Algorithmic monoculture is when a whole industry relies on the same AI models and converges on the same generic answers, and the easiest place to watch it happen is dinner. When Oberhahn asked four AI engines where to eat in New York, a tiny handful of famous restaurants dominated every answer, even when the question begged for something obscure.
We picked restaurants on purpose. Anyone can check the answers, no one has a stake in the outcome, and the pattern that shows up there is the same pattern that shows up when your team asks AI how to run your business.
The study: Where AI Says to Eat in New York
In August 2026, Oberhahn ran a measurement wave called Where AI Says to Eat in New York, the first in a series we call NYCEATS. The setup was simple. We asked 18 diner questions across four AI engine configurations, which produced 108 answers in total. Those answers named 372 distinct restaurants across 693 total mentions.
The questions were the kind a real person asks. Some named a place directly. Most just described a need: a spot for a first date, somewhere quiet for a work dinner, a place off the beaten path. That mix let us see two very different things: what the engines say when you tell them the answer, and what they recall when you only describe what you want.
Finding one: there is no single AI answer, and that is the problem
Here is the result that surprises people. 82% of the restaurants, 305 of 372, were named by exactly one engine. Only 10 restaurants were named by all four.
At first glance that looks like diversity. It is the opposite. The engines do not share one answer. They share a handful of famous names and then scatter into largely separate long tails. The overlap is not broad agreement on good restaurants. It is narrow agreement on the obvious ones.
The consensus list, the restaurants named by all four engines, is exactly the set you would guess: Katz's Delicatessen, Peter Luger, Carbone, and Lucali. The icons. And a small group carries enormous weight: six household names account for 12% of all mentions. A tiny set of defaults dominates the conversation while everything else appears once and vanishes.
Finding two: the defaults survive even when you ask them not to
We deliberately included prompts designed to break the pattern. Questions like "somewhere nowhere near Instagram-famous," "I'm bored of the famous places," and "help me beat the icons."
The engines returned the same famous names anyway. Told explicitly to avoid the obvious, they reached for Carbone and Katz's regardless. This is the tell. The default is not a suggestion the model offers when it has nothing better. It is the gravitational center it returns to even when instructed to leave.
Finding three: known is not the same as recalled
This is the finding that turns a fun study into a serious one. We ran a worked example with a genuinely differentiated restaurant, Ugly Baby, a Thai spot that is exactly the kind of non-obvious answer a good recommendation engine should surface.
When we named Ugly Baby outright, the engines recognized it every time: 24 of 24 answers. Found by name, 100%. The model clearly knows the restaurant, knows it is good, knows why.
When we only described what a diner wanted and let the engine choose, Ugly Baby appeared in just 5 of 84 answers. Found by need, 6%.
The engines know the differentiated option. They rarely recall it unprompted. That gap between what a model knows and what it volunteers is algorithmic monoculture in a single number. The good, distinct answer exists inside the machine. The machine hands you the famous one anyway.
Finding four: this was mostly decided at training time
One more result matters for anyone hoping that fresher content will fix this. The engines named 116 distinct restaurants purely from memory, with no internet access at all. Giving an engine live web search widened the list, 150 versus 116, but the core names stayed the same. The famous defaults did not move.
A lot of what AI recommends was decided at training time. Adding fresh information broadens the edges of the list without dislodging its center. In the short term, you do not out-publish the default. You have to be deliberately surfaced past it.
What this means for your business
Now swap the diner for your team. Replace "where should I eat" with "how should we solve this problem." Every finding transfers.
Everyone using the same model gets handed the same famous, generic, most-likely answer. Even when you ask for something novel, you get the default. The differentiated approach exists inside the model, the same way Ugly Baby did, but it is rarely recalled unless someone names it outright. So a whole industry, all prompting the same models, converges on the same playbook without anyone deciding to. That is algorithmic monoculture, and it is why margins erode.
The restaurant study is a proving ground. The same measurement method applies to any domain, including how your company uses AI. That is exactly what the Advantage Index does. It scores how far your organization's AI usage sits from the generic default, which is the same as scoring how different you are from every competitor who accepted the default the model handed them for free. Oberhahn does not need to see inside your competitors, because the default is public. It is the answer everyone on that model already gets.
Frequently Asked Questions
Can you give a simple example of algorithmic monoculture?
Yes. When Oberhahn asked four AI engines where to eat in New York, 82% of the 372 restaurants named were named by only one engine, while a handful of icons like Katz's, Peter Luger, Carbone, and Lucali showed up in every answer. The AI converges on the same famous defaults, which is algorithmic monoculture.
Why do AI engines recommend the same famous places?
Because a model returns the most likely answer, which is the most-mentioned answer in its training data. Even prompts that explicitly ask for obscure or non-obvious options returned the same icons in Oberhahn's study, showing the default is where the model naturally lands.
What does the Ugly Baby finding prove?
It proves that AI knows differentiated options but rarely recalls them unprompted. The restaurant appeared in 24 of 24 answers when named outright but only 5 of 84 when the diner described only a need. That gap between known and recalled is the core mechanism of algorithmic monoculture.
Does giving AI internet access fix the problem?
Not at its core. In Oberhahn's study the engines named 116 restaurants from memory with no internet and 150 with web search, but the dominant famous names stayed the same. Much of what AI recommends is fixed at training time, so fresh content does not dislodge the defaults short term. Read the full framework on our algorithmic monoculture page, or see how Organizational Intelligence applies the same measurement to your teams. Measure your Advantage Index with Oberhahn.