Walk into an ice cream shop offering 31 flavors. You look at the counter, think for a moment and make your choice. Now imagine asking an AI assistant what to order before you even see the options.
It suggests two flavors. Ask again and you get two suggestions. Someone else asks and the same favorites keep coming back.
The other 29 are still on the menu. Most people simply never reach them.
That is the kind of scenario explored by Sandra Matz, a researcher at Columbia Business School, in her TEDxNewEngland talk, Is AI turning us all into the same person?, delivered in October 2025.
Matz says she asked ChatGPT for ice cream flavor recommendations 100 times. In 96 of those responses, according to her demonstration, the answer selected one of two popular flavors.
Now replace ice cream with a law firm, a marketing agency, enterprise software, clothing or an industrial supplier.
What happens to a business that never makes the AI-generated shortlist?
The research behind the question
The ice cream example gets our attention. The study behind the talk goes further.
In an academic preprint available on arXiv, Matz and her coauthors worked with 110,000 choices made by 1,000 US users. They compared patterns in human preferences with choices produced by agents powered by large language models.
The agents showed a greater tendency to select popular alternatives. Under different experimental conditions, people’s choices became more similar. Giving agents information to personalize their suggestions reduced some of the convergence between different users, while the range of interests explored within an individual profile could also narrow.
Think of your own life. You have favorite musicians, but once in a while you discover someone new. You return to the same restaurant, until a walk takes you past a place you have never tried. You prefer a familiar brand, then a friend introduces you to another.
The research raises an important possibility: when we routinely delegate our decisions to systems optimized to generate plausible recommendations, some of those discoveries may disappear.
| What the researchers observed | What it could look like in everyday decisions |
|---|---|
| A preference for popular options | The same widely recognized brands are repeatedly suggested. |
| Greater similarity between people’s choices | Different customers receive similar recommendation lists. |
| Less variety within individual profiles | People explore fewer options outside familiar interests. |
There is a potential economic consequence. If millions of decisions start with short lists, being included on those lists becomes commercially valuable.
The brand that never makes the list
Try a simple exercise. You need to hire a provider and ask an assistant: “Which three companies would you recommend for this service?”
The AI returns three names with a short explanation. You visit their websites, compare prices, review their reputations and get in touch.
What about provider number four? Perhaps its team is better, its pricing is competitive and it has experience in your region. Yet it never entered the conversation.
The customer may still reject all three suggestions or find an alternative through a referral. Even so, the initial selection has already given the recommended companies a head start.
Being available in the market and being considered by a buyer are two different stages.
A business can invest in quality, build a strong product and develop a capable sales operation. As customers begin their journey with a ready-made answer, discovery also depends on which information the system can find, interpret and choose to present.
That could change how companies think about digital visibility, reputation, differentiation and competition, particularly businesses still building recognition.
The market has seen this before. AI brings the question into the conversation
Anyone who uses a music streaming app knows the feeling. Listen to one track and it suggests something similar. Then another. Before long, the service is choosing much of the soundtrack to your day.
Traditional recommendation systems have long displayed what researchers call popularity bias, documented in academic work on recommender systems.
Conversational AI can compress that process into a single request.
A user asks a question, compares a few suggestions and requests help making a decision. In some products, the next step can already involve checking availability, selecting a provider or carrying out part of a transaction.
The shop window, once spread across dozens of pages, fits into one paragraph.
A famous brand may appear because there are more public records, references and reviews about it. Another provider may be recommended because it genuinely fits the customer’s request. Businesses should ask a straightforward question: are these suggestions reflecting the quality of the available options, or merely repeating the most visible corner of the market?
Better results for individuals, more similar results overall
Another study adds a useful perspective.
In a 2024 Science Advances paper, researchers Anil Doshi and Oliver Hauser examined short-story writing with the help of generative AI.
Participants who received AI-generated ideas produced stories that scored higher in creativity and quality, particularly those who initially performed less well. Across the entire group, however, the AI-assisted stories became more similar to each other.
Imagine ten people developing an advertising campaign. Each starts with a different idea. If they all consult similar tools, the final work might become more polished while losing some of the differences that once made those ideas distinctive.
The experiment examined writing rather than shopping. It still sharpens the question raised by Matz: how much human variety could disappear as more of us turn to the same systems to create, decide and choose?
How to find out whether AI recommends your business
A business owner can begin this investigation without a massive project. The first requirement is to ask the right questions.
What does a customer ask when looking for a supplier? Which attributes matter when comparing products? What criteria matter when choosing a professional?
From there, create a documented test using real customer questions across different AI systems. Record the date, model version, language and context.
| Metric | What to observe | Why it matters |
|---|---|---|
| Brand visibility | How often does the company appear? | Shows whether it enters the first selection. |
| Concentration | Which three brands are mentioned most often? | Reveals which names dominate the answers. |
| Diversity | How many different providers appear overall? | Indicates how much of the market is presented. |
| Reason for recommendation | What evidence supports the selection? | Helps identify decisive information and potential errors. |
| Cross-model comparison | Do different AI systems suggest the same companies? | Reveals repeated patterns and different selection criteria. |
One simple way to measure concentration is to divide mentions of the three most frequently suggested brands by total brand mentions.
Illustrative example: out of 100 mentions, the top three businesses receive 68. The top-three concentration ratio is 68%.
That is a sample calculation, not an observed NEODY result. A real investigation would also account for market size, supplier eligibility and whether the recommendations fit the stated requirements.
It is worth testing whether the results change as questions become more specific. A provider overlooked in a generic search might appear when a buyer adds location, budget, technical requirements or specialization.
That is how a list of names can become meaningful commercial intelligence.
What about Brazil?
AI is already part of business routines in Brazil. The Cetic.br ICT Enterprises 2025 survey found that 17% of surveyed companies used AI technologies, compared with 13% in 2024.
Those figures describe business adoption. To understand purchasing behavior, a different kind of study is needed: one connecting AI-generated recommendations, the alternatives available and the decisions customers actually make.
Matz’s study draws on US users and experimental agents. Its findings offer a strong starting point for research on Brazilian markets, especially local services, small suppliers and emerging brands.
The next step is to examine specific industries, cities and purchasing scenarios.
What should companies do with this information?
If you sell products or services, study the questions that come before a sales inquiry. Check how your business is described in AI responses. Make information about products, qualifications, locations, services and differentiators clear and verifiable.
If you purchase with AI assistance, request alternatives and the criteria behind the shortlist. A convenient answer can still overlook a supplier that may be a better fit.
If you build recommendation systems, consider diversity and reasoning alongside speed and relevance when judging quality.
Competition depends on new businesses having opportunities to be discovered. Technology that connects buyers and suppliers can influence who gets through that first door.
Sandra Matz and the original sources
Sandra Matz studies human behavior, data and technology at Columbia Business School. Her official website features research and academic work, and her professional LinkedIn provides another direct reference.
Her talk, Is AI turning us all into the same person?, available on TED, is an accessible introduction. The research preprint provides the methodology and detailed results.
Who chooses what you get to discover?
The ice cream shop stays with us for a reason. All 31 flavors are still there. One recommendation, however, can change which of them a person ever considers.
Now imagine the same dynamic across thousands of businesses.
Before competing on price, quality and service, a brand has to enter the customer’s field of vision. Part of that field is increasingly being organized by systems that confidently answer which options are worth considering.
As AI enters the purchasing journey, making the shortlist could become as important as making a compelling offer.
The next questions are who appears, why they appear and how many competitors are invisible before negotiation even begins.
