AI Hotel Memory

What do AI models actually remember about hotels with the web turned off? Pick a place and a model below to see the hotels it names from memory — and whether it knows their real website (verified live). Toggle to the weakest model to watch the memory degrade.

Read the full study

Model

71%

named hotels with a working website

177

distinct hotels named (80 runs)

Mostly real, but a notable share of websites are fabricated.

#	Hotel	Recalled	Website (verified live)
1	Hôtel Ritz Paris	100%	ritzparis.com
2	Hôtel Plaza Athénée	100%	dorchestercollection.com
3	Hôtel de Crillon	100%	rosewoodhotels.com
4	Four Seasons Hotel George V	100%	fourseasons.com
5	Shangri-La Paris	100%	shangri-la.com
6	The Peninsula Paris	100%	peninsula.com
7	Park Hyatt Paris-Vendôme	100%	hyatt.com
8	Saint James Paris	100%	saint-james-paris.com
9	Hôtel Lutetia	100%	hotellutetia.com
10	Cheval Blanc Paris	100%	chevalblanc.com
11	Hôtel Madame Rêve	100%	madamereve.com
12	Le Bristol Paris	99%	oetkercollection.com
13	Hôtel d'Aubusson	98%	hoteldaubusson.com
14	Mandarin Oriental, Paris	94%	mandarinoriental.com
15	Hôtel Costes	88%	hotelcostes.com

“Recalled” = share of 80 memory-only generations that named the hotel. A red, struck-through domain is a website the model returned that doesn’t resolve — a fabricated address for a (usually real) hotel.

Caught inventing the website

The signature failure: the model knows a famous hotel exists, then makes up a plausible-but-dead web address.

Le Bristol Paris

Paris · GPT-5.4-nano

bristolparis.com

oetkercollection.comreal

Hôtel Plaza Athénée

Paris · GPT-5.4-nano

plazaathenee-paris.com

dorchestercollection.comreal

The Ritz Paris

Paris · GPT-5.4-nano

theritzparis.com

ritzparis.comreal

Pod Times Square

New York · GPT-5.4-nano

pod-hotels.com

thepodhotel.comreal

Pilot data: 3 models, global chains + Paris/Dubai/London/New York, web search off, websites verified by DNS. Full methodology & findings →