Method
Taumin reads a website the way a patient investigator would: it notes what the site claims, looks in public sources for what confirms or contradicts it, and records every finding with its evidence and the time it was observed.
How an analysis works
- Reading the site. The main pages, the legal notice, the “about”, team and transparency pages, the reviews on display and the chat window when there is one.
- External checks. Company registers (in France, the national business directory; in the UK, Companies House; in the EU, VIES for VAT numbers), geocoding of the address, domain registration date, archives of the site, and a search for the team photos elsewhere on the web.
- Recording signals. Each signal found is filed under a family, with its reliability and its evidence.
- Three separate scores. Use of AI, transparency and risk, each calculated on its own.
- Report. One card per signal, then the trust indicators found, the signals we could not assess, and the method used.
25 signals in eight families
Taumin’s reference framework describes 25 signals. For each one, it sets out the method, the reliability, the rules potentially relevant and the cases where the signal can mislead.
- Verifiability (6): domain age, company identity shown on the site and found in registers, verifiable address, consistent site history.
- Provenance (2): “AI-generated” markers read from the files (C2PA, IPTC), visible labels.
- Interaction with an AI (2): chat assistant presented as a human, or clearly identified as an AI.
- Identities (3): team photos that are probably synthetic or found elsewhere, directors missing from registers.
- Reviews (5): reviews that follow the same template, publication spikes, testimonials that cannot be traced, undisclosed insider reviews.
- Generated content (3): texts, articles and images that are probably generated but presented as written or photographed.
- Transparency (2): use of AI observed with no mention at all, or a page that explains how AI is used.
- Network (2): template reused on other domains, contact details shared with sites already flagged by an authority.
Reliability and safeguards
Each signal has a high, medium or low reliability, set from our measurements and from what the method can miss. A few rules apply to every report:
- No serious flag without two independent signals, or one high-reliability signal.
- A low-reliability signal never triggers a flag on its own.
- A missing provenance marker proves nothing: most generated images carry none, and many sites strip them.
- For faces, we only check whether a photo is probably synthetic or published elsewhere. Taumin never identifies a person.
- A site protected against bots is not bypassed. The report marks it “cannot be assessed” and says why.
What we have measured
We publish our error rates, because a tool that does not say where it goes wrong does not deserve to be believed.
| Test | Result | Consequence |
|---|---|---|
| French-language text detector, 900 human texts and 55 generated texts | At the threshold that flags no human text of at least 100 tokens (a short paragraph), 16 of 43 generated texts caught | Low-reliability signal, never used for risk |
| Image detectors, calibrated on 2,500 real images | About 1 real image in 100 wrongly flagged; 90% of Flux portraits caught, 10% of product photos from the most recent generators | At least two flagged images before any signal; an image that is not flagged proves nothing |
| Risk on 64 websites of businesses certified by the French state | Wrongly rated high risk: 34% before correction, 8% after | Incomplete legal notices count as a single signal; “incomplete identity” reported separately |
Rules we cite
A report may cite a rule as “potentially relevant”, for example Article 50 of the EU AI Act or the rules on fake reviews. We only cite rules whose text and date of application we have checked, and we say so. A cited rule does not establish a breach: only a lawyer or an authority can do that.
Right of reply
Every audit report includes a right of reply. A business that disputes a finding can write to us; we check, correct what needs correcting and show the date of the correction.