We scanned 883 EU websites for Article 50 compliance

Name: ClearLabel EU AI Act Article 50 compliance scan, August 2026
Creator: ClearLabel
License: https://creativecommons.org/licenses/by/4.0/

Article 50(1) has been binding since 2 August 2026 and machine-readable marking of AI content is due 2 December 2026. We wanted a number rather than an assumption, so we measured. Raw data and the scanner are open.

97%

of EU online shops running a chat widget showed no AI-disclosure wording (32 of 33)

24%

of EU online shops run a chat or AI widget at all — five times the rate across general EU sites

0.4%

of all sites carried any machine-readable AI-content marking, with the duty landing 2 Dec 2026

92.3%

no-disclosure rate across the broader Tranco sample, confirming this is not an e-commerce quirk

What we found

Two findings are worth separating, because they point at very different populations.

The chat-widget finding is narrow but sharp. Only 4.9% of readable sites ran a detectable chat or AI widget. But of those that did, 92.3% had no AI-disclosure wording anywhere in the page copy. If the vendor's AI mode is switched on — and every vendor we found ships one — that is the exact gap Article 50(1) addresses.

The content-marking finding is broad and stark. Just 2 sites out of 532 carried C2PA Content Credentials or IPTC synthetic-media metadata. Article 50(2) requires machine-readable marking of AI-generated audio, image, video and text, with a transition ending 2 December 2026. This one is not limited to sites with chatbots — it reaches every site publishing AI-assisted media, which in 2026 is most marketing sites.

Vendors detected

Salesforce Embedded Service / Agentforce13

HubSpot Chat3

Zendesk / Zopim3

Intercom2

Freshchat / Freshworks1

Landbot1

LiveChat1

JivoChat1

Userlike1

Every vendor detected ships an AI-agent mode. None of the sites in this sample used a purpose-built LLM-only widget such as Chatbase or Voiceflow, which we read as those tools being more common further down the long tail than in Tranco's upper ranks.

Where the widget actually lives

Of the 26 sites with a widget, 18 had nothing on the homepage — the widget only appeared on a contact, help or support page. Any Article 50 audit that checks the homepage alone will under-report by roughly two thirds.

A second sample, because the first frame was wrong

The Tranco sample above is dominated by news, banking, government and infrastructure — not the businesses that actually run chat widgets. We said so in the limitations, then went and tested it: 180 independent EU online shops, every domain DNS-verified before scanning, 135 readable.

Measure	Tranco EU sample	EU online shops
Runs a chat or AI widget	4.9%	24.4%
Of those, no AI-disclosure wording	92.3%	97.0%
Machine-readable AI content marking	0.4%	0.7%
Readable sample size	532	135

Widget density is five times higher among online shops, and the disclosure gap is worse, not better — 97% of shops with a widget showed nothing. The earlier 4.9% figure was a property of the sampling frame, not of the web.

One vendor dominates

Vendor	Shops
Gorgias	16
Freshchat / Freshworks	4
HubSpot Chat	3
Salesforce Embedded Service / Agentforce	3
Zendesk / Zopim	2
Google Dialogflow	1
Tidio	1
Olark	1

Gorgias accounts for roughly half of every widget we found on these shops. It is Shopify-native and ships an AI Agent that resolves tickets autonomously, which puts a large, concentrated population of merchants squarely inside Article 50(1) — most of them, on this evidence, without a disclosure.

Sample construction: candidate shops were compiled from public retail directories and country e-commerce listings, then every domain was DNS-checked and 32 unverifiable entries discarded before scanning. This frame is curated rather than probabilistic, so treat it as indicative of online retail, not as a random sample of it.

By country

ccTLD	Readable	With widget	Rate	No disclosure
`.be`	24	3	12%	3
`.pl`	33	4	12%	4
`.ee`	10	1	10%	1
`.es`	31	3	10%	2
`.dk`	21	2	10%	2
`.lt`	11	1	9%	1
`.nl`	44	4	9%	4
`.eu`	12	1	8%	1
`.se`	26	2	8%	1
`.fr`	43	3	7%	3
`.fi`	20	1	5%	1
`.de`	43	1	2%	1
`.at`	24	0	0%	0
`.bg`	12	0	0%	0
`.cz`	17	0	0%	0
`.gr`	13	0	0%	0
`.hr`	10	0	0%	0
`.hu`	14	0	0%	0
`.ie`	14	0	0%	0
`.it`	35	0	0%	0
`.lv`	11	0	0%	0
`.pt`	14	0	0%	0
`.ro`	18	0	0%	0
`.sk`	13	0	0%	0

Countries with fewer than 10 readable sites are omitted. Per-country counts are small; treat the ordering as indicative, not as a ranking.

Method

Sampling frame. The Tranco top-1M list, filtered to 26 EU ccTLDs, taking the highest-ranked domains per country up to a per-country quota. 703 domains in total. Tranco was chosen because it is a published, reproducible ranking rather than a hand-picked list.
Collection. 16 August 2026. Each domain's homepage was fetched; where no vendor was found, up to three country-appropriate contact/help paths were tried (/kontakt, /contacto, /contatti and so on).
Detection. Regex fingerprints for 36 conversational-AI vendors against page source, plus AI-disclosure phrase matching in 12 EU languages, plus C2PA and IPTC marking signals. The scanner is open source; the vendor dataset is CC-BY-4.0.
Coverage. 532 of 703 domains (76%) returned readable HTML. The rest were bot-protected (HTTP 403), did not resolve, or refused connections.

Limitations, stated plainly

This measures signals, not compliance. We cannot see inside anyone's vendor console. A site running Zendesk with the AI agent switched off has no Article 50(1) duty, and our scan cannot tell the difference. "No disclosure wording found" is evidence of a gap worth checking, not a finding of breach.
The frame skews away from e-commerce. Tranco's upper ranks are dominated by news, banking, government and infrastructure. The 4.9% widget rate is therefore a floor for the general web, and almost certainly understates online retail, where chat widgets are far more common.
Disclosure detection is a proxy. We search visible page copy. Wording that appears only inside the chat window after it opens will be missed, so the 92.3% figure may overstate the true gap.
Bot protection is not random. The 171 unreadable domains skew toward larger sites, which likely have more compliance resource. This may bias findings pessimistically.
We sell a product in this space. Treat the framing accordingly, and check the raw data — it is published precisely so you do not have to take our word for it.

Data and reuse

Per-domain results, the sample frame and the scanner are in the public repository. Licensed CC-BY-4.0 — reuse it, and corrections by pull request are welcome.

ClearLabel (2026). Article 50 compliance signals across 883 EU websites. https://clearlabel.eu/study/

Check your own site against the same scanner

Same detection engine, same 36 vendors, same 12 languages. Free, no signup, runs in your browser.

Scan your site