Research protocol · version 1.0
Methodology for measuring AI visibility
This page pre-registers Citeora’s 72-prompt weekly monitoring protocol. It defines the study before results are published. No benchmark result or customer outcome is claimed here.
Quick answer
Quick answer
The protocol treats one prompt-engine response as one observation. A versioned panel of 72 fixed prompts is run across ChatGPT, Gemini and Perplexity, with a 12-prompt sentinel subset repeated during the week. Valid responses feed transparent mention and citation metrics; failed calls remain reliability events and never become negative brand observations.
Scope and unit of observation
The monitoring question is whether and how Citeora appears for a fixed set of AI SEO, GEO, visibility and citation-monitoring questions relevant to English- and Chinese-language buyers, with Hong Kong represented explicitly. One observation is one exact prompt submitted to one named engine in one recorded run.
The protocol measures responses, not a universal domain rank. It does not combine organic search positions, API observations and consumer-interface screenshots into one score.
Prompt sampling and version control
Panel version citeora-72-v1 contains 72 prompts: 40 English, 22 Traditional Chinese for Hong Kong and 10 Simplified Chinese. Intents are allocated before collection: 14 category, 12 problem-to-solution, 10 feature or workflow, 10 comparison, 10 audience or local fit, 8 educational, 4 branded controls and 4 negative controls.
Fifty-four prompts use a natural answer mode and 18 explicitly request researched answers with sources. Discovery prompts do not name Citeora. Every prompt has a stable key, locale, intent and mode. Wording changes create a new panel version and are not backfilled into old runs.
Engine run protocol
The scheduled full panel is 72 prompts across three engine families, producing up to 216 response attempts per weekly wave. A 12-prompt sentinel subset may run once on each of three additional days, producing up to 72 further attempts. Provider-call and token caps, a production-only feature flag, authentication and an idempotent week-and-panel key must all pass before a run can start.
Official grounded APIs are used for repeatable collection where configured. Provider, model identifier, locale, request ID, timestamps and failure class are recorded. Consumer-interface spot checks are a separate monthly validation sample and are never merged into the API time series.
Brand and entity resolution
Citeora and reviewed spelling variants form the measured entity set. Exact and normalized matches are extracted from the raw response; ambiguous references require review. Discovery prompts are checked for accidental brand leakage before a panel is released.
Recommendation position is recorded only when a response provides a defensible ordered shortlist. General prose order is not automatically interpreted as rank.
Citation extraction and URL normalization
Raw citation fields and the response are retained. Derived records preserve observed URL, normalized URL, registrable domain, source ownership and source type. Fragments and approved tracking parameters are removed; redirects are recorded rather than silently erasing the original reference.
Metric formulas and denominators
Mention rate = valid responses containing the resolved entity ÷ valid responses. Citation rate = valid responses containing at least one qualifying citation ÷ valid responses. Owned-domain citation rate = valid responses citing an approved owned domain ÷ valid responses.
Citation persistence = comparable repeated runs retaining the same normalized cited domain ÷ comparable repeated runs. Mean reciprocal rank may summarize valid ordered placements as the mean of 1/rank, with zero for eligible non-appearances. Every output also shows eligible count, attempted count and provider failure rate.
A fixed-test “top five” status, if reported, means rank five or better in four-week rolling mention share of voice for three consecutive weekly waves with at least 90% panel coverage. It is a rank inside this declared test panel—not a universal GEO rank.
Quality control, limitations and change log
Automated extraction keeps a review state. Material claims are withheld when coverage is inadequate, failure rates are excessive, entity resolution is uncertain or required review is incomplete. Raw responses are immutable; parser and scoring versions are stored so derived results can be reproduced.
Answers can vary because of model sampling, retrieval indexes, personalization and provider changes. The protocol can measure repeated observations but cannot prove that one content change caused an answer change. Version 1.0 was published on 10 October 2026. Baseline results will be released only after the scheduled runner is enabled, the first complete waves pass quality review and the eligible dataset is available.
Primary sources and further reading
- Overview of OpenAI crawlers — OpenAI
- Perplexity crawlers — Perplexity
- Creating helpful, reliable, people-first content — Google Search Central