Making your company readable by AI. Outside, and in. 01
Give a direct answer under decisive headings
We recast headings around exact questions and open each passage with a short answer. This structure helps reading without promising a citation.
Published example · AI citation diagnostic
The €1,500 diagnostic, with its evidence.
We are publishing this diagnostic because Real French is the school of Théo’s father, a client of the house, co-directed by Théo. A client’s diagnostic remains confidential. You can therefore read the real substance of the €1,500 excl. VAT deliverable here, with its findings, limits and evidence.
Why this one is public
Real French is owned by Bernard Hénusse, Théo’s father; Théo is its co-director and the school is a client of the house. We can therefore show the pages we observed, the official-tool captures and the findings about the site.
What remains protected
The four schools in the comparison are anonymous. Their figures are unchanged, but we publish no name, address or image from their sites.
One-page summary
The site is sound. Its presence among linked sources remains rare.
Real French receives an overall score of 66 out of 100, computed over the 9 measured dimensions out of 15. Measured AI visibility is 57 out of 100, while conventional web health is 81 out of 100. This contrast is the main finding: a technically sound site does not automatically become a source that answer engines link often.
6 answers out of 270
The realfrench.co domain is linked in 6 of the 270 usable answers, or 2%: 2 answers out of 90 on Claude CLI, 3 out of 90 on Codex CLI with web search, 1 out of 90 on Gemini via Antigravity. The 10 purchase questions carry 3 of these 6 answers; the 20 other questions carry the remaining 3.
4 readings out of 90
The domain appears in 4 question-by-engine readings out of 90: one reading cites it on all 3 passes, three readings cite it on 1 pass out of 3. Three passes do not measure a frequency: these counts describe this reading, these surfaces and these dates, nothing more.
26 pages read
The editorial crawl covers 26 pages. Their form is generally clear, but definite coverage of buyer questions still needs strengthening.
Our reading. Real French has a sound technical and editorial base. The first task is to connect pages more directly to buyers’ exact questions, then ask the same questions again. This measurement predicts neither a citation, a visit nor a sale.
Measurement summary
30 questions, 3 surfaces, 3 runs.
We fixed 30 buyer questions before reading any answers. Each question was asked 3 times through Claude CLI, Codex CLI with web search and Gemini via Antigravity. The measurement therefore contains 270 usable answers, grouped into 90 question-by-engine combinations.
These surfaces are the official tools of the same vendors: Claude CLI from Anthropic, Codex CLI from OpenAI, Gemini through Google's official tool. The web search used is each vendor's own native tool; vendors do not publicly document whether that infrastructure is identical to the consumer app's, and the model served by default to an average user cannot be observed from the outside. The exact surface is named so the reading can be replayed. The gap with the consumer application is measured too: at the control reading of the night of 14 to 15 August 2026, the ChatGPT application and our instrument returned the same verdict on all 30 questions, that of Real French’s absence. This concordance shows the instrument invents no visibility; it does not yet show it detects a genuinely cited brand equally well. The published concordance study carries its method, evidence and limits.
A site counts as linked only when its domain appears in the answer’s links or sources. A name appearing in the prose remains a separate measure: on this reading the instrument records it without yet counting it reliably, so no mention figure is published here. The control reading of 14 to 15 August 2026 does distinguish it, with three verdicts. A failed measurement remains unknown and never becomes a zero.
| Observed site | Answers containing a link | Question-by-engine combinations containing a link | Role in the comparison |
|---|---|---|---|
| Real French | 6 of 270 | 4 of 90 | This is the audited site. |
| School A | 34 of 270 | 14 of 90 | It is selected for the main comparison. |
| School B | 28 of 270 | 12 of 90 | It is selected for the main comparison. |
| School C | 0 of 270 | 0 of 90 | It remains an additional comparison point. |
| School D | 10 of 270 | 6 of 90 | It remains an additional comparison point. |
Question by question
The 30 questions and 90 readings.
Each cell shows how many of the 3 runs linked realfrench.co. The questions remain in English because this is the language in which they were asked.
| Buyer question | Claude CLI | Codex CLI with web search | Gemini via Antigravity |
|---|---|---|---|
| 1. best French immersion programs in France for adults | 0 of 3 | 0 of 3 | 0 of 3 |
| 2. one-on-one French immersion homestay in France | 0 of 3 | 3 of 3 | 0 of 3 |
| 3. live-in French immersion with a host family in France | 0 of 3 | 0 of 3 | 0 of 3 |
| 4. intensive French immersion week in France for adults | 0 of 3 | 0 of 3 | 0 of 3 |
| 5. French immersion course in France for professionals | 0 of 3 | 0 of 3 | 0 of 3 |
| 6. residential French course living with the teacher in France | 0 of 3 | 0 of 3 | 0 of 3 |
| 7. how to become fluent in French through immersion in France | 0 of 3 | 0 of 3 | 0 of 3 |
| 8. best way to learn French fast in France for English speakers | 0 of 3 | 0 of 3 | 0 of 3 |
| 9. French immersion in France with accommodation included | 0 of 3 | 0 of 3 | 0 of 3 |
| 10. small-group French immersion retreat in France for adults | 0 of 3 | 0 of 3 | 0 of 3 |
| 11. why can I read French but freeze when I try to speak it | 0 of 3 | 0 of 3 | 1 of 3 |
| 12. why can't I understand French people when they speak quickly | 0 of 3 | 0 of 3 | 0 of 3 |
| 13. I understand written French but can't follow a spoken conversation | 0 of 3 | 0 of 3 | 0 of 3 |
| 14. why don't French people understand my French when I speak | 0 of 3 | 0 of 3 | 0 of 3 |
| 15. why can't I hear the difference between similar French words | 1 of 3 | 0 of 3 | 0 of 3 |
| 16. why does spoken French sound like one long blur to me | 1 of 3 | 0 of 3 | 0 of 3 |
| 17. I studied French for years and still can't speak it | 0 of 3 | 0 of 3 | 0 of 3 |
| 18. why do I go blank and panic when I try to speak French | 0 of 3 | 0 of 3 | 0 of 3 |
| 19. best way for an adult to finally speak French confidently | 0 of 3 | 0 of 3 | 0 of 3 |
| 20. how to become fluent in spoken French as an adult | 0 of 3 | 0 of 3 | 0 of 3 |
| 21. fastest way to improve your spoken French | 0 of 3 | 0 of 3 | 0 of 3 |
| 22. how to understand fast, natural spoken French | 0 of 3 | 0 of 3 | 0 of 3 |
| 23. how to get conversational in French quickly | 0 of 3 | 0 of 3 | 0 of 3 |
| 24. how to improve my French speaking before moving to France | 0 of 3 | 0 of 3 | 0 of 3 |
| 25. how to prepare for a French oral exam | 0 of 3 | 0 of 3 | 0 of 3 |
| 26. how to pass the DELF speaking test | 0 of 3 | 0 of 3 | 0 of 3 |
| 27. best way to prepare for a GCSE French oral | 0 of 3 | 0 of 3 | 0 of 3 |
| 28. are language-learning apps enough to become fluent in spoken French | 0 of 3 | 0 of 3 | 0 of 3 |
| 29. language apps vs immersion for learning to speak French | 0 of 3 | 0 of 3 | 0 of 3 |
| 30. do group French classes actually help you speak | 0 of 3 | 0 of 3 | 0 of 3 |
The citation shelf
The sources the engines actually used.
Across the 270 answers, the engines linked 262 distinct domains; a domain counts at most once per answer. Here are the domains linked in at least 18 answers, with realfrench.co for comparison. The four compared schools remain anonymised, here as in the rest of the report. A technical Google redirect domain, linked 21 times on Gemini, is excluded from this ranking and remains counted in the file.
| Domain | Linked answers, of 270 | Claude CLI, of 90 | Codex CLI, of 90 | Gemini, of 90 |
|---|---|---|---|---|
| frenchtoday.com | 36 | 18 | 2 | 16 |
| School A (anonymised) | 34 | 17 | 6 | 11 |
| ohlalafrenchcourse.com | 32 | 17 | 2 | 13 |
| fluentu.com | 29 | 16 | 6 | 7 |
| School B (anonymised) | 28 | 19 | 8 | 1 |
| nacel.org | 27 | 11 | 11 | 5 |
| hellofrench.com | 25 | 12 | 7 | 6 |
| commeunefrancaise.com | 20 | 11 | 3 | 6 |
| lingoo.com | 20 | 5 | 8 | 7 |
| cia-france.com | 18 | 12 | 4 | 2 |
| ila-france.com | 18 | 14 | 4 | 0 |
| realfrench.co | 6 | 2 | 3 | 1 |
This shelf shows who occupies, today, the space these 30 questions open. It is published as measured: realfrench.co sits around 40th of 262 domains, and that gap is exactly what this diagnostic measures. The full ranking remains in the delivered file.
Technical comparison
The same page roles, the same tools, the same limits.
Each site is observed through its home page and one editorial page. PageSpeed Insights is run 3 times on mobile and 3 times on desktop for each page. The table shows the median and the observed range in brackets. The Rich Results Test counts the valid and invalid items detected by the tool.
Home pages
| Site | Mobile PageSpeed, out of 100 | Desktop PageSpeed, out of 100 | Rich Results, valid and invalid items |
|---|---|---|---|
| Real French | 96 [96 to 96] | 97 [97 to 98] | 7 valid and 0 invalid. |
| School A | 39 [38 to 42] | 52 [50 to 59] | 2 valid and 0 invalid. |
| School B | 92 [92 to 98] | 99 [99 to 99] | 3 valid and 0 invalid. |
| School C | 56 [56 to 61] | 67 [67 to 67] | 0 valid and 0 invalid. |
| School D | 52 [49 to 52] | 67 [65 to 67] | 0 valid and 0 invalid. |
Editorial pages
| Site | Mobile PageSpeed, out of 100 | Desktop PageSpeed, out of 100 | Rich Results, valid and invalid items |
|---|---|---|---|
| Real French | 99 [96 to 99] | 97 [97 to 98] | 7 valid and 0 invalid. |
| School A | 70 [70 to 70] | 70 [69 to 70] | 2 valid and 0 invalid. |
| School B | 97 [86 to 97] | 99 [99 to 99] | 3 valid and 0 invalid. |
| School C | 80 [65 to 80] | 99 [99 to 99] | 0 valid and 0 invalid. |
| School D | 62 [54 to 62] | 97 [95 to 97] | 0 valid and 0 invalid. |
The Real French pages
The observed pages are published as they appeared.
Official-tool captures
The table scores can be checked in the captures.
We publish every selected Real French capture in full. The chosen PageSpeed capture is the first run equal to the median for its profile. Eligibility in the Rich Results Test does not say what Google will display.
Site audit
The content is readable. Coverage of the questions is the work ahead.
The 26 observed pages have a generally clean structure. We find direct answers after 376 headings out of 557, or 68%, and 770 short, self-contained passages out of 1,054 paragraphs.
| Editorial signal | Measurement | Reading |
|---|---|---|
| Direct answer after a heading | 376 of 557, or 68% | The answer is often visible at the start of the passage. |
| Short, self-contained and well-placed passage | 770 of 1,054 | Most passages make sense without the preceding paragraph. |
| Page containing a table or list | 26 of 26 | Every observed page has at least one reading aid. |
| Page with a clean heading hierarchy | 26 of 26 | Every observed page retains a usable hierarchy. |
| Overlong block | 2 of 1,054 | Two passages need a cut or an opening answer. |
| Passage dependent on preceding context | 0 of 1,054 | No passage falls into this category. |
Question coverage
Of the 30 questions, 1 is certainly handled by a dedicated page, 0 are classified as certainly absent and 29 still need checking or strengthening. The report details 8 examples from those 29 questions.
Technical base
The 2 comparison pages have an aligned canonical address and readable structured data. The robots.txt file allows 8 identities out of 8. The access check receives 13 responses without a challenge out of 14 and encounters 1 challenge. The /llms.txt path returns HTTP 200, while /.well-known/llms.txt returns HTTP 404.
The types observed on Real French include Article, BreadcrumbList, EducationalOrganization, FAQPage, ListItem, LocalBusiness, Person, WebPage and WebSite. Their presence helps technical reading, but it does not by itself explain a citation.
Complete score table
The score orders the findings. It does not replace the evidence.
The overall score is a weighted average of the measured dimensions only, with renormalised weights. An unmeasured dimension is shown as 'not measured' and does not enter the calculation: neither a success nor a neutral score.
Why these weights: the scale scores what the website controls, not what the engines will decide. No engine requires any particular markup or file to cite a page; the weight given to markup measures technical readability, never a citation lever. The citation actually observed is a result rather than a property of the site: it is reported separately, at the top of this report, and no composite should be read as a citation forecast.
| Dimension | Family | Score out of 100 | Weight | Status |
|---|---|---|---|---|
| AI crawler accessibility | AI | 92 | 12 | It is measured. |
| Citation by AI engines | AI | 19 | 20 | It is measured. |
| JSON-LD structured data | AI | 68 | 18 | It is measured. |
| Structured data useful to AI | Web health | 100 | 3 | It is measured. |
| Structured data validity | Web health | Not measured | 10 | It is not validated. |
| Content depth and form | AI | 70 | 10 | It is measured. |
| Essential metadata | Web health | 99 | 8 | It is measured. |
| Readable trust signals | Web health | 70 | 7 | They are measured. |
| Legal pages and consent | Web health | 75 | 9 | They are measured. |
| Off-site authority | Web health | Not measured | 6 | It is not measured. |
| Search Console | Web health | Not measured | 5 | It is not measured. |
| Sitemap and robots | Web health | 75 | 10 | They are measured. |
| Sitemap stability | Web health | Not measured | 7 | It is not measured. |
| Field Core Web Vitals | Web health | Not measured | 10 | They are not measured. |
| Editorial consistency | Web health | Not measured | 8 | It is not measured. |
The measurement also finds 85 addresses in the sitemap, a crawl of 26 pages, a median of 1,227 words per page, 115 question-form H2 headings out of 193, 2 titles outside the expected pattern out of 26 and 2 descriptions outside the expected pattern out of 26.
The scoring scale, rule by rule.
Each note out of 100 follows a rule written in the measurement engine. The formulas date from late spring 2026 and predate this reading, with two exceptions declared below the list. Scale published on 19 August 2026.
- 1. Citation by AI engines: the note starts at 15, adds 190 times the share of answers linking the domain, and removes up to 20 points for failed queries. Here: 15 + (6/270 × 190) = 19.
- 2. JSON-LD structured markup: 35 points if a JSON-LD block exists, plus 65 points pro rata across six useful families (FAQ, HowTo, Product, Review, LocalBusiness, AggregateRating). Three of six: 68.
- 3. AI-useful markup: the percentage of pages carrying at least one of those families. 26 of 26 pages: 100.
- 4. AI robot accessibility: twelve robot identities tested against robots.txt and against real access; the note is the percentage of identities that pass. 11 of 12, Bytespider blocked: 92.
- 5. Content depth and form: 70% direct-answer rate, 20% question-form headings rate, and up to 10 length points, maxed at a 1,200-word median. Here: 70.
- 6. Essential metadata: starts at 100 and removes points pro rata for faulty pages, from missing titles to duplicates. 2 titles and 2 descriptions outside the pattern of 26: 99.
- 7. Readable trust signals: seven signals counted, from the about page to linked profiles; this check records presence without judging expertise, so the note is capped at 70. All seven present: 70.
- 8. Legal pages and consent: two shares per mandatory item, one per optional; 10 of 12 shares with unconfirmed consent: capped at 75.
- 9. Sitemap and robots: 90 when the sitemap reads, with caps by actual coverage. 26 addresses examined of 85 declared: capped at 75.
Two exceptions, declared: the exact list of twelve robot identities was committed to code just after the 7 August measurement started; and the composite computed on measured dimensions only is the 10 August 2026 revision, described above. A scale should not move after a reading is seen; when it does, the revision is dated and published, as here.
Action plan
We start with what the site can control.
02
Check technical access
We investigate the challenge encountered on 1 request out of 14, then decide which crawlers to allow according to their role. GPTBot is used for model training, not real-time search.
03
Strengthen buyer-question coverage
We start with the 29 questions that remain uncertain, choose the closest pages and decide whether to clarify, extend or create a page.
04
Strengthen evidence of trust
We link the official profiles, references and external evidence that corroborate the site’s claims.
05
Correct the counted exceptions
We correct the 2 titles and 2 descriptions outside the expected pattern, then check the canonical addresses and the pages concerned again.
An authority page keeps six simple parts.
- 1. It gives the direct answer and its limits.
- 2. It compares the options and helps the reader decide.
- 3. It connects every important claim to evidence.
- 4. It names the expert responsible for the content.
- 5. It keeps dated, reusable sources.
- 6. It remains maintainable after publication.
Methodology
The measurement can be repeated, not predicted.
The question set was fixed on 7 August 2026. The measurement ran from 7 to 8 August 2026 in UTC. The pages and official tools were captured on 8 August 2026.
| Measured surface | Declared model | Runs per question |
|---|---|---|
| Claude CLI | claude-sonnet-5 | 3 |
| Codex CLI with web search | gpt-5.5 | 3 |
| Gemini via Antigravity | gemini-3.6-flash-medium | 3 |
AI answers
The protocol contains 30 questions, 3 surfaces and 3 runs, producing 270 individual answers and 90 aggregated readings. The measured OpenAI surface is Codex CLI with web search, not the public ChatGPT interface.
Official tools
The technical panel covers 5 sites and 2 pages per site. PageSpeed was run 3 times on mobile and 3 times on desktop for each page, producing 60 runs. Rich Results was run on all 10 pages.
Site reading
Content form and question coverage use 26 pages that returned HTTP 200 with more than 500 bytes. The sitemap declared 85 addresses, but the report does not turn unread pages into faults.
Limits
The measurement sees neither sales, visitor behaviour nor training data. It does not turn correlation into cause. Eligibility in the Rich Results Test does not promise a display, and a laboratory PageSpeed score does not replace field data.
The overall score is the sum of each score multiplied by its weight, divided by the sum of the weights. The captures published here are retained and can be checked on request.
What the fee buys
An honest answer, not an engine promise.
The €1,500 excl. VAT Diagnostic buys a dated protocol, measured answers, a human reading and an order of action. It can conclude that you should proceed, correct something first, wait or buy nothing.
Frequently asked questions
Five answers before you write to us.
What does the €1,500 excl. VAT Diagnostic contain?
A reading of 30 buyer questions, each asked three times on three surfaces: Claude CLI, Codex CLI with web search for OpenAI, and Gemini via Antigravity, so 270 answers, the list of sources the engines cite, a technical reading of the website and a decision note. This page publishes a complete copy.
Why is this diagnostic public?
Real French is our own website. We can therefore show the real substance of the offer without betraying anyone. A client’s diagnostic stays confidential.
Are the competitors named?
In the document a client receives, yes, measured with the same instruments on the same dates. On this public page they appear as School A to School D, with their real numbers but without their names.
Can the measurement be replayed?
Yes. The questions, the engines, the number of passes and the dates are written on this page. Anyone asking the same questions to the same engines can compare their results with ours. Engine answers vary between passes; that is why every question is asked three times.
Is competitor analysis included in the price?
The citation shelf is: it already shows who gets cited on your questions. The technical panel of a named competitor, measured with the same instruments on the same dates, is an option at €400 excl. VAT per competitor, up to four.