The AI Referral

Methodology

FirmRanker must be able to show its work.

If AI visibility is going to be measured seriously, the methodology matters. A published finding should ultimately be traceable back to the observations that produced it.

Study → Experiment → Observation

01

Study

The research question and defined universe.

02

Experiment

The exact prompt, location, practice area, model and repetition.

03

Observation

The immutable answer returned by the AI system.

Derived analysis sits on top of the observation. It does not replace it.

Prompts and models

Prompts should be predefined, preserved historically and classified by intent. Model identity should preserve provider, requested model, returned model/version where available, execution time and search/grounding configuration.

Why repetition matters

A single answer can be useful as an example. It is not necessarily a stable measurement. Repeated observations allow FirmRanker to investigate recommendation stability and distinguish persistent signals from stochastic variation.

Law-firm extraction and entity resolution

AI answers are unstructured text. FirmRanker extracts firm references while preserving the exact raw answer. Different names for the same firm can be resolved to a canonical entity without changing what the AI actually said.

Example“Slater & Gordon” and “Slater and Gordon” may resolve to the same canonical entity, while the original extracted names remain preserved.

Response treatment

Not every mention is a recommendation. FirmRanker separates categories such as recommended, suggested/consideration, mentioned, source-only and negative/cautionary. Prompt intent remains separate from response treatment.

Sources and searches

Search results, provider grounding/citation sources and visibly cited sources should not be collapsed into one concept. Where possible, raw URLs, resolved destinations, domains, titles, providers and relationships to observations are preserved.

Human validation

Machine extraction should be capable of comparison against manually reviewed ground truth, preserving the original machine output, proposed corrections, final human decision, reviewer and timestamp.

Limitations

  • AI outputs can vary.
  • Models change.
  • Providers expose different metadata.
  • Search grounding differs by provider.
  • Observed associations do not prove causal influence.

FirmRanker measures behaviour under defined experimental conditions. That is not the same as claiming to expose a model’s complete internal reasoning.