claude skills / Agencies and freelancers pack / audit-de-visibilite-ia / en français

AI Readability Audit.

What AI tools really answer about your client, measured, dated, then measured again after the fixes.

See what it produces
  • ChatGPT
  • Gemini
  • Perplexity
  • Google AI Overviews
  • Google AI Mode
  • Spreadsheet
12 to 20fixed questions in four groups, same words in every survey
3 runsper question and per engine, in fresh conversations
4 to 6 wksbetween the fixes and the next survey, then every month

orders of magnitude, depending on your volume

the problem

What gets in the way today.

  • 01The client asks "does ChatGPT talk about us?" and the answer changes every time.
  • 02A single screenshot is used as proof, taken from an account that already knows the brand.
  • 03An AI gives the client's old address, and nobody knows where the error comes from.

input, output

You give this. You get this.

The client, their facts and the previous survey

Client: Menuiserie Vallat, 8 people. Custom windows, doors and fitted joinery, for homeowners and architects, one département (county).
Reference facts: workshop moved to a business park in 2025, fitted joinery page created in September.
Known competitors: three joinery firms in the département.
Engines: ChatGPT with search, Gemini, Perplexity, Google.
Previous survey: no. 1 of 1 September, list MV-Q1.

$ claude "Run survey no. 2 with the same question list and compare it with no. 1."

Survey no. 2, before and afterexample

Menuiserie Vallat | Survey no. 2, 3 to 5 November 2026

List MV-Q1, 12 questions | ChatGPT, Gemini, Perplexity, Google

Read this first

These answers change from one day, user and place to another. This survey is a dated sample following a fixed protocol, not a guarantee.

Result

Score: 24/100 (survey no. 1: 13/100)

Search: 17 of 96 | Price and method: 6 of 24 | Brand: 12 of 24, 1 penalty

The client now appears in Perplexity and Google for custom fitted joinery, after the dedicated page went live. It is still absent from ChatGPT for windows, where two trade directories dominate the sources.

Error found

ChatGPT, question C2: "The workshop is on rue des Tanneurs."

Correct fact: the workshop moved in 2025. Likely source: an outdated directory listing.

Sources cited (group A, 96 answers)

Trade directory no. 1: 31 answers, client absent

Trade directory no. 2: 22 answers, incomplete listing

Client's website: 9 answers

Fixes

1. Create the listing on directory no. 1 | Agency | 15/11

2. Correct the address on directory no. 2 | Client | 15/11

3. "Custom wooden windows" page with pricing method | Agency | 30/11

Next survey: week of 7 December, list MV-Q1 unchanged

the method

What Claude does, in order.

  1. step 1Fix the questions

    Searches without the client's name, price and method, brand, multi-turn conversation. The client approves the list.

  2. step 2Survey by the protocol

    Private browsing, fresh conversations, 3 runs per engine, answers word for word, date, location and screenshots.

  3. step 3Score

    0 to 3 per question and per engine, accuracy penalty down to -2, a score out of 100 and three sub-scores.

  4. step 4Read the sources and fix

    Directories and sites cited, errors linked to their likely source, fixes ranked, date of the next survey.

demo in French / click, try it

Your turn.

An interactive demo of the skill. It is in French, the pack works in any language: ask Claude to answer in English.

ChatGPT

the film / 44 seconds, with sound

The pack film.

the pack film

In 8 seconds.

the guardrails

What the skill never does.

  • No promise of citation, ranking or timing: the disclaimer opens every report.
  • Search questions never contain the client's name.
  • Nothing is simulated: without access to an engine, the skill prepares the protocol and scores your surveys.
  • A result is never used to tell a prospect "ChatGPT does not cite you" without a same-day test.

Questions

Why three runs per question?

Because a single answer proves nothing: it changes from one conversation to the next. We record a rate, "cited 2 times out of 3", across three fresh conversations.

Can we promise the client they will be cited by ChatGPT?

No. Nobody controls these answers. The skill measures, links each finding to a cause the client can fix, and measures again with the same questions.

How is this different from the Website Audit?

The Website Audit measures AI readability: whether crawlers can read the site. This skill measures what assistants answer to a buyer. The two work together.

Is an llms.txt file enough?

No, it is a complement. The fixes that matter first: errors at the source, one clear page per service and per area, the same identity everywhere, a presence on the sources the engines cite.

the pack

In the pack: Agencies and freelancers pack.