When a whole industry asks the same AI models the same questions, everyone converges on the same generic answers, and the differences that let a company charge a premium quietly disappear. We call that algorithmic monoculture, and this year we ran a study that shows exactly how it forms.
I want to walk you through what we found, because the study was not about restaurants. It was about your company.
Why we asked AI where to eat
We needed a neutral proving ground. Nobody has a financial stake in which taco joint an AI recommends, and anyone can check the answers against their own experience. So in August we ran a measurement wave we called "Where AI Says To Eat in New York." We posed 18 diner questions across four AI engine configurations. That produced 108 answers, 693 total mentions, and 372 distinct restaurants named.
I expected variety. Four engines, 18 different questions, a city with thousands of restaurants. The results should have sprawled. Instead they collapsed toward a handful of names, and the way they collapsed is the whole lesson.
The consensus is smaller than you think
Here is the number that reframes everything. Of the 372 restaurants named, 305 were named by exactly one engine. That is 82 percent. Only 10 restaurants were named by all four engines.
There is no single AI answer. There are largely separate answers that happen to agree on a tiny shared core. And that shared core is exactly what you would guess: Katz's Delicatessen, Peter Luger, Carbone, Lucali. The obvious icons. The famous four showed up no matter what we asked.
We tried to break it. We wrote prompts that explicitly demanded the opposite of famous. "Nowhere Instagram-famous." "I'm bored of the famous places." "Beat the icons." The engines returned the icons anyway. Six household names carried 12 percent of every mention we collected. A market of thousands of restaurants, and a tiny set of defaults dominated the conversation.
The known-versus-recalled gap
The sharpest finding is the one that maps most directly onto business, so I want to give it a name. Call it the known-versus-recalled gap.
We ran a worked example with a specific restaurant, Ugly Baby, a differentiated Thai place that a knowledgeable local would recommend. When we named it outright and asked the engines about it, it appeared in 24 of 24 answers. A hundred percent. The engines clearly know it.
Then we stopped naming it. We described only what a diner wanted, the flavor profile, the vibe, the neighborhood, and let the engine choose. Ugly Baby appeared in 5 of 84 answers. Six percent.
The engine knows the differentiated option. It rarely recalls it unprompted. Found by name, 100 percent. Found by need, 6 percent. That gap between what the model knows and what it volunteers is where algorithmic monoculture lives.
What this has to do with your margins
Now do the swap. Replace the diner asking where to eat with your team asking AI how to solve a problem. Replace "the famous four" with the most likely playbook in your industry.
Every competitor prompting the same model gets handed the same famous, generic, most-likely answer. Ask for something novel and you still get the default, because the default is what the model reaches for first. The differentiated approach exists inside the model, the same way Ugly Baby did, but it stays there unless someone already knows to name it. So a whole industry, all prompting the same models, drifts onto the same generic strategy at the same time.
Interchangeable output kills the premium. When your product, your positioning, and your operating decisions look like everyone else's, buyers have no reason to pay more for yours. Margins erode across the board, and no single quarter is where you can point to the cause.
The training-time problem
One more finding, because it predicts how long this lasts. We ran the engines with no internet access, from memory alone, and they still named 116 distinct restaurants. Giving them live web search widened the list to 150, but the core names did not move. The icons stayed the icons.
A large share of what AI recommends was decided at training time. Publishing fresh content does not dislodge it in the short term. For a business, that means you cannot out-market your way to distinctiveness inside the model. The default is sticky. If your industry's default becomes the model's default, buying more content and more ads mostly reinforces the same middle everyone already occupies.
How to know where you actually stand
The restaurant wave used a measurement method, and the same method applies to any domain, including how your own company works with AI. That is what the Advantage Index measures. High means your AI usage sits far from the generic default, still distinct and still defensible. Low means you have drifted into the generic middle with everyone else. The sentence you do not want to hear from your own dashboard is simple: our Advantage Index is falling.
People ask how we can score your distinctiveness without spying on competitors. We do not need to. The model's most-likely answer is the answer every competitor using that model already gets for free. Measuring how far you sit from the generic default is the same as measuring how different you are from everyone who took the default. The restaurant study is the proof: the default is knowable, it is stable, and it is shared.
My prediction is direct. Within a few years, the companies that treated AI as a source of the most likely answer will look interchangeable to their own customers, and they will spend the following decade trying to explain why their premium disappeared. The companies that measured the gap early, and defended the approaches sitting far from the default, will be the ones still charging more.
Frequently Asked Questions
Does AI give every company the same competitive advantage?
No. AI hands every company the same most-likely answer, which is the opposite of an advantage. Our restaurant study found that 82 percent of recommendations came from a single engine and only 10 names were shared by all four, meaning the generic core is small and everyone converges on it. Real advantage comes from the differentiated approaches the model knows but rarely recalls on its own.
What did the Oberhahn restaurant study actually measure?
We asked four AI engine configurations 18 diner questions about where to eat in New York, producing 108 answers and 372 distinct restaurants. We measured how often each restaurant was named, how much the engines agreed, and how often a differentiated option surfaced when it was described by need rather than named outright.
What is the known-versus-recalled gap?
It is the difference between what an AI model knows and what it volunteers. In our study, the restaurant Ugly Baby appeared in 100 percent of answers when named and only 6 percent when described by need. Applied to business, the differentiated approach exists in the model but is rarely recalled unprompted, which is how algorithmic monoculture forms.
How does Oberhahn measure this without seeing my competitors?
The model's most-likely answer is the answer every competitor using that model already receives. Measuring how far your AI usage sits from that generic default is the same as measuring how different you are from everyone who took the default. The Advantage Index scores that distance. See how your organization actually works with AI on the Organizational Intelligence page, or request a demo to measure your Advantage Index.