Blog

What AI does with a company's earnings release: 8,671 conversations about 39 CAC 40 companies

Bubbling · November 17, 2026

We asked ChatGPT and Gemini to talk about the results of 39 CAC 40 companies: 8,671 conversations, run from August 25 to 28, 2026, with Reputation Age. On average, AI repeats 68% of the messages and indicators the company highlights, but the gap between companies is very wide, and so is the gap between the two assistants. What AI says about you depends first on what you publish and how you structure it: 78% of AI citations point to the companies' own websites.

Why this study

An earnings release is written for analysts and journalists. Recently, it is also read by AI assistants, which shareholders, journalists and job candidates ask before reading anything else. The question facing a financial communications team is changing: it is no longer only about what the press says, but about what ChatGPT and Gemini tell people when asked to present the company.

This study, conducted by Bubbling with Reputation Age, was picked up by Boursorama (op-ed of September 28, 2026) and then by INfluencia (October 6, 2026). It drew attention mainly for one figure: about 40% of the figures AI gives on results are wrong. That is not the only finding, and for a brand it may not be the most useful one. This article covers the others.

GEO (Generative Engine Optimization) refers to the set of practices aimed at making a brand more visible and better understood in AI assistants' answers, as SEO did for search engines.

Method

  • Scope: 39 CAC 40 companies.
  • Assistants queried: ChatGPT and Gemini.
  • Corpus: 8,671 conversations, run from August 25 to 28, 2026.
  • What we measured: for each company, three equity story messages (what the company wants people to remember about it) and three indicators highlighted in its earnings communication, 114 messages in total. We recorded which ones come back in AI answers, with which figures, which sources and which nuances.
  • How: simulated conversations with personas, as the Bubbling platform does, rather than a single isolated question.

1. On average, 68% of the story gets through, with very large gaps

Out of each company's three equity story messages and three indicators, ChatGPT and Gemini repeat 68% on average. The average hides a large gap: Michelin sees 97% of its story repeated, another CAC 40 company only 33%. The questions and the assistants are the same. Only the release changes.

2. The two assistants do not tell the same story

Of the 114 equity story messages measured, 28 (one in four) show a pickup gap of more than 40 points between the two assistants: well repeated by one, barely or not at all by the other. The gap goes both ways: depending on the message, ChatGPT or Gemini repeats it better. An investor who asks one gets the key message, one who asks the other does not, even though the source is the same.

An average measure of "AI" therefore hides what your audiences hear. You need to measure engine by engine and message by message.

3. AI points out discrepancies and picks its own topics

In 43% of answers, AI points out at least one discrepancy or nuance compared with the figures the company highlights. The topics it raises on its own:

Topic raised by AIOccurrences
Quality of earnings (adjustments, base effects, scope)3,007 (top topic for 36 of 39 companies)
Currency effects1,633
Debt1,337

Conversely, litigation is marginal, ESG is almost absent and lack of transparency is hardly ever mentioned. AI does not follow the company's agenda: it reads the figures as a business journalist would and goes straight to the adjustments.

4. The company holds the pen: 78% of citations come from its own websites

78% of the citations AI displays point to the companies' own websites. What AI says about you therefore depends first on what you have published, and how you have structured it: a figure buried in a PDF, an indicator redefined from one quarter to the next or a poorly identified period gets repeated as is, or repeated wrongly.

5. The figures: an average accuracy of 60.3%

Among answers that give a figure, 60.3% on average are correct, within the chosen tolerance, across the 38 companies measured for this indicator. The most frequent errors come from one period being confused with another, for example the previous half-year's figure given as the current one.

What this changes for a brand

  1. Choose the three messages AI should remember: write them as standalone sentences and repeat them where AI reads: release, results page, corporate website.
  2. Structure your figures. One period, one scope and one definition per indicator, on the page that serves as the source.
  3. Measure by assistant. An average does not tell you which one your audiences ask. A message can get through on one engine and disappear on the other.
  4. Anticipate sensitive topics. If AI raises quality of earnings, currency or debt on its own, it is better to address these topics in your own communication before it does so for you.
  5. Start again with every release. A release is a new corpus.

To see what ChatGPT, Gemini and other assistants say about your brand, which messages get through and which disappear, you can request a free preliminary diagnostic from Bubbling.

Frequently asked questions

What is the 68% pickup rate?

It is the share of a company's highlighted equity story messages and indicators that ChatGPT and Gemini repeat in their answers, averaged across the 39 companies.

Does AI often get the figures wrong?

About 40% of the figures given are not within the chosen tolerance. Errors mainly come from confused periods. The study does not measure why each error occurs.

Why don't ChatGPT and Gemini tell the same story?

They do not rely on the same sources, nor in the same way. Of 114 messages, 28 are repeated with a gap of more than 40 points between the two.

Does this apply to a company outside the CAC 40?

Not necessarily. The study covers large, heavily covered companies. The principle, however, applies to any brand: the measurement must be redone on your own scope.

Can you influence what AI says about you?

Partly. 78% of citations point to the companies' websites: what you publish and how you structure it matter first.

Sources

  • Reputation Age × Bubbling study, "The CAC 40 results under the scrutiny of ChatGPT and Gemini" (September 2026): 39 CAC 40 companies, ChatGPT and Gemini, 8,671 conversations, August 25-28, 2026.

Keep going with the pages our readers use most to get visible inside AI answers.