Methodology and AI results disclaimer

What an audit measures

• Brand mentions and explicit recommendations.
• Observed answer position, top-three and top-five placement where a ranked structure can be detected.
• Citations, normalized source domains and whether a source supports brand or competitor evidence.
• Competitor visibility and prompt-level performance.
• Historical change only between compatible benchmark versions.

Provider execution and surfaces

The default monitoring adapters call OpenAI, Anthropic, Perplexity and Google Gemini provider APIs. Provider availability depends on plan entitlement and production configuration. A provider API answer is not equivalent to a signed-in ChatGPT, Claude, Gemini or Perplexity consumer experience. Google AI Overviews, AI Mode and Copilot are separate surfaces without enabled default production adapters. Each provider run records requested and returned model metadata when available, response identifiers, raw answer text, citations, token usage, latency, errors and estimated cost when pricing coefficients are configured.

Mentions, recommendations, citations and competitors

A mention records a detected reference to the monitored brand. A recommendation requires additional evidence that the answer endorses or proposes the brand, rather than merely naming it. Citation extraction records returned source URLs when available and classifies first-party citations separately. Competitor observations use entity matching and contextual classification; ambiguous names can require manual review. Detection is inspectable but heuristic and can miss or misclassify a response.

Prompt design and repeatability

A benchmark preserves prompt text, provider and model settings, language, market, scoring version and repetition count. Repeated runs can reduce reliance on one variable answer, but do not make the result deterministic. For an internal brand authority panel, unbranded category and problem questions belong in organic discovery; questions naming MonitorMyGEO belong in a separate brand accuracy panel. These panels must not be combined into a single organic visibility score.

Search and Agent Readiness

AI Readiness evaluates technical and static signals that can affect Search Readiness and Agent Readiness. Findings are versioned and retained with issue history. Browser-runtime WebMCP inspection is not claimed in this release; static discovery signals are reported only when they can actually be verified.

Scoring and comparison

Customer scoring weights and benchmark configuration are copied onto each audit. Historical comparisons require the same project and compatible benchmark identity, including prompt set, provider/model configuration, locale, scoring version and repetition count. Failed runs are not counted as successful responses or treated as brand absences. The internal 100-prompt authority panel is versioned separately from the frozen Day-0 dogfood benchmark and does not change customer scoring. Provider-specific valid coverage should accompany any aggregate.

Important limitations

• LLM outputs are probabilistic and can vary between repeated runs.
• Results may differ by time, model version, location, language, provider search availability and other provider-controlled factors.
• Entity, position, sentiment and citation extraction is deterministic and inspectable in this version, but heuristic.
• A detected correlation after an implementation does not prove that the implementation caused the change.
• Audit results do not guarantee future visibility, ranking, traffic, revenue or placement.

Website crawling

MonitorMyGEO fetches public HTTP(S) pages for project context, readiness checks and optional change tracking. Requests follow a limited number of redirects, apply timeouts, and reject local, loopback, link-local and private network destinations. It does not create an archival copy of every page.

Questions

For methodology questions or to report an extraction issue, contact support@monitormygeo.com and include the audit identifier. Do not email provider credentials or sensitive raw data.