{"@context":"https://schema.org","@type":"BlogPosting","headline":"AI Hotel Memory 2026: What Do Chatbots Remember About Hotels Without Searching?","description":"A parametric-recall study with web search turned OFF, isolating what cheap LLMs have memorised about hotels rather than what they retrieve. GPT-5.4-nano, GPT-5.4-mini and Gemini 3.1 Flash-Lite were asked, repeatedly and in JSON, to name hotels and return each one's website and address — ~150 runs per model for global chains and ~80 per model for Paris, Dubai, London and New York. Returned domains were checked for DNS resolution, turning recall into a confabulation lie-detector. Findings: hotel chains are known cold — every model returns Marriott/Hilton/Four Seasons with correct websites ~99% of the time. Individual-hotel website accuracy ranges from 97% (Gemini, Dubai) down to 47% (nano, Paris). The signature failure mode is the model knowing a real hotel exists but inventing its web address — Le Bristol Paris returned as the dead bristolparis.com versus the real oetkercollection.com. Paris is the hardest city because its top hotels are independent palaces on unpredictable collection domains. Counter-intuitively the cheapest model, Gemini 3.1 Flash-Lite, had the most accurate and most consistent hotel memory. Total study cost under €20. Method adapted from Dejan Marketing's AI Brand Authority Index, extended with website verification.","datePublished":"2026-06-09","dateModified":"2026-06-09","url":"https://nicolassitter.com/research/ai-hotel-memory-2026","category":"research","keywords":["AI hotel memory","do AI models know hotels","ChatGPT hotel recall","AI hotel hallucination","parametric memory hotels","AI hotel website accuracy","Gemini hotel recall","LLM hotel knowledge"],"articleSection":"Research","wordCount":1900,"readTime":"8 min","articleBody":"June 2026AI Memory\n\n# AI Hotel Memory 2026:What does a chatbot remember about hotels — with the web turned off?\n\nEvery other study here measures what AI _recommends_ after it searches the web. This one measures the opposite: what it has actually **memorised**. We turned web search **off** and asked three cheap models to name hotels — and to give each one’s website. Verifying those websites turns recall into a confabulation lie-detector.\n\n**TL;DR.** Hotel _chains_ are known cold — every model names Marriott, Hilton, Four Seasons and returns their correct website ~99% of the time. For _individual hotels_it cracks: models confidently return a **dead or wrong website 3% to 53% of the time**, worst in Paris. The cheapest model tested (Gemini 3.1 Flash-Lite) had the _best_ memory. And the failure mode is unsettling — the model knows the hotel exists, then **invents its web address**.\n\n### Summarize with AI\n\n0\n\nweb searches\n\npure parametric memory\n\n99%\n\nchain websites correct\n\nevery model knows chains\n\n47%\n\nworst city accuracy\n\nnano, Paris hotels\n\n~1,400\n\ngenerations\n\n3 models · chains + 4 cities\n\nPeople increasingly ask chatbots for hotels. Usually the model searches the web first — so its answer reflects what’s online, not what it knows. We removed that crutch. With tools disabled, an LLM can only answer from the patterns baked into its weights during training: its **parametric memory**. So we asked, repeatedly: _name hotels you know_ — in JSON, with each hotel’s **website and address**. Because a website is checkable, we can do something a pure recall test can’t: measure whether the memory is _correct_, not just present.\n\nThe result is a clean gradient. Chains live in every model’s memory perfectly. Famous palace hotels mostly do too. But the long tail of real, individual hotels is where models start to confabulate — and they do it with total confidence, returning a tidy JSON record for a hotel whose website doesn’t exist.\n\n## 1\\. The experiment\n\nThe design borrows from Dejan Marketing’s [AI Brand Authority Index](https://dejan.ai/blog/brands/), which ranks brands by how often a model names them unprompted. We adapt it to hotels and add one twist that fixes Dejan’s biggest limitation — he could only normalise raw name strings, never verify them.\n\n-   **Web search OFF.** No tools, no retrieval. Pure memory. (In normal use, ChatGPT searches the web — we deliberately don’t.)\n-   **Three cheap models:** GPT-5.4-nano, GPT-5.4-mini, and Google’s Gemini 3.1 Flash-Lite.\n-   **Two cuts:** “name hotel chains” (global), and “name hotels in {city}” for Paris, Dubai, London and New York.\n-   **JSON output:** each run returns `{name, website, address}`. The website becomes the hotel’s identity — no fuzzy matching against a database needed.\n-   **Verification:** we check whether each returned domain actually resolves. A live domain ≈ real memory; a dead one ≈ confabulation.\n-   ~150 runs per model for chains, ~80 per model per city, temperature 1.0, repeated to surface what’s consistently top-of-mind.\n\nAsking for the **website** is the whole trick. “Name a hotel” is unfalsifiable — almost any string could be a hotel somewhere. “Name a hotel _and its website_” forces a checkable commitment, and a model with only a fuzzy memory of a place will reliably invent a plausible-but-wrong domain.\n\n## 2\\. Chains are known cold\n\nThe base layer of hotel memory is rock-solid. Asked for hotel chains, every model returns the majors — and their correct websites — essentially every time. GPT-5.4-mini hit a **99% live-website rate** on chains; even nano managed 88%. Chains appear in training data on a single, predictable domain millions of times, so the association is overlearned.\n\nTop hotel chains by recall (GPT-5.4-mini, 150 runs, web search off). 'Recalled' = share of runs that named the chain. All websites shown resolve.\n\nRank\n\nChain\n\nRecalled\n\nAvg rank\n\nWebsite (verified)\n\n#1\n\nMarriott Hotels\n\n100%\n\n19.1\n\nmarriott.com\n\n#2\n\nHilton Hotels & Resorts\n\n100%\n\n16\n\nhilton.com\n\n#3\n\nHyatt Regency\n\n100%\n\n9.6\n\nhyatt.com\n\n#4\n\nInterContinental Hotels & Resorts\n\n100%\n\n12.7\n\nihg.com\n\n#5\n\nBest Western\n\n100%\n\n22.6\n\nbestwestern.com\n\n#6\n\nRadisson Blu\n\n99%\n\n23.7\n\nradissonhotels.com\n\n#7\n\nRitz-Carlton\n\n98%\n\n15.7\n\nritzcarlton.com\n\n#8\n\nQuality Inn\n\n93%\n\n25.1\n\nchoicehotels.com\n\n#9\n\nFour Seasons Hotels and Resorts\n\n91%\n\n16.2\n\nfourseasons.com\n\n#10\n\nSuper 8\n\n85%\n\n27.4\n\nwyndhamhotels.com\n\n## 3\\. What AI remembers, city by city\n\nDrop to the individual-hotel level and a city’s “memory leaderboard” emerges — the properties a model names again and again, unprompted. These are the hotels with the strongest grip on the model’s mind. Tables below are GPT-5.4-mini; “Recalled” is the share of 80 runs that named the hotel.\n\n### Paris\n\nMost-remembered hotels in Paris (GPT-5.4-mini, 80 runs, web search off).\n\nRank\n\nHotel\n\nRecalled\n\nAvg rank\n\nWebsite (verified)\n\n#1\n\nHôtel de Crillon, A Rosewood Hotel\n\n100%\n\n6.3\n\nrosewoodhotels.com\n\n#2\n\nFour Seasons Hotel George V\n\n100%\n\n3.8\n\nfourseasons.com\n\n#3\n\nMandarin Oriental, Paris\n\n100%\n\n6.6\n\nmandarinoriental.com\n\n#4\n\nShangri-La Paris\n\n100%\n\n5.1\n\nshangri-la.com\n\n#5\n\nHôtel Plaza Athénée\n\n100%\n\n6.1\n\ndorchestercollection.com\n\n#6\n\nThe Ritz Paris\n\n96%\n\n3.3\n\nritzparis.com\n\n#7\n\nHôtel Molitor Paris - MGallery\n\n83%\n\n20.3\n\nall.accor.com\n\n#8\n\nLe Bristol Paris\n\n81%\n\n5.6\n\noetkercollection.com\n\n### Dubai\n\nMost-remembered hotels in Dubai (GPT-5.4-mini, 80 runs, web search off).\n\nRank\n\nHotel\n\nRecalled\n\nAvg rank\n\nWebsite (verified)\n\n#1\n\nBurj Al Arab Jumeirah\n\n100%\n\n8.2\n\njumeirah.com\n\n#2\n\nAtlantis, The Palm\n\n100%\n\n2.9\n\natlantis.com\n\n#3\n\nAddress Downtown\n\n100%\n\n11.3\n\naddresshotels.com\n\n#4\n\nThe Ritz-Carlton, Dubai\n\n100%\n\n10\n\nritzcarlton.com\n\n#5\n\nW Dubai - The Palm\n\n100%\n\n18.5\n\nmarriott.com\n\n#6\n\nHilton Dubai The Walk\n\n99%\n\n22.2\n\nhilton.com\n\n#7\n\nOne&Only The Palm\n\n95%\n\n11.4\n\noneandonlyresorts.com\n\n#8\n\nRaffles Dubai\n\n95%\n\n16.3\n\nraffles.com\n\n### London\n\nMost-remembered hotels in London (GPT-5.4-mini, 80 runs, web search off).\n\nRank\n\nHotel\n\nRecalled\n\nAvg rank\n\nWebsite (verified)\n\n#1\n\nThe Savoy\n\n100%\n\n1.6\n\nthesavoylondon.com\n\n#2\n\nThe Dorchester\n\n100%\n\n5.3\n\ndorchestercollection.com\n\n#3\n\nShangri-La The Shard, London\n\n100%\n\n8.2\n\nshangri-la.com\n\n#4\n\nThe Ned\n\n100%\n\n14.5\n\nthened.com\n\n#5\n\nRosewood London\n\n100%\n\n8.3\n\nrosewoodhotels.com\n\n#6\n\nThe Ritz London\n\n99%\n\n2.6\n\ntheritzlondon.com\n\n#7\n\nClaridge's\n\n99%\n\n3\n\nclaridges.co.uk\n\n#8\n\nThe Langham, London\n\n99%\n\n5.5\n\nlanghamhotels.com\n\n### New York\n\nMost-remembered hotels in New York (GPT-5.4-mini, 80 runs, web search off).\n\nRank\n\nHotel\n\nRecalled\n\nAvg rank\n\nWebsite (verified)\n\n#1\n\nThe Plaza Hotel\n\n100%\n\n1\n\ntheplazany.com\n\n#2\n\nThe St. Regis New York\n\n100%\n\n13.4\n\nmarriott.com\n\n#3\n\nPark Hyatt New York\n\n100%\n\n16\n\nhyatt.com\n\n#4\n\nConrad New York Downtown\n\n100%\n\n12.2\n\nhilton.com\n\n#5\n\nThe Langham, New York, Fifth Avenue\n\n99%\n\n8\n\nlanghamhotels.com\n\n#6\n\nMandarin Oriental, New York\n\n98%\n\n5.8\n\nmandarinoriental.com\n\n#7\n\nFour Seasons Hotel New York Downtown\n\n98%\n\n5.3\n\nfourseasons.com\n\n#8\n\n1 Hotel Central Park\n\n98%\n\n14.9\n\n1hotels.com\n\n## 4\\. The website lie-detector\n\nHere is the headline. For each model and place, what share of the hotels it named came with a website that actually _resolves_? Chains: near-perfect. Individual hotels: a different story — and it gets worse the weaker the model and the more independent the city.\n\nai-hotel-memory-website-accuracy-2026\n\nModel\n\nGlobal chains\n\nParis\n\nDubai\n\nLondon\n\nNew York\n\nGemini 3.1 Flash-Lite\n\n92%\n\n71%\n\n97%\n\n96%\n\n89%\n\nGPT-5.4-mini\n\n99%\n\n58%\n\n94%\n\n92%\n\n82%\n\nGPT-5.4-nano\n\n88%\n\n47%\n\n77%\n\n67%\n\n68%\n\nThe failure mode is the interesting part. The model usually _knows the hotel exists_ — it just fabricates the web address, often a clean, plausible guess that happens to be dead:\n\nHotel (real)\n\nWhat nano invented\n\nThe real website\n\nLe Bristol Paris· Paris\n\nbristolparis.com ✗ dead\n\noetkercollection.com ✓\n\nHôtel Plaza Athénée· Paris\n\nplazaathenee-paris.com ✗ dead\n\ndorchestercollection.com ✓\n\nThe Ritz Paris· Paris\n\ntheritzparis.com ✗ dead\n\nritzparis.com ✓\n\nPod Times Square· New York\n\npod-hotels.com ✗ dead\n\nthepodhotel.com ✓\n\nGPT-5.4-nano returns `bristolparis.com` for Le Bristol Paris — a dead domain. The hotel is one of the most famous in the world; its real site is`oetkercollection.com`. The model didn’t fail to recall the hotel — it recalled the hotel and **hallucinated the URL**. That is exactly the kind of confident-but-wrong detail that slips past a reader.\n\n## 5\\. Why Paris is the hardest city\n\nParis is the worst-remembered of the four cities for every model (down to 47% live-website rate on nano), and it also produces the _longest tail of invented names_ — nano emitted 469 distinct “Paris hotels” across 80 runs, versus a tight 177 for Gemini. The reason is structural: Paris’s top hotels are independent palaces on collection domains the model can’t predict — Le Bristol on `oetkercollection.com`, Plaza Athénée on `dorchestercollection.com`. A chain trains the model on one obvious domain; an independent does not, so the model guesses — and misses.\n\nai-hotel-memory-invented-tail-2026\n\nA tighter set with more live websites (Gemini) signals a sharper, more reliable memory; a sprawling set with dead domains (nano) signals a model padding its answer with invention.\n\n## 6\\. The cheap-model surprise\n\nThe counterintuitive finding: the **cheapest** model had the best hotel memory. Google’s Gemini flash-lite returned the highest share of working websites in cities — **97% in Dubai, 96% in London**— beating the pricier GPT-5.4-mini and far ahead of GPT-5.4-nano. It’s not about price; it’s about how much specific, rare detail a model retains. This is also why Dejan’s original brand index could run on cheap Gemini at all: that family punches above its cost on real-world entity recall. Two cheap models, two very different memories.\n\n## 7\\. What it means for hotels\n\n-   **If you’re a chain or a famous palace**, the model knows you — name and website. You’re in the memory layer, not just the search layer.\n-   **If you’re an independent or boutique hotel**, the model may know your name but invent your web address. When a model answers from memory (or a user copies its output), that’s a wrong link pointing away from you.\n-   **Memory ≠ retrieval.** With web search on, models ground their answers and this mostly disappears. But the memory layer still shapes which hotels a model reaches for first, and what it “believes” before it searches.\n-   **The fix is the same as the rest of AI visibility:** a consistent, well-linked web presence is what turns a hotel from a fuzzy memory into a correctly-remembered one.\n\n## Methodology\n\n**Models:** gpt-5.4-nano, gpt-5.4-mini (OpenAI), gemini-flash-lite-latest — which resolved to Gemini 3.1 Flash-Lite (Google), called with no tools and temperature 1.0. **Runs:** ~150 per model for the chains cut; ~80 per model for each of Paris, Dubai, London, New York. **Output:** JSON, requesting name + website + address per hotel.\n\n**Identity & scoring:** hotels are keyed by their returned website domain (falling back to a normalised name). “Recalled” is the share of runs that named the hotel; “avg rank” is its mean position in the list. **Website verification:** each distinct domain is checked for DNS resolution. A resolving domain is treated as a (conservative) signal of real memory; a non-resolving one as confabulation.\n\n**Caveats:** DNS-resolves is a lower bound on accuracy — a domain can resolve yet still be the wrong hotel, so true confabulation is somewhat higher than reported. This measures the memory of three specific cheap models, not the full ChatGPT or Gemini consumer products (which use larger models and web search). Cost of the entire study was under €20.\n\n## FAQ\n\n[← All research](/research)","author":{"@type":"Person","name":"Nicolas Sitter","url":"https://nicolassitter.com/about","sameAs":["https://www.linkedin.com/in/nicolassitternolleau/","https://github.com/Nicositter88","https://hotelrank.ai"]},"publisher":{"@type":"Person","name":"Nicolas Sitter","url":"https://nicolassitter.com"},"image":"https://nicolassitter.com/api/og/ai-hotel-memory-2026","mainEntityOfPage":{"@type":"WebPage","@id":"https://nicolassitter.com/research/ai-hotel-memory-2026"},"tags":["AI Search","AI Memory","Hotels","Confabulation","GEO"],"sameAs":["https://hotelrank.ai/research/ai-hotel-memory-2026"],"alternateFormat":{"html":"https://nicolassitter.com/research/ai-hotel-memory-2026","json":"https://nicolassitter.com/api/post/ai-hotel-memory-2026","rss":"https://nicolassitter.com/rss.xml"},"datasets":[{"name":"summary","contentUrl":"https://nicolassitter.com/data/ai-hotel-memory-2026/summary.csv","encodingFormat":"text/csv"}]}