16 days before the EU AI Act's transparency rules apply, compliance is invisible

What this means

The striking result is not that few sites disclose — it's that compliance can barely be audited by machine at all. Nearly a quarter of major EU sites block identified crawlers before any check can happen. Another 38% load their entire tooling stack, chat widgets included, through tag managers like Google Tag Manager, where nothing is visible in the served HTML. Where chat platforms are visible (Intercom, HubSpot, Zendesk, Freshchat, LiveChat, Ada and others), we found not a single AI-disclosure sentence in the page source.

For large brands this doesn't necessarily mean non-compliance — a disclosure can be rendered inside the chat window at runtime. But it does mean that if a market surveillance authority ever wants to verify Article 50 at scale, current websites give them nothing to check without a full browser and a manual conversation with each bot. Meanwhile, the Commission's draft guidelines are explicit that disclosures buried in terms & conditions don't count, that a bot name alone isn't enough, and that the notice must appear at the latest at the first interaction.

For small and mid-size businesses, the picture is more direct: most SMB sites embed their chat vendor's snippet straight into their HTML, where our scanner can see it — and where the absence of a disclosure is equally visible. Several major chat vendors ship their AI labels off by default and place the legal responsibility for end-user disclosure on you, their customer, in their terms.

Method & limitations

One polite HTTP GET per homepage on 17 July 2026, from an identified crawler (user-agent "Article50Scanner"), no JavaScript execution, 2.5 MB read cap. Detection = signature matching against 38 chat/AI platforms plus disclosure-phrase heuristics in English. Limitations: static analysis cannot see widgets injected at runtime; non-English disclosures outside our phrase list may be missed; 403-blocked sites may well run compliant disclosures we cannot observe. We publish the numbers with those caveats — they measure machine-verifiable transparency, not legal compliance.