At the IREP Forum on November 3, 2026, Emmanuel Dollé, co-founder of Bubbling, made a simple case: to measure what AI says about a brand, you should not count isolated prompts, but replay full conversations, in large enough numbers to give a result with its margin of error. This article walks through the reasoning, with the figures from our two studies.
Why a single prompt is not enough
GEO (Generative Engine Optimization) refers to the set of practices aimed at getting a brand to appear and be recommended in the answers of assistants such as ChatGPT or Gemini. To know whether these practices work, you need a measurement. Many tools ask a list of questions, note whether the brand appears, and display a percentage.
That percentage hides two problems.
- People do not stop at the first question. A buyer, an investor or a journalist reads the answer, then follows up: price, competitors, reviews. The decision often takes shape at the third or fourth answer. A measurement that stops at the first message misses this part.
- An answer is not reproducible, an interval is. Ask an AI the same question twice: you do not get the same sentences. What can be reproducible is the share of answers in which the brand appears, provided you repeat enough times and state the margin of error.
Our method: full conversations
Together with the brand, we define who is looking for what (a buyer, an investor, a candidate). We create personas, virtual profiles that are validated before launch. Each persona then holds a full conversation with the AI: an opening question, then follow-ups that depend on the answer received. The same conversation is replayed on several assistants. Finally, we count what comes out: brand mentioned, recommended, sources, tone.
Two recent studies apply this method.
| Electric-car study | CAC 40 study | |
|---|---|---|
| Partner | Autoactu | Reputation Age |
| Period | August 2026 | Collection from August 25 to 28, 2026 |
| Assistants | ChatGPT and Gemini | ChatGPT and Gemini |
| Volume | 2,016 conversations, 8,064 questions | 8,671 conversations, 39 companies |
| Unique questions | 96% of follow-ups | 92% of prompts |
How many answers do you need?
Sample size determines precision. For a brand presence of around 67%, the standard formula (95% interval, one proportion) gives:
| Target margin of error | Answers needed |
|---|---|
| ± 10 points | about 85 |
| ± 5 points | about 340 |
In other words, doubling precision takes four times as many answers. A result drawn from five answers could just as well be 40% as 100%. That is why we publish the range, not just the figure.
One more caveat: answers to the same question resemble each other. Ten thousand answers to a hundred questions are not worth ten thousand independent answers. A margin of error calculated as if they were is too optimistic, and the real margin over the whole remains wider.
What the conversations show
The electric-car study produced three findings.
- The score depends on the engine and the profile. The same brand is not treated the same way by ChatGPT and by Gemini, nor for two different households. A single score hides these gaps.
- The answer changes as the conversation goes on. What AI says at the first message is not what it says after three follow-ups.
- Being mentioned is not being recommended. A brand can appear in the answer without being the one the AI advises.
The CAC 40 study points the same way in another field: AI does not always repeat the earnings release as is, and the story told changes depending on the assistant.
What this changes for a brand
- Ask any measurement tool for the number of answers and the margin of error behind each percentage.
- Check that the measurement follows the conversation, not just the first question.
- Read the score by engine and by profile before reading the average.
- Distinguish mentioned from recommended.
Want to know how AI talks about your brand, conversation after conversation? Contact the Bubbling team for a first sample on your market.
Frequently asked questions
How is this different from listening to existing conversations (LLM-listening)?
Listening analyzes conversations that already exist. We generate them, with personas and a protocol set in advance, which gives a known denominator and allows comparison from one wave to the next. The two approaches complement each other.
Keep reading
Sources
- Bubbling × Autoactu study, electric cars: 2,016 conversations, 8,064 questions, ChatGPT and Gemini, August 2026.
- Reputation Age × Bubbling study, CAC 40: 39 companies, 8,671 conversations, ChatGPT and Gemini, collection from August 25 to 28, 2026.
- Talk by Emmanuel Dollé, IREP Forum, November 3, 2026 ("GEO, le nouveau paradigme : quels enjeux, quelles mesures ?").
- Sample sizes: standard proportion formula, 95% interval, p ≈ 0.67 (Bubbling calculation).