ChatGPT, Claude, Gemini, and Perplexity Do Not Recommend the Same Local Leaders

The assistant changes the winner

When a buyer asks an AI assistant which local professional to hire, the answer depends partly on which assistant receives the question. That sounds intuitive. The size of the difference is easier to miss once four model results are blended into one visibility score.

Viclaro compared model-level recommendation counts in the latest published Atlas snapshots for eight New York City service categories: divorce, personal injury, immigration, plaintiff-side employment, estate planning, cosmetic dentistry, IVF, and cosmetic surgery. The panel used OpenAI GPT-4o, Anthropic Claude Sonnet 4.6, Google Gemini 2.5 Pro, and Perplexity Sonar Pro.

No category had the same recommendation leader on all four assistants. In estate planning and cosmetic dentistry, each assistant produced a different leader. The practical consequence is simple: “number one in AI search” is incomplete unless the measurement says which AI search market it means.

Even the strongest apparent consensus breaks by model

In IVF, RMA of New York’s Westside location led GPT-4o, Claude, and Gemini. It received 194 of 980 extracted GPT-4o firm citations, 183 of 1,509 Claude citations, and 143 of 1,312 Gemini citations. Perplexity produced a different leader: NYU Langone Fertility Center with 125 of 1,501 citations.

Estate planning was the opposite. GPT-4o’s leader was Katzner Law Group. Perplexity’s was Morgan Legal Group. Gemini’s was Lamson & Cutner. Claude’s was Pierro, Connor & Strauss. Four assistants, four model-level leaders.

Cosmetic dentistry also split four ways: New York City Cosmetic Dentistry Center led GPT-4o; 209 NYC Dental led Perplexity; The New York Center for Cosmetic Dentistry led Gemini; and Park Smiles NYC led Claude. An aggregate leaderboard remains useful for describing the total panel, but no aggregate rank should be mistaken for universal assistant agreement.

The pattern across all eight NYC markets

IVF: RMA of New York led GPT-4o, Claude, and Gemini; NYU Langone Fertility Center led Perplexity.

Cosmetic surgery: Dr. Andrew Jacono led GPT-4o, Perplexity, and Gemini; Dr. Sam Rizk led Claude.

Immigration: Fragomen led GPT-4o and Claude; Cyrus D. Mehta & Partners led Perplexity and Gemini.

Divorce: Aronson Mayefsky & Sloan led GPT-4o and Gemini; Berkman Bottger Newman & Schein led Perplexity; Chemtob Moss & Forman led Claude.

Personal injury: Gair Gair Conason led Claude, Perplexity, and Gemini; Block O’Toole & Murphy led GPT-4o.

Plaintiff-side employment: Outten & Golden led GPT-4o and Claude; Wigdor led Perplexity and Gemini.

Estate planning: all four assistants had different leaders.

Cosmetic dentistry: all four assistants had different leaders.

Models also disagree about how many firms belong in the market

The difference is not limited to first place. Assistants produce shortlists of different breadth. In cosmetic surgery, Perplexity named 296 distinct resolved firms in the reconciled extraction corpus, compared with 177 for Claude, 75 for Gemini, and 69 for GPT-4o. Its leading practice held only 4.2% of Perplexity’s extracted citations, while Gemini’s leader held 27.9% of its pool.

Cosmetic dentistry showed the same breadth ordering. Perplexity named 194 firms, Claude 111, Gemini 86, and GPT-4o 56. A business competing for Perplexity visibility was entering a wide, shallow extracted entity pool; the same business on GPT-4o faced a narrower shortlist.

Raw diversity must be read against response volume and prompt behavior. A model that supplies fewer business names will naturally produce a smaller pool. That is why this article reports counted recommendations and distinct firms rather than claiming one assistant is inherently “more diverse.” A normalized diversity analysis is the next step in the series.

Why aggregate AI visibility is still useful

Model disagreement does not make a panel meaningless. It makes the panel necessary. A buyer population uses several assistants, and no public dataset gives a defensible market-share weight for every local-service use case. An aggregate view supplies a neutral portfolio: the firms that appear most often across the defined models and prompts.

The aggregate should be paired with model coverage. A firm with meaningful share across four assistants has a different risk profile from one whose total is produced almost entirely by a single model. Our reconciled eight-industry benchmark found that the number of top-ten firms with four-model coverage ranges from two in cosmetic dentistry to nine in IVF.

This is also why a single ChatGPT screenshot cannot establish market position. The result is one draw from one model under one prompt. The methodology in Why AI Visibility Needs a Panel explains how repeated prompts and per-model reporting turn those draws into a usable measurement.

What service firms should do with model disagreement

First, preserve the model breakdown. If aggregate share rises while three models remain flat, the strategy has not created broad recognition. It has deepened one model relationship. That may still be valuable, but it should be described accurately.

Second, diagnose the prompts and evidence attached to the missing model. Check whether the assistant retrieved owned pages, directories, editorial sources, professional profiles, or no visible source at all. Do not assume that publishing another generic guide will repair every model gap.

Third, set breadth and share objectives separately. “Reach three-model coverage” and “increase aggregate recommendation share” are different goals. In a low-consensus category, winning a previously absent assistant may be the more durable advance. In a broad-coverage category such as IVF, a challenger may need to establish a specific buyer situation the incumbent does not already own.

Finally, rerun a frozen prompt panel. Model results change between samples and provider updates. The leader counts here describe named Atlas snapshots, not permanent facts about the firms or assistants.

Method and limitations

This analysis selects one most-complete run per taxonomy on the active prompt set, then rebuilds rankings in memory from the live extraction tables. The current-state corpus contains 7,637 responses, of which 6,101 are ranking-eligible, and 13,368 extracted firm citations among 3,172 ranked firm records.

An extracted firm citation is a resolved firm appearance in the ranking corpus; it is not a website citation, impression, click, client, or estimate of assistant usage share. The prompt libraries are category-specific, and response totals differ by category and model. The underlying runs were collected from June 26 through July 13, 2026 and reconciled against live extraction on August 11.

Atlas resolves alternate firm names into canonical entities and excludes known directory and junk entities from firm rankings. Entity resolution can still contain errors, and model output, retrieval indexes, and live web results can change. See the public Atlas methodology and individual leaderboards for the current results.