Methodology

Product boundary

Geoduck measures responses returned by named AI models under recorded API execution conditions. It does not scrape or claim to reproduce the consumer ChatGPT, Claude, Gemini, or other applications. Those products may differ because of system instructions, personalization, geography, browsing, and other behavior Geoduck does not control.

Geoduck uses language such as “visibility in tested model responses,” not “ChatGPT ranking,” guaranteed placement, or consumer-app results.

AI Visibility Rate

AI Visibility Rate =
  eligible responses containing a Confirmed Mention
  --------------------------------------------------
  all eligible responses

• A response counts once even when an entity appears repeatedly.
• Failed and timed-out calls never enter the denominator.
• Only deterministic Confirmed Mentions contribute.
• Branded Diagnostic Prompts never enter unbranded visibility metrics.
• Web-Grounded and Knowledge-Only results are reported separately and never mixed.

Sampling

• Free checks: 1 sample per prompt–model pair (directional snapshot).
• Pro checks: 3 samples per prompt–model pair.
• Each failed model response is retried up to two times without extra credits.
• 3/3 success: complete. 2/3: incomplete (marked). 0–1: insufficient evidence (excluded from aggregates).
• Free single-sample checks are never presented with paid-monitoring confidence language.

Mention detection

At launch, detection is deterministic. Canonical names and approved aliases are matched on token/phrase boundaries (never unrestricted substring). Domains are matched on registrable domains. Ambiguous aliases require an approved disambiguator to co-occur before they can establish a Confirmed Mention.

Parent Brand Mentions (a reference to a product's parent brand without the product itself) are shown as supporting context but do not count as product visibility.

Source evidence

• Only structured provider or gateway citations count as Source Evidence.
• URLs found only in generated prose are Response Links, kept separately.
• Geoduck never infers that a cited source caused the answer.
• A Response Link is never promoted into Source Evidence through retrospective scraping.

Measurement integrity

• Execution is blind: the model receives only the prompt, optional Market Context, and a neutral instruction — never entity names, aliases, domains, or expected answers.
• Historical Execution Evidence is immutable. Reanalysis is appended; configuration changes create a new trend segment.
• Model substitution and third-party web-search fallback are forbidden.
• Every Web-Grounded model must pass a recurring live capability test for downstream Zero Data Retention, provider-native search, structured citations, and exact-model execution.