How AI Recommends NYC Divorce Attorneys

Executive summary

Buyers facing a major life decision — divorce, fertility treatment, plastic surgery, a custody dispute — increasingly consult ChatGPT, Claude, Gemini, or Perplexity before they ever type a search query into Google. For attorneys in high-stakes consumer-services categories, the AI assistant has become the new shortlist generator. Whether you appear on that list is no longer downstream of your Google ranking; it's a separate, parallel, increasingly important visibility problem.

This report measures who AI assistants name when buyers ask about NYC divorce and family-law firms. The current published run contains 141 buyer-style questions across four AI assistants, sampled across 8 conversation-state buckets. It is split into a 544-response ranking layer (questions expected to evaluate or recommend firms) and a 348-response editorial layer (questions capturing advice and discussion where naming a firm is not required). The denominators come from one atomic, quality-checked scan run; retries and historical methodologies are not pooled into the current ranking. Current findings:

• 4 firms currently clear the consensus-tier threshold. The table below derives every rate and interval from the current snapshot rather than preserving figures from an older extraction revision.
• Assistant exposure is disclosed, not assumed equal. The current ranking-layer response counts are ChatGPT 193, Perplexity 93, Gemini 85, and Claude 173. A missing firm citation is not labelled a refusal without response-level classification.
• Conversation state — not keyword intent — is the right axis. A buyer in crisis ("my wife filed yesterday") gets a different response shape than a buyer asking for three firms to interview, even when the underlying need is identical. Splitting the scan along conversation state surfaces firm-citation behavior that keyword-style GEO tools cannot see.
• Process- and crisis-stage queries route through directories. When the prompt is "I want to file for divorce, where do I start?", AI assistants cite Avvo, Martindale-Hubbell, and the NYC Bar Lawyer Referral Service more often than any individual firm. These appear in our editorial layer and are explicitly excluded from the ranking; counting them would deflate every firm's share.

The methodology underlying this report — versioned prompt sets, two-layer ranking/editorial split, honest confidence intervals, immutable snapshots, full verbatim source traceability — is published in full at viclaro.app/leaderboards/methodology. Every citation in the rankings below traces back to a specific AI response, viewable on the public leaderboard.

1. Why measure this at all

The classic SEO measurement playbook — Google rankings, organic search traffic, keyword positions — describes a world in which buyers type queries into a search box and click a blue link. That world is collapsing under a new layer: buyers asking a conversational AI for a recommendation before they ever open a search engine.

The conversational layer behaves differently in three important ways. First, it doesn't return ten links; it returns three to five firm names, often with a recommendation tone ("Berkman Bottger is a leading NYC family-law firm known for..."). Second, the buyer doesn't click — they read the recommendation, internalize it, and then search Google for the firm by name. Third, the AI's choice of who to recommend depends on signals that overlap with but are not identical to Google's ranking signals: schema markup, FAQ structure, quotable headline claims, named partners, and content that mirrors the buyer's actual phrasing.

For attorneys in NYC family law — a category where one new client can be worth $25,000–$200,000 in legal fees, and where almost every buyer is doing weeks of research before they engage — being on the AI's shortlist is now a material distribution channel. This report is the first reproducible measurement of who currently occupies that channel.

2. Methodology

The published snapshot uses the versioned prompt set "divorce-attorneys v3.1" with 141 buyer-style questions grouped into 8 conversation-state buckets. The buckets describe where the buyer is in the recommendation conversation, not merely what keyword they would type:

• Situation, Decision, Validation, Follow-up — the ranking layer: questions where the buyer is asking AI to evaluate or recommend firms.
• Crisis, Advice, Learning, Constraints — emotional / financial pressure states where AI typically does not list named firms.
• Process, Specific need, Conversational, Style/approach, Broad authority — topical questions where the brand may appear as a quoted source rather than as a recommendation.

The two-layer split. Not every question is designed to produce a firm recommendation. When a buyer in crisis asks "be honest, do I have a case," the appropriate AI response is usually advice — not a list of names. Counting those responses toward share would artificially deflate every firm. We therefore split the corpus into two layers:

• Ranking layer — only citations from these scans count toward response mention rate and rank. 544 responses in this run, from 97 distinct buyer questions.
• Editorial layer — non-ranking buckets remain qualitative evidence but never feed the ranking. 348 responses in this run, from 44 distinct buyer questions.

Each question was run against four AI assistants — Claude, ChatGPT, Gemini, and Perplexity — with multiple samples per (question × assistant) to capture how consistently each assistant gives the same answer. Each assistant received each question independently: no system instructions, no memory, no conversation history carried in (except for the Follow-up bucket, which explicitly stages a multi-turn exchange).

Every AI response was parsed into a JSON array of named businesses. Names were normalized (suffix-stripped, sub-group prefixes like "Matrimonial Law Group at X" collapsed onto the parent firm) and matched against a pre-harvested pool of 770 NYC legal entities. When an assistant mentioned a firm we hadn't indexed yet, we looked it up and added it to the directory so future scans would catch it.

We compute three statistics per firm in the ranking layer:

• Response mention rate — distinct eligible responses naming the firm ÷ ranking-layer responses. Rates are non-exclusive because one response can name multiple firms.
• Honest 95% confidence interval on the share. The interval widens as the sample shrinks; a firm with the same share at a smaller N has a wider interval.
• Cross-assistant coverage — how many of the four assistants (out of 4) cited the firm at least once. A firm cited by all four is a structurally different signal from one cited only by Perplexity.

Tier classification: Consensus if the CI lower bound is clearly above the noise floor; Mid-tier if cited by ≥ 2 assistants with share > 3% but the CI still spans the floor; Long tail otherwise. Long-tail firms are still ranked and shown, but the position should be read as ordering, not a statistical claim.

Snapshots are immutable. Every ranking is reproducible against the same prompt-set version and the same assistants, within measured sampling variance. Every citation in the rankings below has a verbatim-AI-response permalink, available to research and commercial partners.

Bucket structure and per-version question counts are published at viclaro.app/leaderboards/methodology. Verbatim question text isn't published — the wording is part of what makes the Atlas measure conversation-state behavior rather than search-engine behavior.

3. The consensus tier — four firms AI assistants recommend across the board

The strongest finding from the scan is the consensus tier: four firms with confidence-interval lower bounds clearly above the noise floor at this sample size. Three are cited by all four assistants; the fourth (Chemtob Moss & Forman) is cited by two but at a share high enough to clear the bar regardless.

Two observations this snapshot supports:

The current leader is Aronson Mayefsky & Sloan, LLP. It was named in 114 of 544 eligible responses (21.0%). The runner-up was named in 97 responses (17.8%). These are non-exclusive response mention rates; they must not be added and described as market share.

Cross-assistant coverage complements the response rate. A 4/4 value means the firm appeared at least once in each tested assistant family. It does not mean each assistant had the same number of opportunities; those denominators are disclosed in the per-assistant section below.

4. The mid-tier — real signal, partial visibility

Below the consensus tier are 3 current mid-tier entities. This tier requires a response mention rate above 3% and appearances from at least two assistant families, while its interval still overlaps the configured noise threshold. The live table—not preserved prose from a prior extraction revision—is the source of truth.

Two patterns recur as we look further down the long tail:

• Single-assistant dominance. Many firms appear repeatedly on Perplexity and almost nowhere else. Perplexity is web-search-grounded and surfaces firms with strong site presence even when their training-data footprint is thin. If a meaningful share of your buyers use Perplexity, the competitive set looks different from the cross-assistant consensus tier.
• Sub-specialty positioning. Firms that publish content explicitly about a niche — collaborative divorce, business-owner divorce, LGBTQ-affirming practice — surface on the questions that match that lane and almost nowhere else. The cost is breadth; the benefit is a defensible territory that the consensus-tier generalists don't compete in.

5. The editorial layer — where AI sends buyers in crisis

The reason the editorial layer is separated from the ranking is that it does something the ranking layer can't show. When a buyer is in crisis ("my wife filed yesterday"), early research ("walk me through how NY divorce works"), or working under hard financial constraints ("I have $10K total"), AI assistants typically don't list named firms at all. They route the buyer toward directories, referral services, and process information — Avvo, Martindale-Hubbell, the NYC Bar Lawyer Referral Service, the New York State court system's self-help portal.

This matters because these are exactly the high-intent moments that historically arrived at firm websites via informational SEO content — "what to expect in a NYC divorce," "how does NY divorce work," "how to find a divorce lawyer." On Google, the buyer lands on a firm-authored explainer page. On ChatGPT, Claude, or Gemini, that same buyer is routed to a directory listing, because AI assistants find structured directory content more quotable than firm prose for these queries.

For firms, there are two responses. The first is to publish process- and crisis-stage content explicitly designed to be quoted: short, structured answer blocks under question-shaped H2s, with FAQPage JSON-LD schema. The second is to accept that directory placement is now its own distribution channel — getting listed prominently on Avvo and Martindale is no longer vanity SEO; it's how AI now routes early-funnel buyers.

The editorial layer is what makes this visible. A ranking-only methodology would either drop these queries (losing the signal entirely) or count them and dilute every firm's share with the directory mentions. Splitting the two layers lets us measure both honestly: the ranking layer reflects who AI recommends when asked, and the editorial layer reflects what AI does instead when the buyer hasn't yet asked for names.

6. Per-assistant divergence — why "AI visibility" isn't one thing

The four AI assistants tested do not behave the same way. Across the ranking-layer responses each one produced for this category:

The implication is that "are you AI-visible?" is not a single binary question. Per-assistant counts need their own denominators. A response with no ranked firm can reflect a genuine non-recommendation, a directory-only answer, an extraction miss, or a refusal; this report does not collapse those states into a refusal statistic.

7. What separates the consensus tier from the long tail

The current snapshot contains 186 ranked entities: 4 consensus-tier, 3 mid-tier, and the remainder in the long tail. The following website patterns are hypotheses for follow-up audits, not causal conclusions established by the recommendation panel alone:

• Question-shaped page titles. The consensus-tier firms have pages titled around the buyer questions — "What is a high-net-worth divorce?", "How does contested custody work in New York?" — not generic service pages ("Family Law", "Divorce Services"). AI assistants match content to query phrasing; question-shaped titles win.
• Named partners on landing pages. Pages that name specific attorneys by full name in H2s and structured author markup get cited far more than anonymous "our team" pages. AI assistants are pattern-matching on "person + practice area + location," and explicit named attribution is what they latch onto.
• Quotable thesis statements. The cited firms have a position — "we are the only firm in NYC that specializes exclusively in matrimonial law," "we represent business owners and partnership stakeholders," "our practice is built around collaborative outcomes." AI assistants prefer to quote claims, not adjectives.
• FAQPage JSON-LD on relevant pages. Schema markup is not the only signal but is a consistent differentiator. Firms with valid FAQPage schema on their highest-traffic pages are over-represented in the consensus tier.
• Press footprint. Firms cited in the Wall Street Journal, New York Times, or specialty trade press accumulate citation footprints that survive into AI training data. The consensus-tier firms each have multiple recent press mentions; most long-tail firms have none.

None of these signals is sufficient alone. But the consensus-tier firms all exhibit four or more of them; the long-tail firms typically exhibit zero or one.

8. Methodological honesty — what this report does not measure

A report of this kind must be explicit about its limits.

• We measure citation share, not quality. A firm cited often by AI is not necessarily a better attorney than a firm cited rarely. We measure visibility in a specific layer; legal outcomes are a different question.
• We measure four AI assistants, not all AI surfaces. Microsoft Copilot, ChatGPT in voice mode, Claude in mobile apps, and various agent-based wrappers may behave differently than the API-accessed assistants we test. The big four cover the dominant share of buyer behavior today, but the surface is fragmenting.
• Mid-tier rank movement is uncertain. The 95% confidence intervals on mid-tier firms span 5–7 percentage points. Reading rank changes from #15 to #12 between snapshots as a "rise" is not statistically supported at this sample size. Top-7 ranks are more stable; mid-tier requires larger N before movement is meaningful.
• We do not measure click-through or conversion. Being cited by ChatGPT does not necessarily produce a phone call. The recommendation channel exists, but the conversion economics of that channel are still being measured.

9. Conclusions

Three things are true at once:

AI recommendation visibility is now a measurable, ranked, reproducible distribution channel. The methodology underlying this report — versioned questions, multi-sample scans, honest confidence intervals, immutable snapshots — produces ranks that survive re-running and can be defended in writing to a sophisticated reader.

The channel is structurally winnable. The gap between consensus-tier firms and the long tail is large but not random. Firms that publish question-shaped, named-partner, schema-marked content occupy the consensus tier. The pattern is reproducible, and the cost of publishing the right content is bounded.

The conventional "AI optimization" pitch is partially wrong. Most tools in this space measure a single assistant, a single prompt format, or a single "AI visibility score." This report's data demonstrates that those simplifications hide more than they reveal. Per-assistant divergence is real. Conversational questions surface different firms than keyword queries. Process-stage questions route through directories. A firm that optimizes for an averaged score risks optimizing for nothing in particular.

The full live index for this category — updated as we re-scan, with verbatim source traceability per citation — is at viclaro.app/leaderboards/nyc/legal/divorce. Methodology disclosure is at viclaro.app/leaderboards/methodology.