Chapter 7

About one answer in four named a person

A music shop with a wall of guitars. Two robots stand behind a counter. The left robot's speech bubble holds a small shop. The right robot's speech bubble holds one person with no face. A child looks up at them.

Imagine you walk into a music shop and ask a robot who can teach you piano. One robot names a music school. The robot beside it names one teacher.

Your customers can now ask AI who should check a contract or fill in their tax forms, and the answer can name a firm, a person or both.

So I tested how often that happens, with questions about people who sell homes, cars and insurance.

The short version. In three trades, about one answer in four named a person. The share changed a lot by trade and by robot. Read the answers yourself, because a list of names made in advance can miss them.

What I did

I asked about five cities: Warsaw, Dublin, Amsterdam, Vilnius and Tallinn. Every question asked for a person by job, such as the best car salesperson in a named city. Each robot was told to search the web. I also made a list of people to look for. A program then read every answer for people’s names, without that list.

The trade made a big difference

Four blocks of 100 small pictures. Name tags, 26 green. Houses, 35 green. Cars, 33 green. Sheets of paper, 9 green.
26All three trades
35Selling homes
33Selling cars
9Insurance
Robots named a person in 26 of every 100 answers. For insurance, the count was 9.

Picture it. Each small picture is 1 of every 100 answers about that trade. Green ones named a person. White ones named nobody my program could find.

At Mikko’s: like 100 customer notes on buying a car, 33 with a salesperson’s name, and 100 on car insurance, 9 with a person’s name.

The program missed some known names, so the count may be low.

Places could go wrong too: a question that said only Warsaw, meaning the city in Poland, could bring back car dealers in Warsaw, Indiana. I took names like those out of the counts.

The answer depended on which robot you asked

Two boxes of 100 small speech bubbles. In the left box 38 are green. In the right box 9 are green.
GrokA bar filled to 38 out of 100.38
PerplexityA bar filled to 34 out of 100.34
GPT, behind ChatGPTA bar filled to 22 out of 100.22
GeminiA bar filled to 9 out of 100.9
On the same questions, one robot named a person about 4 times as often as another

Picture it. Each speech bubble is 1 of every 100 answers. Green ones named a person. The left box is Grok, the right is Gemini, and the bars show all four.

At Samira’s: like four staff, each writing 100 ads from the same brief. One puts a person’s name in 38, and another in 9.

A list of names can miss what robots say

In 600 answers about Warsaw, my list found no one. An automated reading of the answers found a person in 188.

Picture it. Each square is one answer about Warsaw. Top: names from my list. Bottom: any person’s name. Green means found.

Counting names from my list600 small squares, all white.0
Reading the answers, without my list600 small squares, 188 of them green.188

At Daniel’s: like checking 600 case files against a client list and finding no match, while 188 carry a person’s name that is not on it.

Answers that named a person listed slightly different pages

Robots also list pages as sources. A robot helped sort those pages into kinds using my rules.

Two piles of 100 small pages. Left pile: 4 green with a dot, 11 plain green, 44 striped, the rest white. Right pile: 1 green with a dot, 7 plain green, 45 striped, the rest white.
Listing sites were 11 of every 100 pages when a person was named, and 7 when my program found nobody

Picture it. Each page is 1 of every 100 listed. White pages are other kinds. Left pile: answers that named a person. Right pile: the others. Plain green is a listing site, which shows many sellers in one trade. Green with a dot is the site of someone who works alone: 4 against 1. Striped pages are a firm’s own site: 44 against 45.

At Joost’s: like two shelves of 100 books, with 11 trade catalogues on one shelf and 7 on the other.

The page mix changed with the trade. These pages appeared beside names; they do not show what made a name appear. Firm sites filled almost half of both piles.

How to check this for your business

Portrait of Daniel, the lawyer.Daniel MensahLawyer, GermanyMade-up example

Daniel’s firm is called Mensah. Daniel asks the robots who can check a work contract in Berlin, and reads each answer to the end.

One answer names a firm. Another names a lawyer at a rival firm. A third names a lawyer in Berlin, New Hampshire. And when the name Mensah comes up, does the robot mean Daniel or the firm? The sentence around it may tell.

Try this on Monday, with your sheet from chapter 1

Take one question from your sheet and ask two robots. Put your list of names aside and read each answer to the end. Write F for a firm, P for a person and X for a wrong place. One answer can get more than one. Copy the first link it lists, if it lists one.

The next answer can differ, so this shows what each robot said this one time.

References

Chapter 7. About one answer in four named a person

Preprint. Dmitrij Żatuchin (2026). Who Gets Named: Citation Type Predicts Individual Naming by Grounded Language Models, and a Roster Instrument Captures 0.5% of It. arXiv preprint, version 2, 1 August 2026. arXiv:2607.23893.

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

Fields marked * are required.

Who is it for?

Your choice about cookies

We use cookies this site needs to work, and others only if you say yes. Choose what you allow, and change your answer whenever you like. Read our privacy policy