Public methodology

A small audit with strict evidence rules.

AgentRank tests how AI assistants answer buyer selection questions, then maps the observed gaps to public product facts, source evidence, and repair actions.

01

Audit scope

The early audit is deliberately narrow: one product, one category, 3-5 competitors, five buyer-intent prompts, and up to four available AI surfaces.

Prompt typeWhat it tests
Open discoveryWhether the product appears before the buyer names it.
Feature-specificWhether the product appears when the buyer describes real requirements.
Named comparisonHow the assistant positions the product against direct competitors.
Alternative searchWhether the product appears as an alternative to a default winner.
Final choiceWhich product is chosen, and which risks shape the choice.
02

Surface disclosure

Every AI surface is recorded with access status. Blocked, login-gated, CAPTCHA-gated, or unstable surfaces are disclosed instead of substituted with invented outputs.

Browser output, consumer AI output, and API output are different evidence types. We do not mix them without saying so.
03

Measurement fields

  • Target appeared or not.
  • Target rank or recommendation position.
  • Winner or default competitor.
  • Competitors named in the answer.
  • Description accuracy and risk language.
  • Visible source or citation behavior where available.
  • Commercial signal strength.
04

Scorecard

The score is a triage tool, not a prediction model. It helps the reader see whether the current issue is visibility, description accuracy, evidence strength, competitive position, or fixability.

ComponentMaxMeaning
Visibility30Whether the product appears without being forced into the prompt.
Description accuracy25Whether category, features, limits, and risks are described correctly.
Evidence strength20Whether answers appear supported by useful public sources.
Competitive position15Whether the product is selected, ranked, or framed favorably.
Fixability10Whether the gap maps to specific public asset changes.
05

Repair roadmap

A finding is only useful if it maps to something the product team can change. Recommendations point to product facts pages, buyer use-case pages, comparison pages, FAQ cleanup, evidence registers, agent-readable assets, or source ecology work.

Weak advice says "add more content." Strong advice says which page to create, which facts to include, which concern to answer, and which prompt to retest.
06

Boundaries

  • No guaranteed AI ranking improvement.
  • No fake citations, fake reviews, fake customers, or spam links.
  • No bypassing login gates or human verification.
  • No claim that one dated test represents every user, model, country, and future answer.
  • No implementation work included unless a separate scope is approved.