AI Cross-Platform Consensus 2026:five engines, five different hotel lists

This closes a triptych. The rankings-consistency study measured how much a single engine disagrees with itself minutes apart (50.5% position-1 stability); the volatility study measured how much it disagrees with itself week to week (29.2% top-10 retention). The remaining axis is engines disagreeing with each other at the same moment on the same questions — and it turns out to be the widest gap of the three.

An average recommendation that week was backed by 1.83 of the five engines. In practice that means a traveler who switches assistants gets a mostly new shortlist, and a hotel that monitors only ChatGPT is blind to the majority of its AI surface area. The rest of this piece walks through where the agreement concentrates: which engine pairs, which hotel types, which cities.

The landscape pipeline puts the same 616-prompt library (56 destinations × 11 templates) to each engine every week and entity-resolves every recommended hotel to a canonical ID. For this study I pinned the week of 2026-08-03 — the same week for all five engines, asserted by the analysis script — assigned each hotel to its destination by geo bounding box, and took each engine’s ten most-mentioned hotels per city. That yields up to 280 city-engine lists (26 come up shallower than ten hotels upstream), which flatten into 1,490 distinct (city, hotel) recommendations to intersect.

• “Agreement” is set membership, deliberately generous: two engines agree on a hotel if both place it anywhere in their top 10 for that city, regardless of position. The rank-level bar is much higher — section 7 shows what survives it.
• Grok sat this one out. Its most recent usable data is the week of 2026-05-18, eleven weeks before the pinned week (the scraper has been suspended since 2026-05-25). Comparing a May snapshot against August ones would manufacture disagreement, so the panel is the five engines with current data. The freshness table in the methodology documents this.
• One week is a snapshot of a moving target. Per the volatility study, an engine keeps only ~29.2% of its own top 10 from one Monday to the next, so the specific hotels named here will rotate. The disagreement level is the durable finding; treat the hotel names as examples from 2026-08-03.

Bucket every recommendation by how many engines make it and the distribution collapses toward one: 858 of the 1,490 city-hotel picks (57.6%) live on a single engine’s list. Each additional engine roughly halves the bucket — 18.9% shared by two, 11.2% by three, 7.7% by four — down to the 69 recommendations (4.6%) that all five engines endorse.

Pairwise, the disagreement has a clear geometry. The closest pair is Copilot × Google AI Mode at 34.8 average Jaccard — the only pair sharing more than half of each other’s picks by the containment measure (51.2%). Both lean on Bing/Google-flavored web retrieval, and it shows. At the far end sits ChatGPT × Perplexity at 16.4 — two chat assistants whose hotel taste has less in common than any other combination. Gemini pairs most naturally with AI Mode (31.2), which is intuitive for two Google surfaces, though even that overlap is under a third.

Averaging each engine’s Jaccard against the other four turns the matrix into a personality ranking, and Perplexity finishes last by a wide margin: 252 of its 550 top-10 picks (45.8%) appear on no other engine’s list for the same city. ChatGPT is a distant second at 33.4%; Google AI Mode is the most consensual engine at 22.7%. This rhymes with what the volatility data showed about Perplexity’s relationship with its own past picks — the engine that re-rolls its list weekly also strays furthest from the pack at any given moment.

Split the 1,490 recommendations by whether the hotel carries a chain brand and the consensus concentrates visibly. A chain-branded pick is backed by 2.06 engines on average against 1.65 for an independent; half of chain picks (50.5%) have at least one corroborating engine versus 36.4% for independents; and at the all-five bar the gap is 6.6% to 3.2% — call it 2.1×. Independents supply most of the raw recommendations (850 of 1,490) but they overwhelmingly populate the single-engine tail.

Averaging the ten pairwise Jaccards per destination spreads the 56 cities across a 13× range. Auckland tops the table at 49.6%: its five top-10s contain just 19 distinct hotels, 5 of which every engine lists. Cairo (46.9), Singapore (42.0), Queenstown (39.9) and Sydney (38.9) follow — compact markets where one obvious luxury shortlist exists and every engine finds it. At the bottom sits Bali at 3.8%: fifty top-10 slots holding 44 different hotels, with not a single property on all five lists. Phuket (10.3), Paris (11.8) and Prague (13.2) keep it company — huge, fragmented leisure markets where each engine assembles its own canon.

The number that surprised me most: 17 of the 56 cities have no all-five hotel at all, and the list reads like a greatest-hits of world tourism — Paris, London, Rome, Tokyo, Bangkok and Bali are all on it. The deeper and more written-about a market, the less the engines can settle on a shared answer for it.

So who actually clears the bar? The 69 all-five hotels are spread across 39 cities, and 42 of them carry a chain brand. Then there is the question the whole study builds to: does any hotel hold rank #1 on every engine? Exactly one does. The Taj Mahal Palace in Mumbai is the most-mentioned hotel for its city on ChatGPT, Gemini, Perplexity, Copilot and Google AI Mode alike — a clean 1-1-1-1-1 in a dataset where five-way agreement on anything is rare. Volatility readers may recognize it: it was also one of the few hotels that never left a top-100 on two engines across that study’s eleven weeks.

The nearest misses are instructive too. Hotel Monteleone in New Orleans goes 2-1-1-1-1, denied unanimity only by ChatGPT’s #2. Hyatt Regency Cape Town (1-1-3-1-1) and Mandarin Oriental Wangfujing Beijing (1-3-1-1-1) each take four firsts, and Electra Palace Athens opens with three. Household names appear further down — Marina Bay Sands, Copacabana Palace, Royal Mansour Marrakech — but ranked differently everywhere, which is the study in miniature.

• For hotels: audit all five engines before believing any of them. With 57.6% of recommendations single-engine, your ChatGPT visibility says almost nothing about your Gemini visibility. The pairs to treat as semi-redundant are Copilot & AI Mode (34.8 Jaccard) and Gemini & AI Mode (31.2); Perplexity needs its own check every time.
• Corroboration is the quality signal. A pick backed by three or more engines (23.5% of the total) sits in a different tier from a single-engine special — especially a Perplexity-only one, given its 45.8% uniqueness rate.
• For travelers, a second assistant is nearly a second opinion by construction. The average overlap between two engines’ top-10 lists for the same city is 26.2% (mean pairwise Jaccard across the ten pairs). If you want variety, switch engines; if you want reliability, look for the hotels they name in common.
• Independents compete engine by engine. The consensus layer is 42/69 chain-branded; an independent’s realistic goal in 2026 is being the darling of one or two engines and knowing which ones they are.
• In fragmented markets, city-level AI rank is close to meaningless. A Bali or Paris hotel can be #1 on one engine and invisible on four. In Auckland-type markets the shortlist is shared, and cracking it means displacing a hotel all five engines already agree on.