On 2 August 2026 the transparency obligations in Article 50 of the EU AI Act (Regulation (EU) 2024/1689) began to apply. For marketing teams the whole thing compressed into four words, AI content must be labelled, and that compression has already produced bad decisions. The rules are narrower, and stranger, than the headline suggests. This article sets out what the AI content disclosure regime actually asks of a marketing operation, which parts bite now and which have already slipped to December 2026, and what the likely effect on search and AI-answer visibility is. The evidence base is the regulation text, the European Commission's Article 50 guidelines published in 2026, and analysis from Sidley, Stibbe and Herbert Smith Freehills Kramer. Teksyte sells SEO services and operates several of the sites it cites here, which is disclosed now and again wherever it bears on the point.
What Article 50 actually requires
Article 50 places two different duties on two different parties. Providers of systems that generate synthetic audio, image, video or text must mark those outputs in a machine-readable format that a detector can read as artificial. Deployers, meaning companies that use those systems, must disclose deepfakes and must disclose AI-generated text that informs the public on matters of public interest. That is the whole surface area, and most marketing output sits outside three of the four triggers.
The load-bearing word is deployer. According to the European Commission's Article 50 guidelines, published in 2026, the disclosure duty on AI-generated text applies only where the text informs the public on matters of public interest, and it falls away entirely where the text has undergone human review with a natural or legal person holding editorial responsibility for publication (Confirmed, Commission guidelines and Article 50(4)). A product page is not a matter of public interest. A reviewed opinion piece on a policy change is, and then the exemption decides the outcome. Most marketing teams are arguing about a duty that, read closely, they can discharge with a documented editorial process they should already run.
Why the marking obligation already slipped to December
Start with the date that has quietly moved. The marking obligation in Article 50(2), the machine-readable one that sits on the model providers, was not left alone. Under the Digital Omnibus provisional agreement reached between the Council and the Parliament on 7 May 2026, marking for systems placed on the market before 2 August 2026 is deferred to 2 December 2026 (Confirmed, Sidley analysis of the Omnibus, 24 June 2026). The obligation the press treated as a hard Sunday deadline had a four-month hole cut in it before that Sunday arrived.
That single fact reframes the risk calculation. The Article 50 transparency obligations are law, and they carry fines of up to 15 million euros or 3% of worldwide annual turnover (Confirmed, Article 99). But an obligation whose most technical component was postponed under commercial pressure, three months before application, is not one being enforced with maximal appetite. The enforcement machinery also runs through national market-surveillance authorities that, in August 2026, are still being designated and resourced (Observed, across member-state implementation reporting). The rule is real. The idea that it detonates uniformly on day one is not.
The editorial review exemption changes the whole game
The common belief, repeated across marketing coverage since the spring, is that every piece of AI-assisted content published in the EU now needs a visible AI label. That belief is wrong, and the error matters because it is pushing teams either to strip AI out of workflows or to attach disclaimers to everything, both overreactions.
Here is the dismantling. The text obligation is confined to AI-generated text that informs the public on matters of public interest, and even inside that box it does not apply where a human carried out editorial review and a person holds editorial responsibility for the publication (Confirmed, Article 50(4) and Commission guidelines, 2026). Standard newsroom review clears the bar. So does a named editor signing off an agency's analysis post. The exemption was written for exactly the situation most agencies are in. The practical consequence is that the compliant path is not a label, it is a documented human editorial process with a named owner. An agency that cannot name who reviewed a piece has a governance gap, not a labelling gap, and a label would not have fixed it.
How to make an AI-assisted page compliant
Work the triggers in order, because most pages exit at the first or second step.
Ask whether the page is one of the four Article 50 categories at all. Direct chatbot interaction, deepfake media, biometric or emotion systems, or public-interest text. A commercial service page is none of these and needs no disclosure (Confirmed, Article 50 scope).
If it is public-interest text, ask whether a human reviewed it and whether a named person holds editorial responsibility. If yes, record that owner and the review date in the page's own governance log. The exemption applies (Confirmed, Article 50(4)).
If no human review exists, either add it, which is the cheaper fix, or add a clear and distinguishable disclosure at the first point the reader meets the text, which the guidelines require to be accessible (Confirmed, Commission guidelines, 2026).
For any synthetic image or video, keep the provider's machine-readable marking intact and do not strip content-credential metadata on export, because that metadata is the provider's compliance artefact and, increasingly, a signal downstream systems read.
What the restructuring actually looks like
The response now moving through agency operations is not a technology project, it is a governance one, and it follows a consistent shape (Observed across agency workflows adjusting to the August application date). A named editor is assigned to each public-facing property. A dated sign-off is recorded against every piece that could count as public-interest text. Provenance metadata on media is preserved through the content pipeline rather than stripped on export. None of that requires new software, and all of it is defensible under Article 50(4). The firms treating this as a labelling problem are buying disclaimer widgets. The firms treating it as an editorial-accountability problem are writing down who signed off what, which is the thing the law actually asks for and the thing a serious publication should have had already.
What disclosure does to search and AI citation
Disclosure and labelling do not, on current evidence, change how a page ranks in classic search. Google has not documented any ranking effect from an AI-disclosure notice, and there is no confirmed signal tying a visible label to position (Confirmed as an absence, Google Search Central carries no such statement as of August 2026). The live question sits one layer over, in AI answers.
Whether machine-readable marking changes how often generative engines cite or reuse a page is genuinely unsettled (Uncertain). Marking is designed to make synthetic content detectable, and it is reasonable that answer engines will lean on provenance signals as they mature, but no published study yet isolates a citation effect from Article 50 marking specifically, and anyone asserting one is guessing. What would settle it is a controlled test, matched pages with and without preserved content credentials, measured for citation frequency across named engines over a fixed window. The defensible position today is that AI content disclosure is a governance and provenance question first, and only a possible visibility question second, and to build for the former while measuring the latter rather than trading rumours about it.
What to do differently on Monday
Do not label everything, and do not tear AI out of the workflow. Both are answers to a rule that was not passed. Instead, sort your pages against the four Article 50 triggers, and for the only category that touches most marketing work, public-interest text, install a named editorial owner and a dated review record. That single change discharges the AI content disclosure duty for the bulk of what an agency publishes, and it is a process worth having regardless of the law. Keep provider marking and content-credential metadata intact on media exports. Then measure, rather than assume, whether provenance signals move AI citation, because that part is still uncertain and will stay uncertain until someone publishes a real test.