OngoingStudy 01
AI Recommendation Stability
If the same legal query is asked repeatedly, how consistently does the same firm appear?
Measures · recommendation stability
Research · FirmRanker programme
How does AI decide which law firms enter the answer?
Through FirmRanker, I’m building a structured research programme around AI legal discovery — repeatable observations across models, prompts, markets, practice areas and time.
- 4
- Studies in the programme
- 4
- AI platforms observed
ChatGPT · Gemini · Claude · Perplexity - Repeated
- Observations, not one-off screenshots
- Validated
- Before anything is published
Current research programme
Four questions we’re measuring.
Each study turns a common assumption about AI and law firms into something that can be observed, repeated and checked.
ProgrammeStudy 02
Cross-model consensus
To what extent do ChatGPT, Gemini, Claude and Perplexity surface the same law firms?
Measures · cross-model consensus
ProgrammeStudy 03
Prompt sensitivity
How much does recommendation behaviour change when wording or intent changes?
Measures · prompt sensitivity
ProgrammeStudy 04
Source landscape
Which domains are searched, cited or associated with recommendations?
Measures · source landscape
No findings are published on this page until each study has passed validation.
From data to publication
No shortcut from raw export to headline.
Every study passes the same gates before anything is published.
- 01Study complete
- 02Quality assurance
- 03Entities resolved
- 04Methodology frozen
- 05Analysis
- 06Limitations written
- 07Fact-checked
- 08Published
Research principles
How the evidence is handled.
Preserve the evidence
Original AI responses remain immutable.
Separate observation from interpretation
The raw answer is evidence. Recommendation classifications are derived data.
Repeat where variability matters
AI responses can change. Repetition is part of the measurement problem.
Validate automation
Automated extraction should be tested against human-reviewed observations.
Publish limitations
Every serious report should explain what it can and cannot establish.
See the full method
Prompts, repetitions, classification and validation.
Read the methodology→Research Notebook
Notes from building the research system.
Methodological observations — the measurement problems we encounter and how FirmRanker handles them. Not study findings.
- Note 001Why we’re separating “recommended” from “suggested”Read→
- Note 002Why one AI response cannot establish visibilityRead→
- Note 003Why entity resolution matters when measuring law firmsRead→
- Note 004Why AI citations need careful interpretationRead→
Flagship research · In development
The State of AI Search for Law Firms
A future annual research property examining how AI systems discover and recommend law firms across markets, practice areas and models.
Findings will only be published once the underlying research has passed validation.
Start here
I’m a managing partnerUnderstand the commercial exposure and what to prioritise.AEO for Law Firms → I’m a practice managerFind out whether AI recommends your firm — in about a minute.Check your firm → I’m in marketingGet one practical, evidence-led idea every week.Join The AI Referral →Need help now?