AI Content and Google: Enforcement and the EU AI Act

Since 2024, Google and EU lawmakers have been pulling in the same direction: both target mass-produced, low-quality AI content. Google fights it algorithmically and with manual actions, and from August 2026 the EU AI Act adds a legal transparency obligation. For B2B sites in the DACH region, this creates a dual requirement: content has to prove its added value to search engines and, at the same time, comply with labeling law. This article explains where Google draws the line between scaled content abuse and legitimate AI use, what Article 50 of the AI Act actually requires, and what a reliable editorial and compliance workflow looks like.

Why AI content enforcement is now business-critical

AI-written text is now mainstream. According to an international analysis of 900,000 newly created English-language web pages, 74.2 percent of them contained AI-generated content (international survey, April 2025). The distribution is telling: only 2.5 percent were pure AI text, 25.8 percent purely human, and 71.7 percent a human-AI mix (international survey). So the norm is the hybrid workflow, not the purely synthetic document. This exact distinction is central for Google and for the AI Act.

AI has also reached broad adoption in the DACH region. In Austria, 30 percent of companies with 10 or more employees were already using artificial intelligence in 2025, up from 20 percent the year before (Austria, Statistik Austria, survey February to July 2025). If you use AI to create content, you therefore increasingly fall within the scope of the transparency obligations that take effect in 2026.

The risk of getting it wrong is concrete. In the course of the March 2024 core update, independent monitoring found that 837 of 49,345 observed sites were removed from the Google index completely, around 1.7 percent (international survey). An accompanying analysis found signs of AI-generated content on 100 percent of the affected sites, and on 50 percent of them, 90 to 100 percent of the posts came from AI (international survey). Complete deindexing means zero organic traffic. For a B2B site whose pipeline depends on search, that means losing the channel entirely.

How Google draws the line: scaled content abuse vs. legitimate AI content

Google does not treat AI content as spam across the board. Its long-standing position is that using automation, including generative AI, is spam only if the main purpose is to manipulate rankings. With the March 2024 core update, Google introduced three new spam policies: scaled content abuse, expired domain abuse, and site reputation abuse (international). What counts is intent and outcome, not the creation method.

Google defines scaled content abuse as generating many pages with the primary purpose of manipulating rankings rather than helping users. The policy applies regardless of whether content is created through automation, human effort, or a combination (international). Publishing thousands of thin pages is therefore a violation even when humans were involved. Conversely, an AI-assisted article with genuine added value is not a violation.

Google itself documents the economic effect of its quality push: after the March 2024 update, it saw 45 percent less low-quality, unoriginal content in search results, more than the expected 40 percent (international, rollout completed April 19, 2024). In practice, Google decides along three axes:

• Added value and originality: Does the page offer information, analysis, or experience that does not exist in the same form elsewhere? Reworded platitudes at scale are the core pattern of scaled content abuse.
• Primary purpose: Is the page built for users or for rankings? Programmatic SEO templates without standalone value per page quickly tip into abuse territory.
• E-E-A-T: Are experience, expertise, authoritativeness, and trustworthiness recognizable? Missing authorship, missing evidence, and missing hands-on experience are warning signs.

The distinction between a manual action and an algorithmic demotion matters. A manual action is triggered by a reviewer, reported in Google Search Console, and requires a reconsideration request once the issue is fixed. An algorithmic demotion from a core update does not appear as a notification. It shows only as a traffic decline and recovers only with a later update.

Site reputation abuse: the underestimated risk for B2B publishers

Site reputation abuse, colloquially known as parasite SEO, affects sites that rent out their own authority to third parties. Google lists it as one of the three new spam policies of 2024 (international). It covers third-party content published without close oversight by the host and primarily to exploit the domain's reputation. For B2B publishers with guest posts, partner content, or sponsored sections, this is a real risk area.

A clean setup here means editorial oversight of all third-party content, clear separation and labeling of sponsored content, no hosting of outside subportals just to pass on ranking signals, and the same quality requirements for external authors as for internal ones. Guest posts are unproblematic as long as they fit the topic, are reviewed, and deliver standalone value.

EU AI Act Article 50 in practice

The EU AI Act creates a legal obligation that is independent of how search engines rate content. Article 50 governs transparency obligations and applies from August 2, 2026 (EU). It addresses two roles with different obligations.

Providers, meaning suppliers of AI systems including general-purpose models that generate synthetic audio, image, video, or text content, must ensure that outputs are marked in a machine-readable format and detectable as artificially generated or manipulated (EU). Technically, this points to standards such as C2PA Content Credentials, which embed cryptographically signed provenance metadata in the file. For systems that were already on the market before August 2, 2026, this marking obligation applies only from December 2, 2026, according to the Digital Omnibus on the AI Act, Regulation (EU) 2026/1744 (EU, Art. 111(4) AI Act). The Omnibus does not postpone the disclosure obligations of deployers.

Deployers, meaning the organizations that use an AI system, have disclosure obligations. Anyone who generates deepfakes must disclose that the content was artificially generated. And anyone who generates or manipulates text published with the purpose of informing the public on matters of public interest must disclose that the text was artificially generated or manipulated (EU). For publishers, this text obligation is the key point. It applies only to content with a public information purpose. Also exempt is text that has undergone human review or editorial control and for whose publication a natural or legal person holds editorial responsibility (Art. 50(4), subparagraph 2). For B2B publishers with a documented editorial process, this exception decides whether labeling is needed at all. It does not apply to deepfakes. The scope for purely commercial product copy is narrower and has to be assessed case by case.

The penalty structure is substantial but tiered. Violations of the prohibited practices under Article 5 can be punished with up to €35 million or 7 percent of worldwide annual turnover (EU). For violations of other obligations, explicitly including the transparency obligations under Article 50, a lower tier of up to €15 million or 3 percent of worldwide annual turnover (EU) applies. For SMEs and startups, the lower of the two amounts applies in each case. So the often-cited €35 million concerns prohibited practices, not the labeling obligation; for Article 50 violations, the €15 million tier is the relevant one.

Compliance workflow for DACH and Austrian B2B sites

The first step is clarifying your role. As a rule, a B2B site that uses a third-party model such as a commercial LLM is a deployer, not a provider. Machine-readable marking under Article 50(2) is primarily a provider obligation. The deployer has the disclosure obligation for deepfakes and for text that informs the public, unless that text has been editorially reviewed and someone holds responsibility for it.

In practice, the following structure works well:

• Role and inventory review: Record which content is created or edited with AI and which of it conveys factual information to the public. Only the latter can trigger the text disclosure.
• Disclosure texts and labels: For deepfakes and for informative AI texts that are not editorially reviewed and under editorial responsibility, provide a clear, visible notice. This mandatory notice has to be understandable and easy to find, not hidden in the source code. For editorially reviewed texts with assigned responsibility, a note about AI assistance is voluntary.
• Machine readability: Where you operate your own generative systems, check the provider marking via Content Credentials and make sure outputs remain recognizable as AI-generated.
• GDPR and consent: The AI Act labeling is a content transparency obligation and does not replace consent under data protection law. If AI features process personal data or load tracking scripts, GDPR compliance and consent banners are still required in their own right.
• Documentation: Document the editorial and review process. In a dispute, this lets you prove that the exception for text under editorial responsibility applies.

Human-in-the-loop editorial process as protection against scaled content signals

The most effective protection against scaled content demotions is a genuine editorial process, not a label. Google stresses that the spam policy applies regardless of the creation method; what decides is usefulness. A reliable workflow includes clear quality gates:

• Fact-checking: Every statement with a number, date, or claim is checked against a primary source. Generative models invent plausible-sounding but false facts; unchecked output is a direct quality risk.
• Source citations: Key statements link to reputable sources. Evidence increases trustworthiness and makes content citable for AI systems.
• Original data and experience: Your own project figures, practical examples, and assessments lift a text above interchangeable AI boilerplate. That is exactly what Google rates as experience.
• Editorial responsibility: A named person reviews the final content and takes responsibility for it. Review and named responsibility together are the precondition for dropping the labeling obligation for AI text under Art. 50(4).

This process solves both problems at once: it prevents scaled content signals and creates the documented editorial responsibility that can remove the disclosure obligation for AI text that informs the public. This exception does not apply to deepfakes.

AI crawler management: weighing content protection against GEO visibility

If you publish AI content, you have to decide whether AI crawlers may read your content. Block rates are rising: according to an international analysis of around 140 million websites, GPTBot from OpenAI is the most blocked AI crawler, at 5.89 percent of all websites and up to 7.3 percent across all robots.txt files (international survey). The block rate of ClaudeBot grew fastest, at plus 32.67 percent, the strongest growth of any crawler in the past year (international survey).

You control this in robots.txt via the respective user agents: GPTBot (OpenAI), ClaudeBot (Anthropic), CCBot (Common Crawl), and Google-Extended (Google AI training, separate from Googlebot). llms.txt is also emerging as a hint file for AI systems. The trade-off is strategic: a block protects content from use in training and answers but reduces the chance of being cited in LLM answers. If you are betting on GEO visibility, differentiate between training crawlers and retrieval crawlers instead of blocking across the board. Blocking Google-Extended also does not affect normal indexing by Googlebot.

Technical SEO foundation for large content volumes

AI lowers production costs and thereby magnifies the volume problem. If you scale, you need clean technical control over indexing. Otherwise you waste crawl budget and create duplicate content signals that look like scaled content abuse.

The performance baseline remains mandatory. According to HTTP Archive, only 59 percent of pages achieved a good LCP score on mobile in 2024, while INP was at 74 percent and CLS at 79 percent (international survey). That makes LCP the hardest Core Web Vitals metric to pass. With large content volumes, you also need three additional levels of control:

• Canonical and robots: Unambiguous canonical tags and targeted noindex directives prevent template variants from being indexed as standalone thin pages.
• XML sitemaps: Up-to-date sitemaps direct the crawler to the URLs that actually matter and take pressure off the crawl budget.
• IndexNow: IndexNow lets you actively report changes to participating search engines, which speeds up discovery for large, frequently updated inventories.

GEO/AEO as the answer to AI Overviews and zero-click

The classic click logic is eroding. According to Pew Research, US users who saw an AI Overview clicked a traditional search result in only 8 percent of cases, compared with 15 percent without an Overview, and only 1 percent clicked a source link inside the Overview (USA, March 2025). Another international analysis quantifies the effect on position 1: there, the presence of an AI Overview correlates with a 58 percent lower average click-through rate (international survey, December 2025).

This shifts the focus from pure ranking optimization to optimizing for citability. GEO and AEO aim to get you named in the AI answer itself. Effective levers are structured data and schema markup for machine readability, clear extractable definitions, verified statistics with source citations, and a structure that makes individual statements citable. These measures align with good E-E-A-T practice and work in classic search as well as in LLM answers.

Recovery after a manual action or a core update hit

The recovery path depends on the type of hit. With a manual action, Search Console names the cause. The process: identify the violation, remove or substantially improve the affected content, then submit a reconsideration request. With an algorithmic demotion from a core update, there is no notification and no reconsideration request. The only remedy is to raise content quality site-wide and wait for a reassessment in a later update.

The diagnostic core is a content audit. Thin content, redundant template pages, and unsupported mass-produced texts are identified and then deleted, merged, or enriched with genuine added value. The scale of the enforcement risk shows in the finding that after the March 2024 update, around 1.7 percent of the observed sites were deindexed completely, all with AI content signals (international survey). Realistically, recovery is not a quick process: manual actions can be lifted within weeks after a successful reconsideration request, while recovery from a core update can take several months until the next suitable update.

Next.js and rendering architecture for AI-scaled content

For scaled content, the rendering strategy is decisive. Server-side rendering (SSR) and static site generation (SSG) ensure that content is in the initial HTML without client-side JavaScript and is therefore crawled and rendered reliably. For large, frequently changing inventories, incremental static regeneration is a workable middle ground between static delivery and freshness.

Programmatic SEO is technically legitimate, but it tips into scaled content abuse as soon as pages are just templates with swapped variables and no standalone value. The guardrail: every generated URL must serve its own answerable user question and deliver verifiable added value. Clean crawl budget management through canonicals, internal linking, and sitemaps prevents thin variants from flooding the index. That way, you can scale with AI without crossing the quality and compliance line.

Further reading and conclusion

Google and the EU AI Act judge AI content by the same basic principle: the machine is not the problem; missing added value and missing transparency are. Google devalues scaled content abuse algorithmically and through manual actions, and from August 2, 2026, the AI Act requires labeling of AI text that informs the public, unless a human has reviewed it and someone holds editorial responsibility for it. Both requirements can be addressed with a single lever: a genuine human-in-the-loop editorial process with fact-checking, source citations, original experience, and documented editorial responsibility. Sensible next steps for Austrian B2B sites are a content audit for thin and mass-produced texts, clarifying your role as provider or deployer, defining disclosure texts for deepfakes and for informative AI texts without editorial responsibility, and a deliberate decision in AI crawler management between content protection and GEO visibility.