Chapter 5

One robot’s named companies barely changed in 20 days

The toy shop again. A robot behind the counter has a speech bubble with three toy cars in it. A wall calendar shows three weeks of days, with the first and the last in green. A sealed envelope lies on the counter. A child looks on, and a man with a clipboard takes notes.

Imagine a robot in a toy shop names three toy cars for you, and you write them down. You come back three weeks later and ask the same thing. Is your note still good?

Your customers can ask a robot the same question this month and next month. The names they hear can stay or can change.

So I tested one robot on twelve known software companies, and I went back after about three weeks.

The short version. Over about three weeks, this robot kept naming nearly the same known companies on the same questions, and kept leaving out nearly the same ones. In this test, the older count still matched closely. Count question by question, and start a new sheet when the robot’s version changes. A version is the maker’s name for an update.

What I did

The robot was Gemini. Each question asked about software for online shops, projects or sales, and named no company. I checked 4 companies on each question. That made 216 company-and-question pairs. The robot’s version label stayed the same.

Almost every yes or no held for 20 days

Files on shelves, 202 in all. On the top shelves, 69 files are green and 4 are white. On the lower shelves, 125 files are grey and 4 are white.
194 of 202 pairs were the same 20 days later

Picture it. Each file is one pair. The top shelves hold pairs named in all 3 answers of the third round, and the lower shelves hold pairs named in none. Green and grey files stayed as they were. The 8 white files changed by an answer or two, and no file went from always to never or back.

At Giulia’s: like checking the address in 202 client files again after 20 days and finding 194 the same.

Most pairs were all or nothing

216 small parcels in rows. 63 are green, 34 are white and 119 are grey.
Across 12 answers, 182 of 216 pairs were named every time or never

Picture it. Each parcel is one pair. Green ones were named in all 12 answers, and grey ones in none. Only the white ones came and went.

At Marta’s: like 216 products counted 12 times, nine times in one day and three times 20 days later, with 182 in stock at every count or at none.

Count by question, because a company’s total tells you less

A company can be named on one question and missing on the next. So I guessed the last round two ways, each from how often the robot had named the company earlier. One used a count for each question, the other one count for the whole company.

Nine order slips in three rows, all green.
Counting each question separately gave better guesses in all 9 groups

Picture it. Each slip is one group of questions, about one kind of software and one buyer need. Green means the count for each question came closer.

At Javier’s: like 9 nights of 9 when what each regular ordered on their last three visits guesses tonight’s order better than the best-seller list.

The latest three answers guessed nearly as well as nine

Nine books on a shelf. 5 are green, 1 is pale green and 3 are white.
Nine answers beat the latest three in 5 of 9 groups, and tied or lost in 4

Picture it. Each book is one group of questions. Green means nine answers from three rounds guessed the last round better. White means three answers from the third round alone did. The pale one is a tie.

At Joost’s: like guessing orders for 9 book series. Using all nine shop reports beats using the latest three on 5 series, ties on 1 and loses on 3.

How to check this for your business

Portrait of Samira, the media agency founder.Samira BenaliMedia agency founder, FranceMade-up example

A hotel owner asks Samira how often to check the robots. No hotel was in my study, so Samira measures.

Samira saves today’s answers with the robot’s version on top, and asks again a few weeks later. A new version means a new sheet. The same version means reading both sheets line by line, which shows the hotel how much a few weeks changed.

Try this on Monday, then again in three weeks

Choose ten questions, starting with your sheet from chapter 1, and one robot. Write the robot’s version at the top, if it shows one. Ask each question three times, in a new chat each time, and write down the names.

In three weeks, do the same and compare line by line. A matching line means only that you saw no change: three tries a side can miss a real one, as chapter 4 shows. If the version changed, start a new sheet. If the robot shows no version, you cannot tell whether it changed.

References

Chapter 5. One robot’s named companies barely changed in 20 days

Preprint. Dmitrij Żatuchin (2026). Twenty Days Later: A Pre-Registered Test of History-Based Forecasts of Brand Mentions in Gemini Answers. arXiv preprint, version 1, 14 October 2026. arXiv:2610.19608. Data. 10.5281/zenodo.23351313. Register entry. osf.io/x8h3p.

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