Skip to content
Waseit
← All articles
AEOGEO

You made the list. Then the customer asks a second question.

18 conversations, three countries, two turns each. At the moment the assistant has to pick one, a third of the time it names a business that was never on its own list.

August 21, 20267 min read
In this article
  1. The protocol
  2. ChatGPT does not choose. It asks where you are.
  3. Being on the list does not get you to the final
  4. What breaks the list: urgency
  5. Gemini picks from its list - except for lawyers
  6. A number we threw away
  7. What to do on Monday morning

On 14 August, Google put a new model behind AI Mode in Search: Gemini 3.7 Flash, released the day before. Robby Stein, VP of Product for Google Search, announced it in a sentence worth pausing on - the model is better at following instructions and understanding intent.

In a real conversation, intent is not in the first question. It sharpens in the second one. The customer asks for "the best bakeries in Bristol", reads the list, then asks what they actually wanted to know: "which one does proper sourdough?". That is the turn where they choose.

Every AI-visibility measurement published this year stops at the first turn. AirOps' "State of AI Search" report, whose figures are quoted everywhere - 30% of brands stay visible from one answer to the next, 20% across five runs - asks the same question again and again. Nobody measures the conversation. We did, this morning.

The protocol

Six cases, three countries: plumber in Lyon, bakery in Bordeaux, employment lawyer in Paris, plumber in Manchester, bakery in Bristol, encanador in São Paulo. For each one, two turns in the same conversation: the list question, the answer fed back into the history, then the question of a customer who has to decide, with the instruction "give me just one name". Three assistants - ChatGPT, Gemini, Perplexity - so 18 conversations. Every name was recounted by hand.

ChatGPT does not choose. It asks where you are.

Zero times out of six did ChatGPT name a business on the second turn. Five times out of six, it answered with a question of its own:

  • "Tell me your neighbourhood or postcode and I'll give you the best available contact." - the leak in Lyon.
  • "Which part of Bordeaux are you in?" - the sourdough.
  • "Are you the employee or the employer?" - the dismissal in Paris.
  • "Do you mean Manchester, UK or Manchester, NH?", then "Do you mean Bristol, UK?" - it does not even know which city we are talking about.

The sixth time, in São Paulo, it did give a name: Triider - a marketplace that connects you to tradespeople, not a tradesperson.

The most used assistant in the world, at the moment the customer has to choose, names nobody. It asks for a postcode.

Being on the list does not get you to the final

Perplexity does decide: six times out of six it gave one name, and only one. But three times out of six, that name was not on its own first-turn list.

Lyon.

First turn: Bilik, Les Bons Artisans, RingTwice, Yoojo, Travaux.com. Second turn: Accord Artisans, never mentioned before.

Manchester.

First turn: Bee Heated MCR, AJ & RC Plumbing, SR Plumbing, JB Plumbing, About Town Plumbing. Second turn: Kang Plumbing, never mentioned before.

São Paulo.

First turn: Triider, GetNinjas, oHub, Encanador 24 Horas, Play Tec. Second turn: Assisty 24h, never mentioned before.

What breaks the list: urgency

The split is clean, and it is not luck. Our six second questions fall into two families. Three are about availability: someone has to come today. Three are about quality: proper sourdough, real levain, contesting a dismissal.

All three availability questions produced a new name. All three quality questions kept a name from the list - Maison Lamour in Bordeaux, Hart's Bakery in Bristol, Kohen Avocats in Paris.

And the engine says why itself. On Accord Artisans: "they advertise a 30-minute call-out in Lyon, 24/7". On Kang Plumbing: "listed as open 24/7 in Manchester and says it offers immediate assistance". Reputation does not win that turn. A sentence written somewhere does - one that says you can be reached now.

Gemini picks from its list - except for lawyers

Gemini named a business three times out of six, and every time it was a name from its own list: Blicko in Lyon, Le Blé en Herbe in Bordeaux, Farro Bakery in Bristol. In the other three, it sent us elsewhere - in Manchester to United Utilities, the water company, and to Checkatrade, a directory; in São Paulo to GetNinjas, a marketplace.

In Paris it refused, and said so plainly: it is bound by rules of neutrality that forbid it from singling out one professional over another. In a regulated profession, the second turn can have no winner at all.

A number we threw away

Our automatic extraction first reported survival rates of 0%, 19% and 50% depending on the engine. We are not publishing them: they are wrong. Reading the lists back, some of the "businesses" counted are not businesses at all - they are pieces of advice the extractor mistakes for proper nouns: "Word of mouth still rules", "Check the registration number", "Red flags", "Dica de emergência".

That is a hole in our tool, and it is not confined to the writing systems we opened last week: it opens in French and English too, as soon as the assistant puts a piece of advice in bold. We write it down because naming a defect is worth more than publishing a flattering percentage. The nine names in this article were each verified by hand.

What to do on Monday morning

Write your availability out in full

, on your site and in your listings: hours, days, area covered, call-out time. That is the sentence the engine quoted to justify all three of its surprise picks.

Write your address and your neighbourhood.

Five of the six questions ChatGPT asked on the second turn were about location. An assistant that does not know where you are cannot choose you.

Do not measure only the first answer.

Being in it does not put you in the final: a third of the choices we observed came from somewhere else.

If your profession is regulated

- lawyer, doctor, accountant - expect a refusal to name anyone. What matters then are the official directories the assistant points to instead.

The limits, as always, and they matter here: six cases, three countries, one second question each, one model per engine, one morning. The availability/quality split rests on three cases each way - a coherent mechanism, quoted by the engine itself, not a law. What is solid fits in one sentence: the first-turn list is not a shortlist.