EU AI Act and AI visibility: what brands must change

In short

The EU AI Act does not magically make a brand visible. Article 50 introduces transparency obligations in specific cases. For a brand, the editorial task is to check the claims circulating about the company, their sources and their approval. In a non-representative subsample of our barometer, 106 out of 657 responses spontaneously cited the brand, or 16%. Compliance does not buy that presence, but it encourages brands not to leave it as an unmanaged blind spot.

The question is often framed badly. A marketing leader asks, "Do we need an AI Act-ready website to show up in ChatGPT?" Legal counsel rightly answers through roles, use cases and obligations. Between those two answers sits a very practical issue: what assistants already say about your brand, products, prices, commitments and competitors.

This is not another legal guide to the AI Act. Our existing AI Act guide for companies already covers risk levels, timing and obligations. This article is about visibility. Not visibility as a promise of a citation, but as an evidence discipline in an environment where generated answers can influence a decision before a visitor reaches your site.

The AI Act does not rank your website, but Article 50 raises the value of evidence

The EU AI Act does not say that a structured FAQ, product page or well-written article will be recommended more often by an AI system. Nor does it create a compliance badge that makes a brand rise in conversational answers. Saying otherwise would confuse regulation with source-selection mechanisms.

Article 50 has applied since 2 August 2026. It includes informing people when they interact with AI and machine-readable marking of certain generated or manipulated content. For systems placed on the market before 2 August 2026, this marking and detection obligation under Article 50(2) applies only from 2 December 2026. Article 50(4) also requires clear labelling of AI-generated or manipulated text published to inform the public on matters of public interest when it has undergone neither human review nor editorial control. The exemption requires human review or editorial control, together with a natural or legal person holding editorial responsibility for the publication. Simple spelling or grammatical correction is not sufficient. The detail depends on the system, the organisation's role and the context of publication.

For content, communication and product teams, we recommend a simple editorial rule: link every commercially meaningful statement published with AI assistance to a source, a responsible person and an appropriate review. This recommendation is not a general AI Act obligation for every sentence. It helps address use cases covered by the regulation and correct an inaccurate AI answer that appears in a buying journey.

One idea needs to go: "If AI generated the draft, it must be unusable." The AI Act does not say that. The problem is a lack of control. An AI answer can be a draft, a summary or a starting point. It should not become a public claim with no owner, no source and no review, especially when it concerns price, certification, performance, availability or a regulated promise.

The visibility risk: plausible but wrong information

A brand incident does not necessarily look like a spectacular deepfake. It can start with an answer that sounds reasonable: an old address, a discontinued offer, an invented certification, the wrong market category, or a fabricated URL. The reader may not realise they have just received an approximation. They may simply choose a competitor, contact the wrong department or conclude that the brand is unreliable.

The AI Labs Audit barometer covers 104 audits and 6,820 AI responses collected between September 2025 and June 2026, using a non-representative sample. In a subsample of 13 brands with entity extraction, 106 out of 657 responses spontaneously cited the brand, or 16%, in questions where its name was not provided. Separately, 43 out of 65 brands, or 66%, were associated with at least one invented URL. These are not AI Act compliance statistics. They do not demonstrate that a brand has broken a rule. They show why "being mentioned" is not a sufficient KPI.

Team question	Too-short answer	Defensible control
Is our brand mentioned?	Yes, so everything is fine.	In which queries, by which models, with which facts and which sources?
Is the content compliant?	We added an AI label.	What is our role, which use case is involved, who approves the message and what evidence is retained?
An answer is wrong. Does it matter?	It is only AI.	Is the claim being repeated, where can the primary source be corrected, and how will recurrence be monitored?
Is our site well prepared?	It has Schema.org markup.	Is the main content clear, dated, sourced and consistent with the pages that carry authority?

This distinction matters. Marking synthetic content is governed by a defined legal framework. The quality of brand information also belongs to commercial trust, sector-specific law and your editorial governance. The subjects overlap, but they are not interchangeable.

What transparency changes in editorial work

A brand publishing with AI assistance does not need to place a warning above every sentence. It first needs to know what was produced, how, for which audience and with what human control. Article 50 addresses defined situations; an internal policy can be simpler than the legal text if people can actually follow it.

One operating rule works well: no commercially material claim should go live without a primary source or an identified owner. “Commercially material” includes pricing, contract terms, product capabilities, study results, certifications, comparisons, security facts and commitments. This is not a marketing constraint. It reduces gaps between the official website, product documentation, press material and answers that may summarise them.

1. Separate a draft from a declaration

Generated text can make a first draft faster. The final statement still belongs to the organisation that publishes it. Keep the source supporting the claim, and do not ask a writer to “make a fact more compelling” when they cannot verify it. That small discipline prevents a working assumption from turning into a message repeated across the site.

2. Date claims that expire

Prices, product capabilities, delivery rules, executives and benchmark results age. A page without a date can be correct on publication day and wrong a quarter later. On reference content, state the verification date and assign a content owner. AI systems will not always respect that freshness, but you will not be giving them an outdated source yourself.

3. Maintain a correction record

When an AI answer attributes a false fact to a brand, the right response is not always to publish ten articles repeating the error. First verify the official source, correct the page that should carry authority, then record the claim, date, model, query and correction. A validated question-and-answer record is more useful than a publishing sprint.

4. Do not confuse public copy with compliance documentation

A brand page should stay readable. Internal documentation can be more detailed: tools used, validation steps, sources consulted, approvals and page version. One serves the reader. The other makes it possible to demonstrate that the team did not let a machine publish sensitive information alone.

Measure visibility with a framework that works for legal and marketing teams

AI-answer monitoring becomes useful when it moves beyond a vague impression. “We show up quite well” cannot trigger an action. A useful framework answers four questions: does the brand appear spontaneously, is the information accurate, are competitors preferred, and are the sources associated with the answer owned or at least understandable?

At AI Labs Audit, an audit separates spontaneous brand presence, invented claims, hallucinated URLs, share of voice against competitors, and the difference between native-model answers and answers with web search. The purpose is not to announce that a model “knows” you. It is to find discrepancies that need an owner. A missing brand may be a content or authority issue. A false certification may require a source correction and legal validation. Those are different actions.

Our AGS methodology also follows a rule we consider healthy: a flattering answer that invents facts must not inflate a score. A compliance team will find that obvious. An acquisition team may not. Yet a high score built on errors is false comfort. A lower, accurate signal with a correction list is more valuable than visibility no one can defend.

Three mistakes to avoid after 2 August 2026

Turning the AI Act into an optimisation claim

You can explain a real transparency policy. Do not present it as a way to earn citations from assistants. Search systems and models select information through their own systems. Compliance can reduce certain risks; it does not guarantee indexing, recommendation or conversion.

Adding an AI label to human-reviewed copy without analysing the use case

Approximate transparency can cause more confusion than it solves. Article 50 has a specific scope, and the European Commission has published guidelines. Seek analysis that fits your role, particularly if you operate a chatbot, generate synthetic content at scale, or publish content on matters of public interest without human editorial review.

Treating every inaccurate answer as a writing problem

A wrong answer can come from an old local listing, reseller page, directory, legacy indexed page, redirect or third-party source. Before rewriting the homepage, find the origin. The right fix may be a dated product page, structured-data update, correction with an outside publisher, or a clearer help-centre answer.

A control plan that connects compliance and visibility

• Inventory the AI systems your organisation provides or deploys, then separate them from ordinary use of writing-assistance tools.
• Have competent legal advisers classify use cases that may trigger transparency obligations, especially chatbots, synthetic content, deepfakes and public-interest text published without human review.
• Define claim categories that always require approval: price, security, compliance, performance, research, commitments and comparisons.
• Keep a primary source and verification date for every sensitive claim on reference pages.
• Measure what models say when they are not given your brand name, then separate absence, factual error, invented URL and competitor advantage.
• Correct the primary source before multiplying secondary content. An extra article does not repair an incorrect official listing.
• Document corrections and outcomes without promising that an update will produce a citation or remove an error immediately.

This is less glamorous than “becoming visible in AI”. It builds stronger assets: traceable claims, pages that do not contradict each other, a team able to respond to an error, and measurement that does not merge awareness, accuracy and compliance into one flattering number.

The question your brand should ask

The question is not, “Are we AI Act compliant so we can appear in AI?” It combines two different responsibilities. Ask instead: “When our organisation uses or deploys AI, do we know which obligations apply? And when AI systems talk about us, can we distinguish useful presence from incorrect information?”

The first answer should come from legal and operational analysis. The second should come from data, sources and ongoing observation. Their connection is not a visibility trick. It is a consistency requirement. When you publish reliable information, maintain it and monitor what is said about you, you reduce the risk of letting your reputation be assembled from a draft, an archive or an invention.

Useful sources

• Regulation (EU) 2024/1689 on artificial intelligence.
• European Commission FAQ on Article 50 transparency obligations.
• European Commission guidelines on AI transparency obligations.
• AI Labs Audit barometer, data collected between September 2025 and June 2026, and the AI Labs Audit Observatory.