Making your company readable by AI. Outside, and in. Published reading · Five surfaces · 16 August 2026
Is a brand cited in the same way across answer surfaces?
The answer is not quite the same. On 16 August 2026, we put the same ten frozen buying questions to five surfaces whose answers were retained. Depending on the surface, Real French is cited for 2 questions out of 10 or 3 questions out of 10. One precise question is cited on all 5 surfaces out of 5. On Google, however, a single answer returns yes or no depending on what the reader calls a citation.
What was compared
The retained reading covers the OpenAI API with web search, the Claude API with web search, the Claude application, the Gemini application and Google AI Overviews. The ten questions come from the Real French dataset published under the CC BY 4.0 licence. None contains the brand name.
The reading rule
On 16 August 2026, both APIs answered each question three times. A question counts as cited there when realfrench.co appears in the sources of at least 2 answers out of 3. The three visual surfaces have one capture per question. We also publish the answer-level fractions because a majority alone would hide variation between passes.
The five measurements, each with its denominator.
| Surface | Dated result and method |
|---|---|
| OpenAI API with web search | On 16 August 2026, realfrench.co appears in the sources of 5 answers out of 30. A majority of at least 2 passes out of 3 is reached for 2 questions out of 10. The calculation rereads 30 files out of the expected 30 and their source lines. |
| Claude API with web search | On 16 August 2026, realfrench.co appears in the sources of 8 answers out of 30. A majority of at least 2 passes out of 3 is reached for 3 questions out of 10. The name appears in the body of 7 answers out of 30, and the naming majority is reached for only 2 questions out of 10. |
| Claude application | On 16 August 2026, Real French is named and realfrench.co is displayed as a source for 2 questions out of 10. The method rereads at original resolution one new private conversation and one captured answer per question. |
| Gemini application | On 16 August 2026, Real French is named and displayed as a source for 2 questions out of 10. The method rereads one captured answer per question and counts only visible text and sources. |
| Google AI Overviews | On 16 August 2026, an AI Overview appears for 8 questions out of 10. Real French is visible as a source for 2 questions out of 10, or 2 of the 8 displayed overviews. It is named in the prose for only 1 question out of 10. The method rereads the full capture of every result. |
What the totals say, and what they do not say
The range is narrow, from 2 questions out of 10 to 3 questions out of 10, but the positive questions are not all the same. On 16 August 2026, “One-on-one French immersion Brittany” cites Real French on all 5 surfaces out of 5 under the rule above. Similar totals are not identical answers.
On all 5 surfaces out of 5, the set of questions where the brand is visible is also the set where realfrench.co is a source. This finding from 16 August 2026 comes from comparing the source lines of 60 API answers out of the 60 retained and 30 visual readings out of the 30 retained. It does not mean that the name always appears in the prose.
This study extends the control between the ChatGPT application and the command-line instrument. The first control asked whether two surfaces returned the same verdict. This one shows why a verdict must also name its surface and its definition.
“Being cited” has four defensible definitions.
The question “French immersion for diplomats” is enough to show this. On 16 August 2026, the same Google results page was retained in its collapsed state, its expanded state and after opening the image. Each fraction below concerns that single capture, taken in France in results Google described as not personalised.
The name in the prose
On 16 August 2026, the generated text names Real French in 0 captures out of 1. It recommends three other organisations.
The source card
On 16 August 2026, Real French is visible in the sources in 1 collapsed view out of 1. After “Show more”, it remains visible in 0 expanded views out of 1 because the three displayed cards change.
The image used
On 16 August 2026, the Real French illustration appears in 1 overview out of 1. Its attribution is established by the realfrench.co card opened from that image, not by an unreadable watermark.
The organic result
On 16 August 2026, Real French holds the first organic position on 1 page out of 1 below the overview. That presence does not answer the same question as a citation in the generated response.
A prose-based measurement says no. A measurement based on sources, the image or the organic result says yes. Even the source-based answer changes when the reader expands the page. A number without a definition and a display state does not describe a reproducible instrument.
The evidence, followed by the limits.
The source files
For the 16 August 2026 reading, the API fractions are recalculated from 60 files out of the expected 60, api-qNN-pN.txt and claude-api-qNN-pN.txt, then compared with api-releve.json and claude-api-releve.json. The 20 Gemini and Google readings out of the expected 20 come from releve-vision.json. The Claude application captures cover all 10 questions out of 10. The expanded Google case is documented in CONSTAT-apercu-google-diplomats.md.
Instrument identity is missing
The campaign artifacts contain no authoritative session record for model and version identity. Names written in scripts, response headers or folders are therefore not published as instrument identities. The ChatGPT application captures and command-line sessions are also absent from this file set, so no result from either surface enters the fractions above.
What this study does not demonstrate
Ten questions about one brand on one day do not measure an entire market. The applications have one answer per question, while the APIs have three. The captures do not expose every source hidden behind a button. Finally, Real French belongs to the Henusse family and works with HENUSSE. We publish that connection so readers can assess the measurement with the conflict known.
Reading performed on 16 August 2026 and published on 17 August 2026. The ten questions were frozen before measurement. Raw answers, captures and reading sheets are retained.