Chapter 3

One answer misses names, so ask again

The toy shop again. A robot behind the counter says a few names, drawn as stickers in a speech bubble. A child fills an album of stickers on the counter, and one place in it is still empty. The shelves behind them hold sticker packs.

Imagine you ask the shop robot which toy car to buy, and it names a few. You stick each name in an album. You walk out, come back and ask again. Now a new name comes up.

Your customers can do the same when they ask a robot which firm to choose. One answer can show them part of what the robot might say.

So I tested how a list grows with each new try, on buying questions about software, consulting, payments and banking, online shops and health technology.

The short version. One answer shows part of the list. Ten tries found most of what 15 tries found, yet most questions still held a name seen in one try alone. If someone sells you a percentage, ask how many tries they ran and how they counted names.

What I did

Picture each answer as a pack of stickers, one for each company name. I compared what one, five, ten and fifteen tries found.

One try shows part of the list

One try found about 6 to 8 of every 10 names from five tries

Picture it. Each row of 10 stickers stands for all the names from five tries. Green stickers were already there after one try.

The robot with the smallest share10 stickers, 6 of them green.6 of 10
The robot with the largest share10 stickers, 8 of them green.8 of 10

At Mikko’s: like one customer naming 6 to 8 of every 10 dealers that five customers name between them.

Ten tries found most of what 15 tries found

15 sticker packs in two rows. The top row has 10 green packs, and the bottom row has 5 white packs.
Sonar, from Perplexity, the highestA bar filled to 95 out of 100.95
Mistral, the lowest by the new ruleA bar filled to 90 out of 100.90
Ten tries found 90 to 95 of every 100 names that 15 tries found

Picture it. My 15 packs for one question fill one album, and the green packs are ten of them, picked by chance. Each bar shows how much ten packs filled on average, on a robot’s middle question.

At Joost’s: like 10 of 15 book clubs, picked by chance, naming on average 90 to 95 of every 100 titles that the 15 clubs name between them.

The album holds the names found in my 15 tries. A different question can need more or fewer tries.

Most questions held a name from one try alone

On 33 to 48 of 50 questions, a name came up in one try alone

Picture it. Each card is one question, asked 15 times. On a green card, at least one name came up in one try alone. It can be a real firm or a counting mistake.

Sonar, the fewest green cards of my six robots50 cards, 33 of them green.33
Mistral, the most green cards of my six robots50 cards, 48 of them green.48

At Daniel’s: like asking 15 clients to name a lawyer, for each of 50 kinds of case. On 33 to 48 kinds, at least one lawyer is named by one client alone.

Change the counting rule, and the numbers change

My old rule could skip a rare one-word name. A new AI-assisted count read more possible names in context. The two rules often disagreed.

100 answer sheets in four rows. 50 are green and 50 are white.
A new counting rule changed the names on half the answers

Picture it. These 100 sheets stand for all my answers. On the green ones, the names stayed the same. On the white ones, they changed.

At Javier’s: like a new rule for sorting the same 100 order slips, and the sorting changes on 50 of them.

How to check this for your business

Portrait of Joost, the publisher.Joost BakkerPublisher, the NetherlandsMade-up example

Joost asks for Dutch publishers. Each time, another name joins the list. When a report arrives, he asks how many tries it used and what counted as a name.

Try this on Monday: watch one list grow

Ask one question from your chapter 1 sheet 15 times, using one robot and a new chat each time. Write down each new name. Count the different names after try 1, 5, 10 and 15. Note what you counted as a name.

References

Chapter 3. One answer misses names, so ask again

Preprint. Dmitrij Żatuchin (2026). Measuring Brand and Source Discovery under Repeated LLM Queries: A Finite-Sample Audit. arXiv preprint, version 2, September 2026. arXiv:2609.05059v2.

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