Repeat, not one answer
Independent, context-zero observations make output variability visible.
METHODOLOGY · V1.1
If the measurement is not understandable, the recommendation is not trustworthy.
We do not judge AI visibility from one screenshot or an unexplained score. Every finding connects to a query, repeat observation, raw evidence, rule, and limitation.
Separate metrics, not a score
Mentions, recommendations, positions, and citations stay separate.
Evidence trail, not assertion
The records and classifications behind a result are retained.
- 01
Design and freeze the query set
Market, language, treatments, and commercial scope are fixed before the benchmark begins. Queries are generated to cover categories (e.g., best clinic, quality, safety, price) to avoid selection bias. Competitors are selected based on market presence and initial AI output trends.
- 02
Collect independent runs (Context zero)
Each query is observed multiple times in independent sessions with zero previous context. One answer is never treated as definitive evidence.
- 03
Record provider, model, and scope
Language, market localization prompts, date, AI provider (e.g., OpenAI), explicit model version (e.g., gpt-4o), successful observations, and handling of failed runs/refusals are disclosed.
- 04
Preserve raw responses
The outputs behind the review are retained within applicable licensing and privacy terms so classifications can always be verified against the raw text.
- 05
Match known entities deterministically
Clinic names, domains, and verified variations follow explicit matching rules.
- 06
Review ambiguity manually
Unknown clinics, unclear ranking language, and source relationships enter human review.
- 07
Accept explicit ranking only
Text order is not treated as position without clear numbering or recommendation language.
- 08
Re-measure comparably
Market, language, queries, and rules stay as consistent as possible when measuring change.
METRIC FORMULAS
The denominator is never hidden.
Rates are presented using successful observations. Provider network failures are excluded from denominators and retried. Valid refusals or empty responses are treated as successful observations and included.
Mention rate
Successful responses naming the clinicTotal successful responses
Explicit recommendation rate
Responses clearly recommending or ordering the clinicTotal successful responses
Top-position rate
Explicit recommendations in defined leading positions (Top-3)Total successful responses
Citation rate
Responses with a reviewed clinic-associated or 3rd-party sourceTotal successful responses
Average recommendation position
Sum of non-null explicit positionsExplicit recommendations with position data
TECHNICAL DEFINITIONS
Exact Meanings of Concepts
Independent Run
A completely fresh AI session with zero prior conversational context or memory.
Successful Observation
A complete generation without network errors or content policy blocks.
Provider Failure
Instances where the platform fails to respond. These are excluded from the denominator and retried.
Model & Version Logs
The exact LLM version active at the time of testing (e.g., gpt-4o-2024-08-06) is documented for every benchmark.
Market Localization
VPN routing or language locales are used to simulate the target patient's origin market.
Competitor Selection
Competitors are defined directly by the clinic or extracted organically from initial AI responses.
QUERY TAXONOMY
Sample Query Categories
To prevent selection bias, queries are grouped across the following diverse patient intents.
- Best clinic / Top lists
- Direct Comparisons
- Quality-first
- Doctor / Technique-specific
- Safety & Accreditation
- Price & Value
- Market localized (e.g., UK patient)
WHAT THE METHOD CANNOT PROVE
Visibility is not a business outcome by itself.
A benchmark does not prove revenue, patient demand, clinical quality, treatment outcomes, or a permanent AI ranking. It separates observation from causal inference.
A TRANSPARENT START