Pressure-test selected ideas

Discovery ends with a ranked set of candidate ideas. From that set, you can choose one to three exact candidates for a single Deep Research run. The run investigates the saved scope in more depth, compares the selected ideas, and produces a final recommendation. A run started directly from the Idea Catalog begins with one seeded idea.

Before you start

Only the shortlist is required. The selection workspace keeps the path short:

1. Choose ideas. Save one to three candidates.
2. Compare trade-offs. Optional. Read the Discovery evidence side by side. You can also add private build limits and check fit without changing the research scores.
3. Check the evidence. Optional. Re-read the sources already captured for one decision-changing risk. This does not search for new evidence.
4. Review and start. Check the exact candidates, current cost, balance, and balance after starting.

Nothing is charged while you choose, compare, or review. The confirmation screen shows the current price before you start; one confirmation covers the selected scope.

Build limits, questions to resolve, and test plans are optional supporting work. Planning a test is a contextual follow-up to an unresolved question, not another required step. If none of the ranked ideas fit, branching a different direction is an escape hatch: the branch starts unevaluated, and the original candidate and its scores stay unchanged.

Why it's a separate phase

Discovery is broad by design: it looks across a niche to find recurring problems, audiences, and plausible products. (See How the research works.) Deep Research narrows that breadth to the candidates you chose and spends more effort on the questions that can reverse the decision.

What it checks

• Who the buyer is. The report narrows the audience and the job the product would serve.
• Competition and substitutes. It looks for direct products, adjacent products, free alternatives, and do-it-yourself routes.
• Distribution and search demand. It checks whether the proposed acquisition surface is plausible, not just whether broad keywords have volume. Each keyword is graded on whether it is genuinely about the idea before its volume counts toward demand, and the keyword count you see is that graded set. If none pass, the report calls demand unmeasured rather than borrowing volume from the category. In narrow or technical markets the graded volume is often too thin to separate the candidates at all, and demand then carries no weight in the comparison. (See How scoring works.)
• Pricing signals. It uses captured commercial-intent evidence and competitor pricing where available. It does not treat those signals as measured willingness to pay.
• Market size and timing. It separates the initial serviceable slice from a broader market ceiling and labels estimates as estimates. Both are built from the keywords that survived grading, so a market size here can come out well below the size of the niche around it.
• Data access and build feasibility. It checks whether required inputs appear obtainable and whether the implementation assumptions hold up.

The recommendation

The selected candidates continue through the validation stages as one saved scope. The final report records a recommended winner, runner-up context where applicable, and a Go, Conditional, or No-Go verdict with its supporting reasons. That recommendation is research output, not a promise that the business will succeed.

The recommendation is not simply the top of the ranking. A candidate whose core premise the adversarial review could not confirm is marked premise unproven and is passed over for the recommendation, so the highest-scoring idea in a report is sometimes not the recommended one. Its scores still stand — they describe how well it would work if the premise holds — and it stays available to you. (See How scoring works.)

What it can't do

Deep Research still works from public discussion, search data, and point-in-time web evidence. It cannot observe private buyer behavior or predict the future. Growth, demand, and market-size figures are directional estimates with uncertainty. It also does not evaluate candidates you left outside the confirmed scope; changing the scope requires a new run.