If you are like me, you’ve probably asked ChatGPT to validate your startup idea.
The response is always some variation of: “This is a promising market, but you need to validate demand.”
It sounds smart, but it’s completely useless. It’s just an AI guessing based on its training data. It doesn’t tell you if real people are actually complaining about this problem today.
I got tired of this, so I built Ainexa — a startup idea validator that doesn't guess. It scans live data sources (Reddit, Hacker News, Product Hunt, etc.) and forces the AI to link every claim to a real, clickable source.
My Rule: If we can't find evidence for a claim, the AI is not allowed to make it up.
Over the last few months, I’ve been building this in public. I’ve run the first batch of 49 real ideas submitted by early users through the system.
Here is what the data actually surprised me with:
“AI Wrapper” ideas are getting slaughtered
Out of the 49 ideas I scanned, the ones focused on generic “AI + [Something]” had the highest competition scores and the lowest overall viability. The market signal was weak. It’s not that AI is dead, it’s that the “thin wrapper” era is over.Specificity equals demand
The ideas that scored the highest had extremely specific target users. For example, an idea targeting “freelance illustrators who invoice in multiple currencies” scored much higher than a generic “invoicing tool for small businesses.” The narrower the niche, the stronger the community signals we could find.The "Zero Competitor" trap
One user submitted an idea and proudly said, "There is zero competition for this."
Our system scanned the web and found 4 indirect competitors. But more importantly, it found zero demand signals. No one was complaining about this problem.
No competition usually means no market. You want competitors; it proves people are willing to pay.
The hardest lesson after 3 months
Building the scanning logic was hard. Making the AI perfectly cite its sources without hallucinating was even harder.
But the hardest part? Distribution.
I have 0 paying users. I have the tool, I have the data, but getting people to see it is a completely different game.
What’s next?
I am opening Ainexa to the first batch of early adopters on DEV. If you have an idea you are chewing on right now, let me help you verify it.
You can drop your idea in the comments below, and I will run it through Ainexa for free and reply with the actual data and sources it finds. No strings attached.
If you want to run your own scan, you can try it here: https://www.ainexa.xyz/
I’m building this in public. If you have feedback on the UI, the data, or just want to tell me I’m wrong, my inbox is open.
Top comments (2)
Two things here, and the second one is the part I'd argue with.
Distribution first - I hit the same wall this week. My own post: 48 hours, 0 comments, 0 reactions. Five comments I left on other people's posts: four authors replied. Same person, same effort, opposite outcomes. Publishing isn't distribution; showing up inside a thread that already has readers is. Your 49 ideas came from people who found you - that part is solved, even if it doesn't feel like it.
Now the push: 49 points is enough to be interesting and not enough to say "the thin wrapper era is over". Every one of those 49 was submitted by someone willing to paste an unvalidated idea into a free tool they found on dev.to. That's a self-selected sample, and it over-represents people who read dev.to. My 10k run showed the same trap at scale: at n=1 the estimate was wrong by more than 20% four times out of ten, and the aggregate only looked sane because the errors canceled.
The honest version of your finding - "among 49 self-selected ideas, generic AI wrappers scored worst" - is still useful, and it won't embarrass you at n=500.
Thanks — I'll take both points.
On distribution, you're right, and your framing is more precise than mine was: publishing isn't distribution; showing up in threads that already have readers is. My numbers this week looked almost identical to yours — my own post went nowhere, but other people's threads led to real discussion. That asymmetry feels like a finding worth recording on its own. Thanks for sharing your data.
On the second point, I won't push back — you're right. The 49 points are a self-selected sample: they over-represent active dev.to readers, and the submitters are exactly the kind of people willing to paste unvalidated ideas into a free tool. My original wording didn't make that bias clear. Your n=1 simulation is convincing too: in four out of ten runs, single-point estimates were off by more than 20%, and the aggregate only looked reasonable because the errors canceled out. I'll remember that analogy.
So I'll narrow the claim to the version you suggested: "across 49 voluntarily submitted ideas, generic AI wrappers scored lowest" — a description of this sample, not an assertion that "the lightweight wrapper era is over." If I want to talk about trends, I need a sampling method that doesn't depend on voluntary submission, or at least an n large enough to report confidence intervals instead of telling stories with 49 data points.
If you're willing, I'd love to see the setup for your n=1 simulation — especially how you defined "more than 20% error." I'd like to run it myself.