AI Chain Bias Distribution 2026:the luxury inversion

This study completes a pair. The cross-platform consensus study found that the five engines barely share individual hotels: 57.6% of top-10 picks existed on one engine only. The natural follow-up is whether they at least share a shape: when you aggregate each engine’s mentions by chain instead of by hotel, do five different hotel lists turn into five different markets, or one?

The answer is one. The engines disagree hotel by hotel and agree almost perfectly in aggregate, on the chain share, on the ranking of chains, and on each chain’s weight. What they converge on is not the hotel industry as it exists. It is a version of the industry where a 135-property luxury operator outweighs a 7,000-property one, and the largest franchisors by property count get very little space.

Each week the landscape pipeline puts an identical 616-prompt library (56 destinations, 11 templates) to the five engines, and every hotel named in an answer is resolved to a canonical ID carrying a parent-chain attribution (sub-brands roll up: a Ritz-Carlton counts as Marriott, a Sofitel as Accor). This study reads the untruncated weekly resolution table, so the shares below cover every resolved mention rather than a top-N cut.

• Window: 2026-05-25 to 2026-09-07, 16 weeks. The longest span where all five live engines have every week. It ends where Perplexity’s scraper stopped, and it predates both the Copilot pause (2026-10-13) and ChatGPT’s capture-vendor change (2026-10-14), so every ChatGPT row comes from one vendor. Grok is excluded: suspended 2026-05-25, zero rows in the window.
• Chain share = chain mentions / (chain + independent mentions). Unresolved mentions (3.5% to 7.2% by engine) and vacation rentals (10 mentions total) sit outside the denominator and are reported in the methodology.
• A “mention” is an appearance in an answer, so a hotel named by three engines in the same week counts three times. That is the right unit for this question: it measures what AI answers show, not what exists.

Every engine gives chain hotels between 42.7% and 46.9% of its resolved mentions. Copilot leans most chain (46.9%), Google AI Mode least (42.7%), and the whole band is 4.2 points wide. For contrast, the consensus study’s hotel-level agreement between engine pairs averaged 26.2% Jaccard. At hotel level the engines disagree; at chain level they look almost the same.

The agreement goes deeper than the aggregate. Within each engine’s chain mentions, the brand weights are nearly identical: Marriott takes 21.1% to 23.4% depending on the engine, and across the twelve biggest chains the largest engine-to-engine gap is 4.9 points (Hilton: Perplexity 11.1%, Gemini 6.2%). No engine has a house chain.

The shared distribution does not resemble the hotel industry. Marriott on top (22.1% of chain mentions) looks conventional; it is also the industry’s largest company by rooms. The inversion starts at position five: Four Seasons, which operates 135 hotels, collects 6,364 mentions, ahead of IHG’s 5,685 from 7,014 hotels. Mandarin Oriental (2,729 mentions) out-ranks Wyndham (903), which describes itself as the world’s largest hotel franchisor by number of properties.

Put company-reported portfolio sizes next to the mention counts and the scale of the inversion is plain:

Group the brands and the pattern generalises. The 13 standalone luxury operators in the data (Four Seasons, Mandarin Oriental, Peninsula, Langham, Aman, Rosewood, Belmond, Capella, Kempinski, Oberoi, Rocco Forte, Leela, Jumeirah) hold 22% of all chain mentions. The 9 economy and budget franchisors present (Wyndham, Choice, Best Western, OYO, Premier Inn, B&B Hotels, Jinjiang, Huazhu, Motel One) hold 2.8% together, fewer mentions than Mandarin Oriental alone (2,424 vs 2,729).

The per-hotel view makes the mechanism visible. Dividing each chain’s weekly mentions by the number of its hotels that earned them gives a depth score: how hard the engines lean on each property they know. Langham averages 12.1 mentions per mentioned hotel per week and Four Seasons 9.5; Marriott runs at 3.8, Hilton 3.1, IHG 2.8 and Wyndham 1.6. A luxury chain’s few properties are named constantly; a big-box chain’s many properties are named occasionally.

Within the chain half of the market, visibility is top-heavy: the top five companies take 56.7% of chain mentions and the top ten take 73.3%, out of 90 chains the resolver can attribute. Marriott’s 22.1% alone exceeds everything outside the top ten combined. The engines reproduce this concentration almost identically: the Herfindahl index of the chain-mention distribution sits between 830 (Google AI Mode) and 973 (Perplexity), with the panel at 888.

Everything above holds the destination mix constant, and across it the engines differ by 4.2 points. Change the destination and the chain share moves 19.2x. In Doha (QA), 80.5% of resolved mentions are chain hotels; in Zermatt (CH), 4.2%. The gradient tracks what each city’s hotel stock and its coverage in the engines’ sources look like: Doha, Cairo, Toronto and the Chinese gateway cities at the top, Alpine and southern-European leisure markets at the bottom.

The volatility study showed each engine keeps only about 29% of its own top-10 hotels from one week to the next. The chain share is the opposite kind of number: the five-engine aggregate stays between 42.5% and 45.7% across all 16 weeks, a band of 3.2 points. Individual hotels rotate constantly through AI answers while the chain composition of those answers barely breathes. The churn swaps a Marriott for a Marriott.

• For independents: the ceiling is higher than the chain-marketing story suggests. Independents hold 55.2% of resolved mentions, a majority, on every engine. The consensus study showed their problem is corroboration (each engine picks different independents); the volume is there.
• For mid-size and economy chains: a brand flag alone buys nothing. 9 economy franchisors share 2.8% of chain mentions. If the engines mention your segment rarely in gateway cities, AI visibility work for such a chain is property-level work, closer to an independent’s playbook than to Marriott’s.
• For luxury operators: the engines are already your distribution. A Langham property is referenced 12.1 times per week on average once the engines know it, three times Marriott’s per-property rate. The weight of guides, awards and press that luxury properties accumulate in the training corpus appears to compound in AI answers.
• For measurement: engine choice barely matters for this metric, so stop segmenting by it. A chain-share tracker needs one engine to approximate all five (4.2pp band); what it must segment by is destination (19.2x spread).
• The lean is parametric. ChatGPT without web search is 2.5 points more chain-leaning than with it. Fixing a chain’s or an independent’s AI visibility through live-web content works against a prior the model carries from training, which is presumably why the distribution moves so little week to week.