Potential clients are asking AI chatbots to recommend lawyers. Not Googling, not browsing directories. They type questions like "I need a personal injury lawyer in Houston" into ChatGPT, Gemini, Grok, or Claude and receive direct answers, sometimes with specific firm names and sometimes without.
We wanted to know precisely how each major model handles these queries. Not speculation, not anecdote. So we ran 200 queries, 50 per provider, across all four platforms and measured everything: which responses name firms, which cite URLs, which add disclaimers, which reference directories, and how that behavior shifts with what the person asks.
This is what we found.
Methodology
We submitted 50 queries to four large language models: ChatGPT 5.2 (gpt-5.2 via the OpenAI API), Gemini 3.1 Pro (gemini-3.1-pro-preview via the Google GenAI SDK), Grok 4.1 (grok-4-1-fast via the xAI API), and Claude Opus 4.6 (via the Claude Code CLI). All queries used temperature 0 for deterministic, reproducible output.
The 50 queries were distributed across five categories, 10 queries each, designed to represent the full spectrum of how real consumers search for legal services:
- City and practice area: direct lawyer-finding queries such as "I need a personal injury lawyer in Houston."
- State and legal question: jurisdiction-specific process questions such as "How do I file for bankruptcy in California."
- Generic best or top: reputation-seeking queries such as "Best personal injury law firm near me."
- Specific scenario: situation descriptions such as "I was rear-ended and the other driver has no insurance."
- Comparison and evaluation: cost and assessment queries such as "How much does a criminal defense lawyer cost."
We chose these four models because they represent the four most-used consumer AI chatbots as of February 2026. Each takes a different approach to safety, citation, and recommendation behavior, which makes them well suited to comparative analysis.
Each response was analyzed programmatically for whether it named specific law firms (using pattern matching for common legal firm naming conventions), whether it cited URLs, whether it included disclaimers (categorized as "not legal advice," "AI disclaimer," or "consult a lawyer"), which legal directories it mentioned, total word count, and specificity level (generic, location-specific, or firm-specific). Of 200 total queries, 198 returned valid responses, with one Gemini error and one Claude error, for a 1 percent error rate.
Finding 1: which models name specific law firms
The most consequential question for any firm is simple: does the model actually name you?
| Provider | Names a firm |
|---|---|
| Gemini | 53% |
| Grok | 52% |
| ChatGPT | 26% |
| Claude | 12% |
Gemini and Grok are nearly tied as the most willing to name specific firms. When someone asks "I need a personal injury lawyer in Houston," Gemini names actual firms 100 percent of the time, and Grok does so in 80 percent of city and practice-area queries. ChatGPT names firms in 90 percent of those queries, its highest rate for any category. Claude names firms in only 10 percent of city-specific queries.
The gap between providers is wide enough to matter. If you are a firm wondering whether AI is sending clients to your competitors, the answer depends entirely on which model your prospective client happens to use. A Gemini or Grok user will almost always receive specific firm recommendations. A Claude user will almost always receive general guidance instead.
Firm naming by query type
| Query type | ChatGPT | Gemini | Grok | Claude |
|---|---|---|---|---|
| City and practice area | 90% | 100% | 80% | 10% |
| Generic best or top | 20% | 70% | 90% | 22% |
| State and legal question | 10% | 60% | 40% | 10% |
| Specific scenario | 10% | 22% | 30% | 10% |
| Comparison and evaluation | 0% | 10% | 20% | 10% |
The pattern is clear. The more geographic and practice-area-specific the query, the more likely the model is to name an actual firm. Generic evaluation questions ("how much does a lawyer cost") rarely produce firm names from any provider. City and practice-area queries trigger firm names at an overall rate of 70 percent, while comparison and evaluation queries produce them only 10 percent of the time.
Most frequently named firms
Morgan & Morgan was the most frequently named individual firm across all four models, appearing 10 times in 200 responses. No other firm came close. "The Law Offices" and "The Legal" each appeared 4 times, while Arnold & Itkin LLP, Georgia Legal, and Immigration Legal each appeared 3 times. National brand recognition translates directly into AI recommendation frequency.
Finding 2: URL citations and source attribution
Do these chatbots actually link to the firms or resources they mention?
| Provider | Cites a URL |
|---|---|
| Grok | 40% |
| ChatGPT | 26% |
| Gemini | 20% |
| Claude | 2% |
Grok leads URL citation by a clear margin, providing links in 40 percent of all responses, consistent with its overall habit of giving detailed, heavily sourced answers. ChatGPT follows at 26 percent and Gemini at 20 percent. Claude almost never provides URLs, at just 2 percent of responses.
URL rate by query type
| Query type | ChatGPT | Gemini | Grok | Claude |
|---|---|---|---|---|
| City and practice area | 80% | 30% | 80% | 10% |
| State and legal question | 20% | 30% | 50% | 0% |
| Generic best or top | 10% | 10% | 50% | 0% |
| Specific scenario | 20% | 22% | 10% | 0% |
| Comparison and evaluation | 0% | 10% | 10% | 0% |
City and practice-area queries produce the most URLs overall (50 percent), especially from ChatGPT and Grok, which both reach 80 percent for this category. For comparison and evaluation queries, URLs are rare across every provider, only 5 percent overall.
This carries a direct implication for legal marketing. When a model names your firm but provides no link, which is the case in the majority of Claude responses and many Gemini responses, the user has nothing but the name. If they then search for that name and your website is slow, dated, or hard to find, you lose the referral the model already handed you.
Finding 3: disclaimers and hedging
How often do these models tell users to "consult a real lawyer," or disclose that they are AI?
| Provider | Adds a disclaimer |
|---|---|
| Gemini | 71% |
| Grok | 66% |
| Claude | 22% |
| ChatGPT | 10% |
Gemini is by far the most aggressive at disclaiming, attaching some form of disclaimer to 71 percent of all legal responses. The overwhelming majority of those disclaimers, 36 of 38, are AI disclosure notices, essentially telling the user "I am an AI." Grok follows at 66 percent, with a more balanced mix of "not legal advice" warnings (17), AI disclaimers (16), and "consult a lawyer" suggestions (5).
ChatGPT is the least likely to disclaim, at just 10 percent, and when it does it uses "not legal advice" language. Claude sits at 22 percent, drawing primarily on AI disclaimers (9) and "not legal advice" statements (5).
Disclaimers by query type
| Query type | ChatGPT | Gemini | Grok | Claude |
|---|---|---|---|---|
| State and legal question | 30% | 100% | 90% | 60% |
| Specific scenario | 10% | 100% | 70% | 20% |
| City and practice area | 10% | 80% | 70% | 10% |
| Comparison and evaluation | 0% | 30% | 60% | 20% |
| Generic best or top | 0% | 50% | 40% | 0% |
State and legal-question queries trigger the most disclaimers overall (70 percent), which is intuitive: asking about a specific legal process comes closest to requesting actual legal advice. Gemini reaches a 100 percent disclaimer rate for both state legal questions and specific scenario queries. Generic best or top queries produce the fewest disclaimers, at 23 percent overall.
Finding 4: the role of legal directories
Legal directories are the backbone of how these models discover and reference law firms. Across 200 responses, Grok referenced directories in 74 percent of its answers, Gemini in 47 percent, ChatGPT in 22 percent, and Claude in only 10 percent.
| Directory | Total mentions | ChatGPT | Gemini | Grok | Claude |
|---|---|---|---|---|---|
| Avvo | 60 | 6 | 17 | 34 | 3 |
| Martindale-Hubbell | 41 | 9 | 12 | 19 | 1 |
| Super Lawyers | 37 | 4 | 15 | 17 | 1 |
| Yelp | 18 | 0 | 0 | 17 | 1 |
| Best Lawyers | 17 | 5 | 4 | 8 | 0 |
| Justia | 17 | 0 | 0 | 16 | 1 |
| 14 | 3 | 5 | 6 | 0 | |
| Nolo | 14 | 0 | 0 | 12 | 2 |
| LegalZoom | 7 | 0 | 0 | 7 | 0 |
| Lawyers.com | 5 | 0 | 0 | 5 | 0 |
| FindLaw | 5 | 0 | 0 | 3 | 2 |
| BBB | 4 | 0 | 0 | 4 | 0 |
Avvo dominates. It appeared in 60 of 200 responses, making it the single most important external signal these models draw on when discussing firm quality and reputation. Grok alone referenced Avvo 34 times, more than any other provider-and-directory combination in the entire study. A firm without a strong, current Avvo profile is effectively invisible to the most common AI discovery mechanism.
Martindale-Hubbell is second at 41 mentions, and Super Lawyers third at 37. Together, the top three directories accounted for 138 of all directory references, nearly 60 percent of the total.
Grok is the clear directory champion. It referenced Yelp (17), Justia (16), Nolo (12), LegalZoom (7), Lawyers.com (5), and BBB (4), directories the other three providers essentially ignored.
Directory mentions by query type
| Query type | ChatGPT | Gemini | Grok | Claude |
|---|---|---|---|---|
| City and practice area | 40% | 100% | 100% | 20% |
| Generic best or top | 60% | 80% | 100% | 11% |
| Comparison and evaluation | 10% | 40% | 90% | 20% |
| Specific scenario | 0% | 11% | 70% | 0% |
| State and legal question | 0% | 0% | 10% | 0% |
City and practice-area queries and generic best or top queries are the strongest directory triggers. Gemini and Grok both reach a 100 percent directory reference rate for city and practice-area queries. State and legal-question queries almost never mention directories; only Grok does, and only 10 percent of the time.
Finding 5: geographic specificity
We classified each response as generic, location-specific, or firm-specific. The results show how differently each model handles geographic context.
| Provider | Generic | Location-specific | Firm-specific |
|---|---|---|---|
| Grok | 0% | 32% | 68% |
| Gemini | 0% | 43% | 57% |
| ChatGPT | 0% | 66% | 34% |
| Claude | 18% | 67% | 14% |
Grok, Gemini, and ChatGPT never produce a fully generic response; every single answer from these three includes at least location-specific context. Claude, by contrast, returns generic responses 18 percent of the time, answers that reference no specific location or firm.
The firm-specific rate maps closely to the overall firm-naming rate, as one would expect. Grok leads at 68 percent, followed by Gemini at 57 percent, ChatGPT at 34 percent, and Claude at 14 percent. For firms, this means the geographic signals on your website are decisive. When a model produces a location-specific response, it is pulling location data from somewhere. A site that clearly establishes its service area with specific city and county names, structured data with geographic coordinates, and location-relevant content raises the probability of appearing in these location-aware responses.
Side-by-side provider comparison
| Metric | ChatGPT | Gemini | Grok | Claude |
|---|---|---|---|---|
| Names firms | 26% | 53% | 52% | 12% |
| Cites URLs | 26% | 20% | 40% | 2% |
| Adds disclaimers | 10% | 71% | 66% | 22% |
| Mentions directories | 22% | 47% | 74% | 10% |
| Average word count | 408 | 686 | 418 | 153 |
| Firm-specific responses | 34% | 57% | 68% | 14% |
Grok is the most aggressive recommender and the directory champion. It names firms in 52 percent of responses, provides the most URLs (40 percent), and mentions directories more than any other provider (74 percent). It references Avvo, Martindale-Hubbell, Super Lawyers, Yelp, and Justia at rates the others do not approach.
Gemini is the most willing to name specific firms (53 percent) and the most verbose, averaging 686 words per response. It is also the heaviest disclaimant at 71 percent, driven almost entirely by AI disclosure notices. When Gemini recommends a firm it does so thoroughly, but always with a disclaimer attached.
ChatGPT sits in the middle. It names firms at 26 percent, provides URLs at 26 percent, and mentions directories at 22 percent. Its disclaimer rate of 10 percent is the lowest of the four. Given ChatGPT's enormous user base, even a moderate recommendation rate produces meaningful referral volume.
Claude is the most conservative on every metric. It names firms only 12 percent of the time, provides URLs in just 2 percent of responses, and mentions directories only 10 percent of the time. Its responses are the shortest, at 153 words on average, roughly a quarter of Gemini's output. Claude treats legal queries as requests for information, not requests for referrals.
What this means for law firms
The data points to several concrete steps a firm should take now.
1. Your directory profiles matter more than ever
Avvo appeared in 60 of 200 AI responses. Martindale-Hubbell appeared 41 times. Super Lawyers appeared 37 times. These directories are no longer only for consumers browsing directly; they are primary data sources the models use to identify and evaluate firms. A strong, complete, well-reviewed Avvo profile is no longer optional. It is a prerequisite for AI visibility.
2. Geographic and practice-area signals must be explicit
City and practice-area queries triggered firm names 70 percent of the time, the highest rate of any category. Your website needs to make those signals unmistakable through page titles, headers, structured data, and content that names your service cities, counties, and states alongside your practice areas. The models that name firms most often (Gemini at 100 percent and Grok at 80 percent for city queries) are plainly pulling geographic data from web content.
3. Brand recognition compounds in AI
Morgan & Morgan appeared 10 times across 200 responses; no other individual firm came close. In AI-generated answers, brand awareness compounds: the more often a model has encountered a firm name in its training data, the more likely it is to recommend that firm. The result is a flywheel in which established brands draw more AI recommendations, which drive more web traffic, which produces more content, which further raises AI recognition.
4. Your website must convert name-only referrals
Grok provides URLs 40 percent of the time; Claude does so only 2 percent of the time, and Claude is used by millions. When a model names your firm without a link, the user will search for you next. If your website is slow, if your Google Business Profile is incomplete, if your homepage does not immediately say what you do and where you do it, you lose the referral AI already handed you. This is exactly the gap a fast, structured, purpose-built site is meant to close.
5. Different models call for different strategies
Optimizing for Grok, which favors directories and geographic specificity, is not the same as optimizing for Gemini, which names firms often but always disclaims. Both differ again from optimizing for Claude, which rewards authoritative general content over specific recommendations. A serious AI Answer Optimization program accounts for all four major platforms, not one. This is also where the method earns its keep: rather than guess which signals move which model, NitroCMS unifies five data sources, Google Analytics, Search Console, Google Ads, independent SEO and competitor intelligence, and the Constellate Analytics Engine, our first-party analytics, and hands them to Claude Opus, which returns a ranked set of recommendations before any work begins. Evidence over guesswork.
Limitations and future research
This study is a snapshot of model behavior in February 2026. AI models are updated frequently, and response patterns may shift as providers adjust their safety guidelines, training data, and inference configurations.
Our sample of 50 queries per provider (200 total) gives a solid directional signal but is not large enough to claim statistical significance for every sub-category comparison. Some category breakdowns rest on only 10 queries per provider, so a single response shift moves the percentage by 10 points.
Firm-name detection used pattern matching for common legal naming conventions, for example "[Name] Law Firm," "[Name] & Associates," and "[Name] LLP." It is possible some unconventionally named firms were missed, or that a generic reference to "a law firm" produced a false positive despite our filtering.
All queries used temperature 0 for reproducibility. Real user interactions typically use higher temperature settings, which would introduce variation in responses. Our results represent the models' most deterministic output, not the full range of possible responses.
We plan to expand this study with larger samples, additional providers, and longitudinal tracking to measure how AI recommendation behavior changes over time.
Full methodology details
This study was conducted in February 2026 using the following setup:
- ChatGPT: GPT-5.2 (model ID: gpt-5.2) via OpenAI API, temperature 0
- Gemini: Gemini 3.1 Pro (model ID: gemini-3.1-pro-preview) via Google GenAI SDK, temperature 0
- Grok: Grok 4.1 (model ID: grok-4-1-fast) via xAI API, temperature 0
- Claude: Claude Opus 4.6 (model ID: claude-opus-4-6) via Claude Code CLI, default settings
- Total queries: 200 (50 per provider)
- Valid responses: 198 (1 Gemini error, 1 Claude error)
- Error rate: 1%
- Query categories: 5 categories, 10 queries each
- Analysis pipeline: automated JSON extraction, pattern-based firm-name detection, URL parsing, disclaimer classification (3 types: not_legal_advice, ai_disclaimer, consult_lawyer), directory-name matching (12 directories tracked), word count, and specificity classification
All queries were designed to reflect realistic consumer search behavior when looking for legal services. None were designed to game a particular provider's response patterns. The raw data and analysis scripts are available for review.