How Greg Isenberg uses Claude and Perplexity to find and vet startup ideas in real time
Why this story matters
Most AI tool stacks are built for shipping. Greg Isenberg's stack is built for thinking — specifically, for compressing the time between "I noticed something" and "I understand whether this is a real business."
He hosts the Startup Ideas Podcast, which has over 375,000 YouTube subscribers and has published more than 300 episodes. The format is simple: Greg and a guest (or sometimes just Greg) spot an underserved market, talk through the opportunity, and figure out whether someone should build it. What's changed in the past eighteen months is that Claude and Perplexity have become live participants in that process, pulling data and testing assumptions in real time while the cameras roll.
The result is a feedback loop that would have been impossible before: ideas get stress-tested with actual market data during the conversation that generates them, not in a separate research phase that happens days later when the energy has dissipated.
The problem with traditional idea validation
Greg has talked openly about the failure mode that plagued his earlier work: by the time you finished validating an idea, you'd lost enthusiasm for it. The research was too slow. You'd spot something interesting on a Monday, spend a week doing market research, and by the following Monday either you were onto something else or a competitor had shipped.
The other failure mode was confirmation bias. When validation takes effort, you unconsciously look for evidence that supports the idea rather than evidence that breaks it. You want the work to pay off.
AI research tools solve both problems. When Perplexity can pull market size estimates and recent news in thirty seconds, validation happens fast enough that the idea is still warm. And when Claude can argue against your thesis as easily as for it, the temptation to cherry-pick evidence weakens.
How the live research stack works
Greg's setup during recording is straightforward. He has two tabs open alongside his notes: Claude and Perplexity.
Perplexity handles fact retrieval. When the conversation produces a claim — "the market for this is probably $2 billion" — Perplexity verifies it or corrects it. When someone mentions a competitor, Perplexity pulls their pricing and recent funding. When there's a question about whether a regulation exists, Perplexity checks. The answers come back fast enough that they don't interrupt the flow of conversation.
Claude handles reasoning. Once there's a rough picture of the opportunity — the problem, the customer, the rough size — Claude's job is to find the hole. Greg describes it as having a skeptical partner in the room: "Tell me why this doesn't work." Claude's response usually generates the most useful part of the episode, because the counterarguments are specific and grounded rather than generic.
The Notion database collects everything that survives the live vetting. Each entry has the core thesis, the evidence found via Perplexity, Claude's main objection and how it was resolved, and a rough verdict: build now, wait for timing, or abandon.
The three questions Claude answers every time
Greg has converged on a consistent set of prompts. He's shared these across his newsletter and podcast:
1. "Who is losing sleep over this problem right now, and why?"
This forces specificity. "Small business owners" is not an answer. "Solo operators running service businesses under $500K revenue who are manually tracking invoices in spreadsheets" is an answer. The specificity matters because it determines where you find customers, how you price, and what the MVP looks like.
2. "What would have to be true for this business to reach $1M ARR?"
Claude works backwards: how many customers at what price, what's the realistic CAC at that scale, what does the retention curve need to look like. The goal is not to get precise numbers — it's to surface the assumptions. If the model requires 10,000 customers at $100/month, the question becomes whether there are 10,000 addressable customers in the niche. If the model requires a $5,000/year contract, the question becomes whether the buyer has that budget and authority.
3. "What is the most likely reason this fails in year two, not year one?"
Year-one failures are obvious: wrong market, no distribution, can't build it. Year-two failures are subtler: the market turns out to be a feature, not a product; a well-funded competitor enters; the customer segment is too small to grow beyond initial traction. Claude surfaces these with uncomfortable specificity when you ask it to.
From idea to Late Checkout portfolio company
The ideas that survive this process go into a second stage: the Late Checkout studio model. Greg's team does a more thorough pass — customer interviews, a landing page test, sometimes a paid acquisition experiment to validate willingness to pay — before deciding whether to build.
This stage is slower and doesn't rely on AI in the same way. But the AI-accelerated first filter means the ideas that reach the studio have already been stress-tested more rigorously than most ideas that reach the MVP stage at traditional companies.
He's described the combination as "first-draft thinking with Claude, confirmation with actual humans." The AI handles the intellectual work of killing bad ideas fast. The humans handle the emotional work of confirming that someone cares enough to pay.
What breaks the system
Greg has been candid about the failure modes:
Claude hallucinates market data. Perplexity is better for factual claims because it cites sources. Claude will confidently state market sizes that don't match any real research. The fix is to never use Claude for numbers — only for reasoning — and to verify any statistic via Perplexity before treating it as real.
Perplexity is bad at "what is the vibe." It can tell you a market is $4 billion and growing 12% per year. It cannot tell you whether the customers in that market are frustrated enough to switch tools. That requires reading forums, Reddit threads, and customer reviews — which is still manual work.
The speed creates overconfidence. When validation takes an afternoon instead of a week, there's a temptation to act on less signal. Greg's guard against this is the Notion database: ideas don't leave the "vetted" stage until they have at least three customer conversations behind them, regardless of how clean the AI analysis looks.
The practical framework
If you want to apply this to your own work, the minimal version is:
- State the idea in one sentence to Claude: what it does, who it's for, what existing behavior it replaces.
- Ask Claude to identify the single most likely reason it fails.
- Ask Perplexity for the size of the addressable market and the two or three existing competitors.
- Ask Claude whether the numbers Perplexity returned change its failure analysis.
- Write down the result in three sentences: thesis, main risk, what would have to be true for it to work.
If you can do that in twenty minutes and still want to build it, it's worth a landing page test.