Legal AI SEO Jul 8, 2026 9 min read

How ChatGPT Recommends Divorce Lawyers: What We Learned Auditing NYC

ChatGPT recommends divorce lawyers by retrieving named firms from a shortlist that its training data and retrieval layer have built up over months of web crawling — not from lawyer directories, and not from Google search rankings. In NYC divorce, three firms hold 53.5% of every AI recommendation. This piece walks through how AI actually decides which lawyer to name, using data from 7,485 citations across four assistants, and what a firm outside the top ten can do about it.

How does ChatGPT decide which lawyer to recommend?

ChatGPT decides which lawyer to recommend by retrieving a passage from the open web that answers the buyer's question in a clean, quotable form — and citing the firm that owns that passage. It is not scoring domains the way Google does. It is not ranking directory listings. It is matching the buyer's phrasing against passages the model has seen in training data or fetched via a retrieval layer, then repeating the passage that best matches.

That mechanism has three consequences for law firm marketing. First, the firms that get named are the firms whose sites contain sentences shaped like buyer questions with self-contained answers. Second, the shortlist ChatGPT draws from is remarkably stable within any given category — the same handful of firms show up across dozens of variant prompts, because the same handful of pages are the most quotable in the model's corpus. Third, single-model results are not the market. Different assistants have different retrieval layers and different training cuts, and they disagree with each other by margins large enough to flip an entire competitive story.

To see what this actually looks like in the wild, we ran a Viclaro Atlas audit of NYC divorce: 141 prompts, four assistants (ChatGPT / GPT-4o, Claude Sonnet 4.6, Gemini 2.5 Pro, Perplexity Sonar-Pro), two samples per prompt. The result is 7,485 citations of 368 named firms across 544 recommendation slots. It is the most-covered category in the Atlas corpus, four times larger than the next-biggest legal vertical, and the ranking behaves like a category with a real AI consensus. The rest of this piece is what that audit says about how AI recommends divorce lawyers.

Do AI assistants use lawyer directories or law firm websites?

AI assistants use law firm websites far more than directory listings when recommending specific firms in high-consideration verticals like divorce. Directory sites like Avvo, Super Lawyers, and Martindale-Hubbell do surface in AI responses — but usually as generic reference links, not as the source that names one firm over another. When an assistant names Aronson Mayefsky & Sloan or Berkman Bottger, it is pulling from those firms' own pages and from the small number of editorial mentions and profile pages that reproduce their scenario-specific content in citable form.

The mechanical reason: directory listings list firms in bulk. A page listing thirty NYC divorce lawyers with equivalent one-paragraph blurbs contains no quotable passage that names one firm as the answer. A firm's own site, if written correctly, contains many such passages. Retrieval layers preferentially cite the source that produces the most specific, self-contained answer sentence, and firm sites win that comparison against generic directory pages by design.

This is why the "get listed on more directories" playbook that worked for local SEO does not transfer cleanly to AI. Directory presence still helps with authority signals, but it is not what makes ChatGPT name a firm. What makes ChatGPT name a firm is what the firm's own site says about specific buyer situations — and how quotable that content is.

What makes one law firm rank higher than another in AI search?

Three factors separate the firms that win NYC divorce AI recommendations from the firms that do not. First, scenario-shaped content: pages that name specific buyer situations ("hidden assets," "high-net-worth divorce," "business valuation in divorce," "prenuptial enforcement") as headings, each followed by a two-to-three-sentence quotable answer. Firms whose pages read like an FAQ about the buyer's exact problem outperform firms whose pages describe capabilities in marketing prose.

Second, coverage across all four assistants. AI coverage — how many of the four assistants named the firm at least once — is a reliability signal that raw share hides. A firm cited by four of four assistants has a defensible position. A firm at the same aggregate share cited by only two of four is one model update away from losing half its visibility. In our NYC divorce audit, every one of the top three firms is cited by four of four assistants. The fourth-ranked firm by aggregate share is not, and its position is structurally more fragile than the number tells you.

Third, prompt-bucket concentration. Atlas classifies prompts into buckets by buyer intent. For the top NYC divorce firm, Aronson Mayefsky, citations concentrate in the decision bucket (50 citations for prompts like "who should I hire") and validation bucket (35 citations for prompts like "is this firm a good choice"). These are the commercial buckets. A firm can accumulate citations in situation and learning buckets and still lose the actual client — decision and validation are where the retainer gets signed.

Firms that combine scenario content, full four-of-four AI coverage, and citation concentration in decision and validation win. Firms missing any one of the three underperform their apparent share. Our guide to how to rank in ChatGPT covers the specific content patterns that produce each of these outcomes.

The three firms winning NYC divorce (and why)

Across the 7,485 citations, three NYC divorce practices absorb 53.5% of every AI mention. Aronson Mayefsky & Sloan takes 21.0% share (114 citations). Berkman Bottger Newman & Schein takes 17.8% (97 citations). Blank Rome takes 14.7% (80 citations). All three are cited by every one of the four assistants. This is a genuinely concentrated market — one in five AI answers about NYC divorce names Aronson Mayefsky, and more than half of all answers name one of these three.

The consensus does not appear by accident. All three firms publish pages that name specific divorce scenarios in the buyer's vocabulary: hidden assets, high-net-worth divorce, complex property division, business valuation. All three have content structured so that a retrieval layer can extract a self-contained answer paragraph. All three have earned enough independent editorial coverage that their scenario-specific language appears in multiple places on the web, which reinforces the model's confidence in citing them.

What is striking is how much the per-model breakdown varies underneath that aggregate. Claude cites Aronson 57 times. GPT-4o cites Aronson 33 times. Perplexity cites Aronson 13 times. Gemini cites Aronson 11 times. That is a 5x spread on the top firm in the top vertical Viclaro measures. Blank Rome shows an even sharper 8x spread. The three firms hold the market on aggregate, but the specific model that a buyer opens still determines which of the three is named first — and how confidently.

The fourth-ranked firm by aggregate share, Chemtob Moss & Forman, sits at 13.4% share (73 citations) but is cited by only two of four assistants — Claude and GPT-4o. On the surface it looks like a peer of the top three. In practice, 50% of buyers using AI to shortlist a NYC divorce lawyer are effectively blind to it. Cohen Clair Lans Greifer Thorpe & Rotten sits at 7.9% share (rank five) but is cited by all four assistants — a more defensible position than a higher share number with half the model coverage.

The live leaderboard for this category is at the NYC divorce leaderboard.

What lower-ranked firms can do to catch up

A firm sitting outside the top ten in NYC divorce is not competing with Aronson Mayefsky for the number-one slot. That is not a realistic near-term target — the top three are supported by content that took years to accumulate and editorial reach that is expensive to replicate. What is realistic is moving from two-of-four AI coverage to three-of-four, or from mid-tier to top-tier within the two assistants where the firm is already visible.

The most common gap Viclaro sees on audits of firms at ranks 8 through 20 is language mismatch. The firms have partners with strong offline reputations, but their public pages describe practice areas in credentials and jurisdictions rather than in the situational language a buyer actually uses. "Complex litigation matters" instead of "hidden assets." "Sophisticated matrimonial practice" instead of "business valuation in divorce." Same work, different vocabulary — and assistants match the buyer's vocabulary, not the firm's.

The second gap is uneven model coverage. A firm cited by Claude and GPT-4o but invisible to Gemini and Perplexity should treat that as a specific, measurable retrieval problem, not a mystery. Gemini and Perplexity have different training signals and different retrieval preferences — Perplexity in particular weights recent web presence heavily — and the fix is usually a combination of scenario-specific content plus enough independent editorial reach that the newer models learn the firm exists.

The third gap is prompt-bucket weakness. A firm strong in learning-bucket prompts but absent in decision and validation is accumulating citations that do not close clients. Rebalancing content toward "who should I hire" and "is this firm a good choice" prompts, with FAQPage schema on the answer blocks, tends to shift the bucket distribution within a snapshot cycle or two.

How AI recommendations for lawyers will change in the next 12 months

Two directional changes are already visible in the Atlas data. First, the AI shortlist consolidates over time in categories with strong publishing incumbents. In NYC divorce, the top three firms have gained aggregate share across the last several snapshots, not lost it. Consensus, once formed, is self-reinforcing — the more often an assistant names a firm, the more that citation appears in downstream training data, and the more likely the next generation of models will keep naming it. That is the pattern that turned RMA of New York into a 60% category dominant in NYC IVF, and it is running in slower motion in divorce.

Second, model heterogeneity is unlikely to shrink meaningfully. Each provider has its own retrieval architecture, training cadence, and citation preferences. Panels that measure across all four assistants will continue to see 3x-to-10x spreads on the same firm. Buyers will continue to get materially different shortlists depending on which assistant they open. Any AI SEO for law firms strategy that optimizes against one model will remain fragile.

The near-term action for firms outside the top three is: measure the panel, identify which of the four assistants already cite the firm even occasionally, ship scenario-specific content in the buyer's vocabulary on the topics those assistants already weight, wrap the answer blocks in FAQPage schema, and re-measure. That is the loop. It is the same loop that got the top three firms to the top three.

The live NYC divorce leaderboard, refreshed on the Atlas cadence, is at the NYC divorce leaderboard. The methodology behind it is at the methodology page.

Key takeaways

ChatGPT recommends divorce lawyers by retrieving quotable scenario-answer sentences from firm sites, not by ranking directories or Google results.
Three NYC divorce firms hold 53.5% of AI recommendations, and all three are cited by every one of the four major assistants — that is what a concentrated AI market looks like.
Per-model variance is severe: Claude cites the top firm 57 times to Gemini's 11 on the same prompt set, so single-model results are 5x off the panel average.
For firms outside the top ten, the near-term move is scenario-shaped content in buyer vocabulary plus FAQPage schema — not chasing the number-one slot.

Next step

Atlas shows the public map. A Viclaro audit turns that map into the prompts your firm is losing and the page edits most likely to change the next scan.

See the live NYC divorce leaderboard.