Skip to content
Forrest Miller
← Research

Research system · build paused

AudienceValidation

The evidence gate that decides whether an attractive idea deserves code — and records when the answer is no.

Research only · no live product · no user metrics · build paused

41 → 30
certificate-generator scoreas of 2026-03-04revised after the full credential-platform category was mapped
38 → 28
SOW-generator scoreas of 2026-03-04revised after crowding and the private-output mismatch were counted
0
build-authorized winnersas of 2026-08-03the complete 200/20/5 gate has not run end to end

Source: the AudienceValidation research record — demand-first playbook, validation prompts and results, pipeline report, session record · consolidated 2026-08-03

The gate a complete cycle must pass

  1. 200 action-intent queries
  2. 20 manual SERP audits
  3. 5 evidence dossiers
  4. API feasibility
  5. smallest build
  6. explicit authorization

Current state: this gate has not yet been run end to end, so the build stays paused.

What happened

The problem

AI made implementation cheap enough that a polished prototype became the fastest way to get confidently wrong. AudienceValidation was supposed to move the hard work upstream: find a specific high-intent job, prove the existing solutions weak, estimate the traffic upside, and confirm the smallest useful test could ship in days.

Where the process failed

The first cycle did not follow its own method. Search-result counts, surface-level SERP observations, and LLM scores were treated as validation. A certificate generator scored 41/50 and implementation began before the professional credential-platform category had even been mapped.

The missing test

The playbook required a five-signal demand dossier: quantified pain across independent sources, multi-tool workaround complexity, third-party ecosystem formation, thread longevity, and vendor-confirmed gaps. When that test was finally applied, the attractive gap weakened or disappeared.

The decision register

Dated verdicts and the evidence behind them.

killed

Certificate generator

  • Organic pain was shallower than the result counts implied.

  • The category already held more than ten dedicated credential platforms.

  • Unique credential URLs, QR verification, and social sharing were already standard.

  • The started build was paused.

killed

SOW generator

  • AI SOW generation was crowded and increasingly table stakes.

  • The output is private, so it could not feed the public shareable-URL flywheel being tested.

open

Current winner

  • No replacement has completed the five-signal dossier and the explicit authorization gate.

The judgment

What changed

The strongest output was a refusal: build speed, sunk implementation, and an AI-generated score all lost to the evidence. Earlier analyses stay in the record. The started build stays paused. No candidate reaches code without inspectable evidence and explicit authorization.

What the process now requires

Restart at phase one. Generate 200 specific action-intent queries across real expertise areas, hand-audit the top 20 result pages, complete five demand dossiers, check API feasibility, and define the smallest possible build. Code resumes only after one candidate clears the original threshold with inspectable evidence and receives explicit authorization.

Why this belongs in the portfolio

The four live products show what happens after a bet earns code. AudienceValidation shows the judgment before that point: an honest research cycle can end with two kills and a stopped prototype.

Method: manual SERP audits · review-theme counts · workaround mapping · competitor-category mapping · forum longevity · vendor documentation · API feasibility · Claude Code