Chapter 2
Ask robots for lottery numbers, and tickets repeat
Imagine the toy shop runs a lucky draw. A bag holds balls numbered 1 to 49. You pull out six, and those six numbers are your ticket. Now you ask the shop robot for six numbers, again and again. Its answer changes, so it looks as mixed as the bag.
Your customers can do the same with AI. They can ask one question many times, and answers that change can look like a wide choice.
So I tested how wide the choice is, with lottery numbers. It is a plain number question, with no company, product or purchase in it.
The short version. A robot’s answers can change from one try to the next and still keep landing on a few choices. So if a report says the answers varied, ask two things. How many different answers were there? How often did the top one come up?
What I did
A fair bag is a useful yardstick. The published Polish Lotto counts looked much like one at this scale.
A fair bag almost never repeats a ticket
I calculated this yardstick; I did not draw from a real bag.
Picture it. Each small box here is one round of 200 tickets from the bag. In the green round, some ticket came up twice. In the 702 white rounds, every ticket was new.
At Giulia’s: like 703 months of 200 invoices each, with two matching invoices in about one of those months.
Every robot kept returning to a top ticket
Picture it. Each small ticket is one answer from that robot. Green tickets all hold the same six numbers. White ones hold others.
At Marta’s: like 136 of 200 parcels going out with the same six items inside.
None of these tickets is a lottery tip. They show what the robots repeated.
Answers can change and still stay narrow
Picture it. Each small ticket is one answer, and each pile is one set of six numbers. The 2 tall green piles hold 187 answers. The 6 short white piles hold the other 13.
At Samira’s: like filling 200 ad slots from 8 different ads, with 2 of those ads taking 187 slots.
These answers did change, and a few of them side by side would look mixed. A bag can be uneven and still work by chance. Fill one with many balls that carry the same few numbers. Each pull is still chance, and those numbers keep coming up. So change on its own is no proof of a wide choice.
The words you use can change the count
Picture it. Each small ticket is one answer from that robot. Green tickets hold the top set for that wording. The two green sets differ by one number.
At Joost’s: like a helper on a book-fair stall. Ask for “a novel” 50 times, and the top title comes up 4 times. Ask for “a prize novel” 50 times, and the top title comes up 36 times.
Do company names behave like these numbers? I have not run this test on company questions. The next chapter counts names in another way.
How to check this for your business
Marta gets a report on robot answers about gift shops. It says the answers changed each time, so many shops have a chance.
Marta asks two questions. How many different answers came back? How often did the top one come up? The counts will not tell her what buyers do. They will show her if “changed each time” meant many answers or a few.
Try this on Monday, with one report
Take any report someone sent or sold you about robot answers, and pick one question. Write two counts: how many different answers or names came back, and how often the top one came up. If the report gives neither, ask for both.
No report? Use your sheet from chapter 1. Count how many different names came up and how often the most common name appeared. Keep each question and robot separate.
References
Chapter 2. Ask robots for lottery numbers, and tickets repeat
Preprint. Dmitrij Żatuchin (2026). Ask a Language Model for Lottery Numbers: Concentration in Repeated Six-of-49 Outputs. arXiv:2610.00052, version 1. arXiv:2610.00052v1. Data and code. github.com/Rankfor/rankfor-open/tree/main/research/lotto-models.
What does AI say about your business?
Put the book into practice. Tell us your company and website to request an AI visibility audit. We’ll reply by email to agree what to check.
What the audit shows
- Where your company appears in the answers, and next to whom
- The pages AI cites in your market
- What to publish first on your own site
