Skillaeo

SkillAEO Methodology: What We Observe and How We Score It

The prompts, run conditions, evidence rules, failure handling, scoring, limitations, and version behind SkillAEO reports.
Aug 17, 2026

Methodology version: 2026-08-17.1 Last reviewed: August 17, 2026

SkillAEO reports two separate measures: Observed AI Visibility and Site Readiness. They are not combined into a universal score and neither predicts a future ranking or citation.

Generated monitoring prompts

Generated monitoring prompts are test questions created for an audit. They are not a record of all real user searches on the internet. Prompt type is preserved as either direct (the brand or domain is named) or discovery (the brand is not supplied and the question tests category discovery).

The system creates prompts from the submitted public domain and available topic context. Users should confirm that the prompt set reflects their audience, language, country, and business facts before interpreting results. A rerun should preserve the prompt set when the goal is a like-for-like comparison.

Run record

Each run records the exact prompt, prompt type, repeat index, provider, model, product surface, search mode, locale, country, language, start and completion times, raw answer, parsed result, cited URLs, error state, and evaluator version. Current plan settings use one run per prompt. Repeat consistency is shown only when a prompt has enough comparable repeated observations.

The API models are called through OpenRouter. A provider model accessed through an API is not presented as identical to a named consumer chat product. Web search is recorded as enabled, disabled, or unknown; SkillAEO does not imply a separate search tool when none was used.

Valid observations and definitions

  • Mention: the valid answer names the audited brand or domain.
  • Recommendation: the valid answer positively includes the brand among options for the prompt.
  • First choice: the valid answer presents the brand as the first or primary recommendation.
  • Owned citation: at least one cited source belongs to the audited domain.
  • Claim accuracy: a reviewable claim in the answer agrees with the public evidence available to the evaluator.
  • Repeat consistency: comparable repeated runs produce the same classified outcome.

A valid answer that does not contain the brand is not mentioned. A provider failure, timeout, rejected request, or parsing failure is unavailable, not “not mentioned.”

Observed AI Visibility calculation

Only valid observations enter a denominator. Each displayed metric shows its numerator and denominator. Available metric rates are combined using these weights: mention 30%, recommendation 25%, first choice 10%, owned citation 15%, claim accuracy 10%, and repeat consistency 10%. Missing metrics are omitted and remaining weights are normalized. Provider or parsing failures are shown separately.

Citation share is owned citations divided by all citations in valid answers. It is not a market-wide share-of-voice estimate.

Site Readiness calculation

Site Readiness uses direct public-page checks and documented heuristics. Passing checks score 100, warnings 50, and failures 0. Erroring checks are unavailable and excluded. Available categories are combined using crawlability 20%, factual clarity 25%, support coverage 20%, structured data 10%, architecture 10%, and trust consistency 15%; missing categories are omitted and remaining weights are normalized.

llms.txt and agent.json checks are experimental and excluded from Site Readiness. Readiness is not evidence that an AI system crawled, indexed, cited, or ranked a page.

Direct observations and heuristics

Raw answers, citations, response states, response codes, public metadata, canonical URLs, robots rules, and visible structured data are direct observations for that run. Content clarity, support coverage, likely usefulness, and prioritization are heuristics or editorial judgments. Recommendations show an evidence badge: Official, Observed, Established web practice, Experimental, or Editorial judgment.

Randomness, changes, and limitations

Generated answers can vary with prompt wording, model and provider changes, retrieval, time, locale, account state, and randomness. A before-and-after change does not establish causation. Public pages can change between crawl and review. Some provider surfaces do not expose the same sources or behavior as consumer interfaces. Small prompt sets are snapshots, not population estimates.

See Data Sources, Changelog, and the raw evidence in each report.