AI transparency audit for business websites
Talking to an AI is fine. Not knowing it isn’t.
Since August 2026, the EU AI Act requires businesses to make it clear when people are dealing with an AI. Taumin checks what a customer actually sees on your website, and shows you where what you say about AI differs from what we observe. Every finding comes with its evidence, its date and the fix.
The site says
“Our advisers are here to help, Monday to Saturday.”
home page · chat window
We observe
- The chat assistant introduces itself with a first name and a photo, with no mention of AI when it opens.signal S07 · medium reliability · observed 24/10/2026 14:32
- No page or notice describes how the site uses AI.signal S20 · medium reliability · observed 24/10/2026 14:35
- Rule potentially relevant: AI Act, Art. 50(1).Taumin reference v0 · status verified
Fictional example, built on one of our test sites. Taumin never publishes a report on a real website.
Why Taumin
The law is only the starting point. What is really at stake is the trust of the people who visit your website.
Your customers’ trust
A customer who finds out they were talking to a bot feels misled, even when it was allowed. Telling them upfront takes one sentence.
Your reputation
Spot a gap before a customer, a competitor or a journalist does, and fix it in time.
Your sales
A dated report answers buyers’ and partners’ questions about how you use AI, without hours of back and forth.
Your time
An outside view of your website in a few pages, with a concrete action for every point raised.
Three scores, never added up
A site can use a lot of AI and say so clearly. Another can use very little and hide who runs it. A single score would blur the two, so we keep them apart.
Use of AI
What we observe of AI on the site: chat assistant, images or text that are probably generated, provenance markers.
Transparency
What the site says about it: a dedicated page, labels next to content, an assistant that introduces itself as an AI.
Risk
What makes the site hard to verify: company missing from registers, very recent domain, team photos found elsewhere, reviews that all follow the same template.
What a report contains
Each finding fits on one card. It says what we saw, where the evidence comes from, how reliable the signal is, and what it does not prove.
- Observation
- The chat assistant introduces itself with a first name and a photo, without saying it is an AI.
- Evidence
- Screenshot of the window and text of the first message, home page, 24/10/2026 14:32.
- Rule
- AI Act, Art. 50(1) and 50(5), status verified. Cited as a rule potentially relevant, not as an established breach.
- What this does not prove
- A human may take over the conversation, and the notice may appear later in the exchange.
- Suggested action
- State in the very first message that the assistant is an AI.
Fictional example. Real reports follow the same format, with a right of reply.
Who it is for
Compliance and legal
Prepare for the transparency obligations of Article 50 with dated findings rather than impressions.
Marketing and communications
Know what your website really shows about AI, before a customer or a journalist notices.
Procurement and partnerships
Review a supplier’s or a partner’s website with the same grid, signal by signal.
Our limits, measured
1 in 3
Roughly one generated text in three is caught on our French-language test bench, when we accept no false alarms on human texts. A text that is not flagged proves nothing.
34% → 8%
On 64 websites of officially certified French businesses, the share wrongly rated high risk, before and after we corrected our method. We publish these figures.
2 signals
Independent ones at least, or one high-reliability signal, before any serious flag. A missing provenance marker proves nothing.
Our latest research
We publish what we learn while building Taumin, mistakes included.
Our first articles are coming soon.
See your website the way a customer does
The first business audits open soon. Write to us at [email protected] to join the waitlist or ask a question about the method.