What the
Machines
Said

The four most-recommended restaurants in New York by normalized response appearances are, in order: Le Bernardin (791), Eleven Madison Park (631), Daniel (590), and Per Se (524). All four platforms agree on these four. The gap between #4 and #5 — Per Se at 524, The Modern at 417 — is 107 appearances. That gap is larger than the total appearance count of most restaurants in the dataset. New York's fine dining canon, as rendered by AI, is extremely concentrated at the top. The machines agree on who belongs there.

The Modern (417), Gramercy Tavern (395), and Atomix (346) form the next tier. Their platform profiles look nothing like the top four. The Modern is relatively balanced (ChatGPT 106, Claude 100, Gemini 116, Perplexity 95). Gramercy Tavern skews Claude (134) and Gemini (138) with ChatGPT trailing at 63 and Perplexity at 60. Atomix skews Claude heavily (118) with ChatGPT at only 59. Below rank four, every restaurant tells a different story about which platform found it and why.

Gramercy Tavern's ChatGPT count (63) is the lowest of any restaurant in the top six, and the most anomalous. Thirty years of defining New York hospitality. The platform a first-time visitor is most likely to consult. 63 appearances. Something is off.

The pattern across Jean-Georges (243), Masa (245), and Gabriel Kreuther (191) is consistent: the platforms diverge significantly. Jean-Georges has only 32 Claude appearances against 80 ChatGPT and 77 Gemini — old-guard French cuisine surfaces more readily in platforms reading current dining guides. Masa has 105 Claude appearances but only 49 ChatGPT. Both are Michelin 3-star restaurants. Both have decades of press. The machines still disagree on how much attention they deserve. Old-guard New York luxury is not a universal signal. It depends which training data you inherited.

Perplexity's platform profile diverges sharply from Claude's. While Claude leads every top-four restaurant, Perplexity is the most variable — ranging from 166 for Le Bernardin to 104 for Daniel. When asked about difficult-to-book restaurants, special occasions, or where to take an international visitor, Perplexity consistently reaches for web-sourced, citation-backed answers. That means it performs better on restaurants with active booking discourse and worse on restaurants that are primarily known through critical writing.

• Perplexity appears in 166 Le Bernardin responses, 135 EMP responses, 106 Per Se responses — all rely on active booking and review coverage.
• When the question is "difficult to book," Perplexity shifts to Atomix, Carbone, and Tatiana by Kwame Onwuachi — restaurants that dominate the 2024–26 reservation conversation on social and editorial platforms.
• Perplexity's lowest performances are on restaurants with thinner live-web coverage: boutique omakase counters, newer tasting-menu rooms with limited press. Chef's Table at Brooklyn Fare (43 appearances) surfaces through its Michelin documentation despite having no active social campaigns.
• The pattern: Perplexity follows the citation trail. Where that trail is active and well-sourced, it performs. Where the trail goes quiet, it disappears.

Caesar Ramirez's counter in Crown Heights holds a Michelin 3-star. It has 185 normalized response appearances. Claude accounts for 69. ChatGPT contributes 42. Perplexity has 43. Gemini adds 31. Chef's Table doesn't run campaigns. It doesn't dominate the 2026 reservation conversation the way Carbone does. It exists in culinary writing and in the memory of people who have eaten there. Claude's training data includes that writing — and the restaurant's Michelin documentation surfaces it in Perplexity's live-web coverage as well.

Gemini leads Gramercy Tavern (138), The Modern (116), and Ai Fiori (84). That third number is the most telling. Ai Fiori's total is 120. Gemini accounts for 70% of it. ChatGPT has 8. Claude has 8. Ai Fiori is a Midtown Italian fine dining restaurant with real credentials: Michelin recognition, a polished room, a clear culinary identity. It is not in the cultural conversation the way Gramercy or The Modern are. Gemini surfaces it from travel and dining pages where it appears in "best Italian in Midtown" roundups with strong search presence but limited editorial circulation. What Gemini knows often differs from what food critics know. The Gemini map of New York fine dining is partially a search results map.

Every platform has blind spots. The question is whether those blind spots are random or structural. In this dataset, they appear structural. The restaurants AI misses tend to share traits: small, quiet, not aggressively present on booking platforms, without ongoing social or press campaigns. They rely on the kind of knowledge that spreads slowly, through word of mouth and serious culinary writing. That knowledge is in Claude's training data. It is not always in Gemini's search index or Perplexity's live sources.

• Essential by Christophe — 51 total appearances. Claude: 0. Perplexity: 0. Gemini accounts for 73% (37 of 51). The French tasting menu in Midtown is essentially invisible to two of the four platforms — a clean binary split that does not reflect the restaurant's Michelin standing.
• Le Coucou (75 total) — 48 Gemini, 19 ChatGPT, 7 Perplexity, 1 Claude. A celebrated SoHo French brasserie that Claude has essentially never surfaced in its New York fine dining responses.
• Tatiana by Kwame Onwuachi — A top difficult-to-book recommendation in Gemini and Perplexity responses, virtually absent from the overall normalized appearance totals. It surfaces for the right query and disappears from most others. Context-dependent visibility is a different kind of gap.
• The pattern: High-volume restaurants with active media coverage are visible everywhere. Restaurants that earn visibility through critical consensus rather than booking volume have patchy platform coverage. AI does not read the room the way a knowledgeable diner does.

Not everything below the top four is chaos. There are pockets of cross-platform agreement on specific categories where all four machines converge on the same answer. Korean fine dining is the clearest example: Atomix (346), Jungsik (80), and Cote (87) all appear across every platform with meaningful normalized appearance counts. The category is defined clearly enough in the training data, and validated clearly enough in ongoing press, that the machines agree without being forced to. Similarly, Carbone (127) earns cross-platform coverage; every platform surfaces it when the question touches on Italian-American dining or high-energy special occasions. Marea (96) and Aquavit (110) hold moderate cross-platform consensus for seafood and Nordic fine dining respectively. Niche consensus is real. It is also fragile: one platform defecting produces a significantly misleading result for anyone relying on a single source.