The lodging leaderboard
Every property the engines named, resolved against a 74-place Google Maps registry of the Oisans (55 real hotels and B&Bs). The race splits the podium: the valley start town takes #1, the mountain finish takes the next four.
| # | Hotel | Town | Mentions | Engines |
|---|---|---|---|---|
| 1 | Hôtel Oberland | Bourg-d’Oisans | 29 | 3/5 |
| 2 | Royal Ours Blanc | Alpe d'Huez | 24 | 4/5 |
| 3 | Le Castillan | Alpe d'Huez | 23 | 4/5 |
| 4 | Grandes Rousses Hotel & Spa | Alpe d'Huez | 22 | 4/5 |
| 5 | Club Med Alpe d’Huez | Alpe d'Huez | 20 | 3/5 |
| 6 | Hôtel de Milan | Bourg-d’Oisans | 19 | 3/5 |
| 7 | Hôtel Le Pic Blanc | Alpe d'Huez | 19 | 4/5 |
| 8 | B.O. Lodge | Bourg-d’Oisans | 16 | 4/5 |
| 9 | Au Chamois d’Or | Alpe d'Huez | 10 | 3/5 |
| 10 | Hotel Au Bon Accueil | Les Deux Alpes | 10 | 3/5 |
Mentions = recommending answers across all 5 engines, EN + FR, name variants merged (“Hotel Oberland” = “Hôtel Restaurant L’Oberland”). #10 sits in Les Deux Alpes, one valley over — the engines’ enthusiasm doesn’t always match the road book.
Most-recommended hotels — popups show per-engine recommendations
What the engines actually named
901 lodging mentions, sorted into what they really were. The verifiable-hotel core is barely a quarter of the answer stream — the rest is towns (“stay in Vaujany”), packaged cycling holidays, and names we couldn’t match to any geocoded lodging around the route.
“Unverifiable” is a floor, not a hallucination rate: it includes real chains further afield (Grenoble, Vizille) and properties our conservative exact/alias matching missed — but also names we could not find anywhere.
ChatGPT booked you into last year’s race
26 recommendations in our capture set place riders in Albertville and La Plagne — ibis Styles Albertville (10 mentions), Base Camp Lodge, Fasthôtel, even Club Med La Plagne. That geography is perfect… for the 2025 edition, which ran Albertville → La Plagne. It is 100 km from this year’s start. Every single one of those 26 mentions came from ChatGPT — the other four engines never made the mistake. When the event is annual and the route moves, an engine answering from memory instead of the live web sends you to the wrong valley with a non-refundable booking.
The pattern is one we keep measuring: what a model remembers and what the web currently says are different retrieval paths, and lodging queries around dated events are where the seams show. (See our index-vs-live-web study for the hotel-search version of this.)
Two things we didn’t expect
The engines answer in towns, and hedge across the whole valley
76 recommendations were not properties at all but places: Les Deux Alpes (11), Vaujany (9), Oz-en-Oisans (9), Auris, Allemond, Grenoble, Vizille. Sensible base-camp advice for a sold-out race weekend — but nobody can book “Vaujany”. If you run a hotel in one of these villages, the AIs are recommending your town while leaving the actual booking to whoever shows up first in the next query.
Tour operators outrank most hotels
Sportive Breaks (10), Love Velo (9), Sports Tours International (8) and 44|5 Cycling Tours (6) all scored more mentions than most actual hotels — the engines like recommending the package (entry + bed + transfer) over the bed. For event lodging, the aggregator layer we keep finding in city studies is a tour-operator layer here.
The Airbnb-and-campsite gap
Ask anyone who actually rode: L’Étape weekend around the Oisans runs on Airbnbs, campsites and gîtes as much as hotels. The engines barely acknowledge that layer. Across all 901 mentions, Airbnb-style platforms got 11 and campsites, hostels and gîtes together got 16. Even on our four deliberately neutral prompts (“where to stay”, “which town”, “last-minute places to stay”, “chalets or lodges”) — 219 mentions where anything could have been named — the alternative-lodging layer got just 12 mentions, about 5%. Ten of our 14 prompts did say “hotels”, so the panel leans hotel-first by design — but when given the open question, the engines still answered hotels, towns and tour packages. The way a big share of riders actually sleeps on race weekend is close to invisible in AI answers.
Five engines, five personalities
Gemini talks the most and scatters the widest; Perplexity says the least. Copilot names almost as many distinct places as ChatGPT while producing half the volume.
| Engine | Lodging mentions | Distinct names |
|---|---|---|
| Gemini | 338 | 155 |
| ChatGPT | 246 | 136 |
| Google AI Mode | 139 | 74 |
| Copilot | 102 | 86 |
| Perplexity | 76 | 63 |
The 74 geocoded lodging places around the route
How we ran it
- Panel: 14 lodging templates — hotels near the start/finish, cyclist-friendly, bike storage, budget, spa recovery, early breakfast, family, chalets, last-minute — in English and French. The event is named, the towns never are: the engines had to know the route themselves.
- Engines: ChatGPT, Perplexity, Gemini, Copilot, Google AI Mode. 252 captures in the days before the race (AI Mode: 28; the others: 56 each).
- Mentions: lodging names extracted from each answer with a named-entity-recognition pipeline — 901 mentions.
- Registry: 74 geocoded places from Apify Google Maps around Alpe d’Huez, Le Bourg-d’Oisans and the surrounding Oisans villages, 55 classified as real hotels/B&Bs/resorts.
FAQ
Summarize with AI
