Why you got the same five landmarks

You asked for trip ideas and received the same famous cities and sights you already knew. Independent studies show that pattern as concentration toward destinations that were already visible.

Updated 7 October 2026

You asked for ten German Christmas markets, or the best walking tours in a city you have already booked, and the answer felt like a postcard rack.

Nürnberg, Dresden, the old town square, the museum everyone photographs. The assistant may even describe source triangulation, checking one claim against several sources, or quality checks, its own steps for judging a source. It still names the same famous places.

Recommendations for places follow the same pattern as a product recommendation, with different data behind them and a documented bias toward famous destinations. For the overall model, see How AI chatbots build a shortlist.

A small ask, a narrow canon

Dirk H. R. Spennemann ran fifteen repeats of one open prompt on ChatGPT 5.2 over five days: “Recommend me a list of 10 German Christmas Markets.” Against a real pool of roughly two thousand markets, the model kept returning a small set of iconic names: Nürnberg, Dresden, Köln, München, and Stuttgart. Reasoning traces showed reliance on tourism marketing, travel media, and blogs rather than neutral sampling. The paper is Assumptions and Undeclared Selection Criteria in Administrative Sciences (2026).

That case is easy to explain out loud. It is also a stand-in for broader travel asks where “best in Europe” collapses to the same handful of cities. For products, the same concentration on a blank ask is in Ask for a white t-shirt and you get Uniqlo.

Audits that measure the bias

Andreev, Kosmas, Livieratos, Theocharous, and Zopiatis audited ChatGPT-4o and DeepSeek-V3 with 216 traveler personas and 6,480 destination recommendations (Destination (Un)Known, MDPI AI, 2025). They report measurable bias across popularity, geography, culture, stereotypes, demographics, and reinforcement. Both models still favor mainstream destinations even when personas differ.

That study names ChatGPT-4o and DeepSeek-V3. Unconstrained travel prompting is not neutral in those two systems. The paper does not measure Perplexity, Gemini, or Claude.

A 2026 paper in Current Issues in Tourism (digital overtourism, DOI 10.1080/13683500.2026.2654066) analyzed 420 recommendations from ten AI systems across fourteen tourism queries. The authors coin digital overtourism: algorithmic concentration of destination visibility before anyone travels. A list can contain many names while attention still repeats iconic and semi-iconic places. The authors call the long list high nominal variety and the repeated famous places low effective diversity. Sustainability-flavored prompts in their sample often changed wording more than spatial spread.

What asking can change

Joseph Mellors (ChatGPT and the tourist trail; plain-language summary) found that lesser-known or more sustainable locations tended to appear when travelers explicitly asked for them, including off-peak timing and sharper constraints. That connects back to whether a vague ask pulls the tourist list: concentration is partly an ask-side problem, not only a model bug.

A Barcelona study (The Concentrated City) compared heritage sites tourists geotagged on Instagram with ChatGPT recommendations and found the model’s picks even more spatially concentrated than those crowdsourced visits. Useful as supporting color for the same pattern.

If you are the person asking

Overlap across tools and repeated famous names are a reason to widen the ask with a season, a neighborhood, a budget, or “not the cruise-ship circuit,” and still verify hours and access on the ground.

If you promote a place or experience

Fame in the training and retrieval pool still matters. A walking tour with three names on three sites will lose to landmarks unless the assistant can treat it as one place across Maps, your site, and reviews.