Invisible to the Machine: Auditing AI Restaurant, Cafe, and Bar Recommendation Against a Complete Market Census
This paper presents the first census-based audit of AI venue recommendations in Bali, revealing that 85.6% of local restaurants, cafes, and bars are invisible to major AI systems due to a lack of digital documentation rather than poor ratings, while highlighting that the primary failure mode is recommending outdated venues rather than fabricating non-existent ones.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are walking through a giant, bustling city where every single shop, café, and restaurant is listed in a massive, invisible phone book. Now, imagine a new kind of magical guidebook that doesn't just show you a list of options but actually speaks to you, telling you exactly where to go for coffee or dinner. This is what Artificial Intelligence (AI) is becoming for finding local spots. But here's the tricky part: this magical guidebook doesn't show you everything. It picks a few favorites and ignores the rest. Scientists call this "recommendation," but to the shops that get ignored, it feels like being invisible.
For a long time, we've known that if a shop gets a good rating or shows up on a map, more people visit. But we didn't know how this new AI guidebook works. Does it pick the best places? Does it pick the most famous ones? Or does it just pick places that have a website? To find out, a researcher needed to check the guidebook against the entire city, not just a few famous spots. They needed to know: if you are a real, open shop, what are the odds the AI will even know you exist?
The Great AI Restaurant Hunt
A researcher decided to play a giant game of "hide and seek" with four popular AI assistants (ChatGPT, Claude, Gemini, and Perplexity). They chose two busy areas in Bali, Indonesia—Canggu and Ubud—as their playground. First, they did something no one had done before: they created a complete census of every single café, restaurant, and bar in those areas. They counted 4,776 venues. Think of this as making a master list of every single player in the game.
Then, they sent the AI assistants on a mission. They asked the AI 96 different questions, pretending to be different types of people: a digital nomad looking for Wi-Fi, a couple on a date, a family with kids, or a budget backpacker. They asked these questions over seven days, running the test 2,208 times in total. The goal was simple: see which of the 4,776 real shops the AI would actually recommend.
The Shocking Result: Most Places Are Invisible
The findings were startling. Out of the 4,776 real places, the AI assistants never recommended 85.6% of them. That's right: more than 8 out of every 10 places were completely invisible to the AI, no matter how many times they asked. Even among the "established" places—those with at least 50 reviews and a solid reputation—72.6% were never mentioned.
It wasn't that the AI was picking the "best" 10% and ignoring the rest. The places that did get recommended were spread out. The single most popular spot only got about 1.9% of all the recommendations. It wasn't a "winner-take-all" situation where one famous place got everything; it was more like a lottery where most tickets never get drawn at all.
The Two-Step Filter: Getting In vs. Getting to the Top
The researcher discovered that getting recommended by AI happens in two very different steps, like a two-stage rocket launch.
Stage 1: Getting a Seat at the Table (The Entry Margin)
Before an AI can even rank a place, it has to decide to mention it at all. The study found that star ratings don't matter here. A place could have 5 stars or 3 stars, and it didn't change its chances of being picked. Instead, what mattered was documentation.
- Review Volume: Places with more reviews were more likely to be seen.
- Own Website: Having your own website made you nearly twice as likely to be recommended.
- Price Info: Listing your prices helped.
- Web Mentions: Being talked about on other websites helped.
It's as if the AI is looking for a "paper trail." If a place has a website and lots of reviews, the AI says, "Okay, this place exists, I'll put it on the list." If it doesn't have those digital footprints, the AI acts like it doesn't exist, regardless of how good the food is.
Stage 2: Winning the Race (The Ranking Margin)
Once a place is on the list, then the star rating matters. Among the places the AI actually decided to recommend, the ones with higher ratings were more likely to be listed first. So, ratings help you win the race, but having a website and reviews is what gets you into the stadium.
What Didn't Work?
The researcher tested a few ideas that people in the industry thought were important, and they found they were wrong.
- Foursquare Presence: Many people thought that if a place was listed on the open map dataset Foursquare, the AI would find it. The study proved this false. Being on Foursquare didn't help at all.
- Fake Places: People worried the AI would make up fake restaurants (hallucinations). While this happened, it was incredibly rare (only 0.08% of the time).
- The Real Problem: The AI's biggest mistake wasn't making things up; it was recommending places that were permanently closed. It recommended 93 closed venues. The AI wasn't lying; it was just outdated.
The Unstable Guidebook
Finally, the researcher found that the AI is a bit jittery. If you ask the same question twice in a row, you might get a different list of restaurants. If you ask the same question but change the words slightly (paraphrase), the list changes even more. However, when they tested the AI two weeks later, the results were just as jumpy as they were on day one. This means the AI isn't slowly forgetting things over time; it's just naturally inconsistent, like a friend who gives you different directions depending on their mood.
The Bottom Line
This study shows that for a restaurant or café to be seen by AI, having a great rating isn't enough. You need to be "documented" in the digital world first. If you don't have a website, clear prices, or lots of reviews, the AI might as well not know you exist. And for the AI itself, the biggest issue isn't that it's making things up, but that it's recommending places that have already closed their doors. The magic guidebook is powerful, but it's currently blind to the majority of the real world.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.