Methodology
StartupBench evaluates the idea visible from public evidence—not revenue, traction, team quality, or private company data. The website is scored separately.
Each area is a shifted weighted geometric mean. This makes weak criteria matter instead of letting one exceptional criterion fully cancel them out.
i = 1, …, n·ri ∈ [0, 5]·Σwi = 1·G ∈ [0, 10]
Idea score
⅔ overallWebsite score
⅓ overallMethod 1.1 requires a decoded, usable screenshot for all three website judges. A failed capture or access challenge stops evaluation. Legacy results without verified visual evidence remain readable but are excluded from the board and recent activity. Area calculations use full precision; rounding happens at the final score.
ScrapeBadger retrieves public page content with JavaScript and anti-bot handling. Screenshots are captured separately. Incomplete content or unusable images stop evaluation; capture sources are recorded in the audit. External context comes from ScrapeBadger search results; snippets are not full-page verification. Research gaps are disclosed, not treated as proof of uniqueness.
How a score is produced
Evidence profile
The site is read as untrusted source material. Public competitor research adds context; unsupported assumptions are excluded.
Three judges per area
Three Luna judges per area use medium reasoning and the same anchored 0–5 rubric. Idea judges do not see visual-design signals.
Deterministic math
The median judge score is taken per criterion. The published weights and formula produce the final score.
Anchored ratings
Every criterion uses the same evidence scale. Missing evidence is uncertainty—not an invitation to invent facts.
Research basis
The structure draws on research about opportunity evaluation, multi-item measures, and composite indicators. It does not claim to predict company success. Human validation and reliability testing are still required, so the method remains beta.