AI legal discovery across four engine groups: different firms feature in sampled answers
Pinsent Masons and Shoosmiths are closely matched in the OpenAI group, while CMS, Irwin Mitchell and Harper James feature in other historical samples.
Citations.io’s AI Legal Discovery Index examines more than 127,000 recorded AI visibility data points in commercial legal questions for England and Wales. Produced in association with Legal Directory, this point-in-time research snapshot compares selected national neutral-market results across four engine-labelled groups using selected sampled AI-generated answers.
Selected recommendation rates by engine-labelled group
- OpenAI group: Pinsent Masons 31.3% and Shoosmiths 31.0%. They are closely matched, so the sample does not support declaring a decisive winner.
- Claude group: CMS 25.6%.
- Gemini group: Irwin Mitchell 28.6%.
- Perplexity group: Harper James 22.7%.
Recommendation rate is the percentage of analysed relevant answers in the specified group containing a recommendation. It is not recommendation conversion, the original sector report’s measure of active recommendations as a percentage of a firm’s mentions or appearances.
These selected figures come from different collection modes and are not a controlled head-to-head evaluation of consumer interfaces. They describe a designed historical question mix, not a live ranking or an assessment of legal service quality.
Read the related research
- AI Legal Discovery Index 2026
- Commercial legal discovery in Newcastle upon Tyne
- Original UK Legal AI Visibility Study, including its separate recommendation conversion definition
Methodology and limitations
Point-in-time snapshot. These results describe selected sampled AI-generated outputs from a single historical collection. They are not live rankings and have not been refreshed or recollected.
Recorded data points. Recorded data points are extracted attributes of sampled answers: entity mention records, recommendation flags, sentiment classifications, numeric position measurements and extracted source references. Several attributes can come from the same answer, so they are not independent observations and are not equivalent to an answer count. Repeated outputs for the same question are not independent.
Metric denominators. AI visibility is the share of relevant analysed answers in the stated scope that mention an entity. Recommendation rate is the share of relevant analysed answers in that scope containing a recommendation of the entity. Positive, neutral and negative sentiment are shares of classified mentions of the entity; unknown labels are excluded. Recommendation conversion, used only in the original sector report, is the share of an entity’s mentions or appearances that actively recommend it.
The full England and Wales commercial legal study, produced in association with Legal Directory, comprises 127,349 recorded attributes: 26,259 extracted entity mention records, 26,259 recommendation flags, 26,259 sentiment classifications, 14,225 numeric position measurements and 34,347 extracted source references. The regional analysis covers 20 cities and contains 53,896 attributes within that total, not in addition to it: 11,337 mention records, 11,337 recommendation flags, 11,337 sentiment classifications, 6,319 numeric positions and 13,566 extracted source references.
Each engine-labelled figure uses that group’s own neutral-market subset as its denominator, so percentages are comparable only within a group. National comparisons use pooled neutral-market questions and stored canonical labels. Local analysis uses limited provisional grouped labels, so national and city percentages must not be compared directly. Entity aliases and matching remain provisional and subject to review.
The question mix was designed to cover commercial legal discovery and is not weighted to actual buyer behaviour. Coverage is limited to the selected services, industries and places. The study includes OpenAI and Gemini model/API groups and saved native Claude and Perplexity groups; collection modes differ, so engine groups are not a controlled head-to-head comparison. Extracted source references are not necessarily native citations and do not by themselves prove that a source supports a claim. Duplicate signal rows were resolved to the most recently updated record per answer.
Sentiment describes generated language, not client opinion, customer satisfaction or an assessment of legal service quality. The findings do not establish expertise, regulatory status, suitability, market share, actual instructions or service quality, and they are not legal advice. Check current firm identity and regulatory status independently.
Detailed study provenance, collection timing and sample breakdowns are retained in Citations.io’s internal audit records and are available from the research team on request for editorial or scientific verification, where they can appropriately be shared. Contact support@citations.io.
Media enquiries
For supporting data and commentary, contact Josh Tulip, founder of Citations.io, at support@citations.io.

