run”.
- 9Can it launch without licences or compliance audits? — shown on a dropped idea as “Needs heavy compliance (licenses, certifications)”.
- 10Can it sell without legal review and a sales team? — shown on a dropped idea as “Needs enterprise sales and legal review”.
Failing one is enough. A candidate that clears nine and fails the tenth is dropped, and the run tells you which one — that is what the rejection list on a result page is.
An idea you bring us yourself goes through a different list — 12 checks rather than 10. It drops the first one above (there is nothing to source a documented problem from when you hand us the idea), splits “is there a path to money that repeats?” into two, and adds two more the sourced path doesn't ask:
- 1Can it make money?
- 2Can it charge on repeat?
- 3Can it charge enough?
- 4Can you reach buyers?
- 5Low hands-on support?
- 6Safe from one platform's control?
- 7A real product, not just an AI wrapper?
- 8Can you sell and run it?
- 9No heavy compliance blocker?
- 10Sells without 'book a call'?
- 11Pays for itself within a year?
- 12Can it be sold or handed off?
You can see all 12 applied, one by one, on the published sample report.
What that does not mean is that two runs of the same market return the same list. The ratings come from a language model and they move; a thin market can come back with four ideas one day and none the next. What stays fixed is the standard they are measured against, and the fact that the model never marks its own paper.
We don't claim to be accurate, and neither can anyone else: there is no way to check whether an idea nobody built would have sold. What we claim is that we show our work and we're willing to hand back nothing.
5What it costs, and what the free credits cover
A Discovery run costs 2 credits for the first market and 1 more for each additional market, with no ceiling — every market is searched separately, so the price follows the work.
The 2 credits in every new account cover a founder read and one idea check — real runs, full output, no card. They do not cover a Discovery run. If you want to see one before paying for one, the published run below is a complete one, and it needs no account.
6Why not just ask a chatbot
You can prompt a chatbot to be harsh. You can't prompt it to go and look. A chat answers from what it already knows, so it produces plausible competitors and plausible complaints; the published run made 60 searches, read 24 pages, and carries 41 sources you can open.
It also throws almost everything away: one run considers up to 150 candidates and keeps at most 5. On the published run, 36 of 69 were killed and each one says which check killed it. A chat gives you back about as many ideas as you asked for, and starts from zero the next time you open it — where each run here checks against the problems every earlier run found.