AI Watermarking and SEO: What the EU Law Actually Says

Published August 20, 2026 by AldoMedia

On August 2, Europe's AI transparency rules took effect, and the marketing world started predicting that hidden AI watermarks would sink your rankings. We read the regulation. It does not say what people think it says.

There is a version of this story going around that reads roughly like this: the EU now forces AI companies to hide watermarks in generated text, detectors can spot them with over 90 percent accuracy, and search engines will start pushing watermarked pages down the results. Adapt or lose your rankings.

Almost every load-bearing part of that is wrong. Not spun, not exaggerated, just factually not what the documents say. And the parts that are true are more useful than the scare version, because they point at something you can actually act on.

Here is what the regulation says, what the AI companies have actually done, what Google has actually said, and the one clause that genuinely touches a business blog.

What the Law Actually Requires

The relevant text is Article 50 of the EU AI Act. The operative obligation on providers is that they must ensure the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated.

Read that again, because the wording matters. It says marked in a machine-readable format. It does not say watermarked. The word "watermark" does not appear anywhere in Article 50.

Watermarking does appear once in the surrounding material, in Recital 133, and it appears there as one item on a list of options: watermarks, metadata identifications, cryptographic methods for proving provenance, logging methods, fingerprints, "or other techniques, as may be appropriate." That trailing phrase is the giveaway. The list is illustrative, not mandatory.

The duty is also explicitly qualified. It applies as far as it is technically feasible, taking into account the specificities and limitations of different content types, the cost of implementation, and the generally acknowledged state of the art. In plain terms, the law describes an outcome and lets the industry work out the method.

That distinction is not pedantry. A rule that says "embed a statistical watermark in every sentence" and a rule that says "make it machine-detectable somehow, where that is technically feasible" lead to completely different worlds, and only one of them is the actual law.

The Deadline Almost Nobody Is Mentioning

August 2, 2026 is a real date. The Article 50 transparency obligations did apply from that day.

But there is a second date that matters more right now and it is largely missing from the coverage. Under the Digital Omnibus regulation that entered into force on July 27, 2026, a four month transitional period runs to December 2, 2026 for the machine-readable marking duty, and it applies only to generative AI systems that were already on the market before August 2. Anything placed on the market on or after August 2 had to comply immediately.

So the industry is mid-transition right now, not finished. If you are judging what providers "are doing" based on August headlines, you are reading a snapshot of a period that does not close until December.

For scale, the penalty tier is worth knowing precisely, because it also gets inflated. Breaches of Article 50 fall under the fine bracket of up to 15 million euros or 3 percent of worldwide annual turnover, whichever is higher. The much quoted 35 million and 7 percent figure belongs to a different article covering outright prohibited practices, not transparency.

Who Has Actually Signed What

Alongside the regulation there is a voluntary Code of Practice on Transparency of AI-generated Content. The Commission concluded in July 2026 that the Code adequately covers the relevant obligations, and it published the signatory list on July 31.

Roughly 190 organizations signed. The provider section, which covers marking and detection, includes Anthropic, Google, Meta, Microsoft, OpenAI, Mistral, Cohere and others. The deployer section, covering labeling, drew names like Getty Images, Lenovo and Lufthansa.

That is a real, citable fact, and it is the accurate version of "the big AI companies are on board." What it is not is evidence that any specific one of them has shipped text watermarking. Signing a voluntary code that describes marking obligations is a different act from deploying a particular technique.

What Is Genuinely Being Watermarked Today

This is where the popular version goes furthest off course, in both directions. The honest picture as of today:

• Google does watermark some of its own text. Google's SynthID documentation states it has been expanded to watermarking and identifying text generated by the Gemini app and web experience. So "nobody watermarks text" would be false.
• OpenAI publicly separates the two cases. It has said it is researching more effective provenance techniques for text, while having committed to deploy mechanisms for audio and visual content. Its current content provenance tooling covers images and audio. Text is absent.
• For other providers, there is no public documentation of text watermarking. Anthropic and Meta have not published anything describing a text watermark on their model output. An absence of documentation is not proof of absence, but it is also not a basis for asserting that it is happening.

Notice what none of that supports: the claim that all the major labs quietly switched on text watermarking on a particular date this month. Image, audio and video watermarking is genuinely widespread. Text watermarking is the exception, not the rule.

The Detection Claim You Should Not Repeat

The version circulating says detectors identify AI text with over 90 percent accuracy, even after paraphrasing or editing. Be very careful with that number, because the best evidence available runs the other way.

OpenAI built an AI text classifier and then withdrew it in July 2023, stating plainly that it was no longer available due to its low rate of accuracy. The published figures: it correctly identified 26 percent of AI written text, while incorrectly labeling 9 percent of human written text as AI. The company also stated that it is impossible to reliably detect all AI-written text and that such text can be edited to evade a classifier.

That 9 percent false positive rate is the part that should worry any business owner. It means a detector can flag work your team wrote by hand. Research has repeatedly found these tools are hardest on writing by non-native English speakers, which turns a "quality check" into a fairness problem.

Watermark detection is a different and more reliable technique than blind classification, because it looks for a signal that was deliberately inserted. But it only works on content from a system that inserted one, and published research shows the signal degrades under paraphrasing and can be stripped or forged. It is a provenance tool, not a lie detector.

Will Any of This Hurt Your Google Rankings?

The short answer is that there is no evidence for it, and a fair amount of evidence against.

Start with Google's own position, which has been consistent for years: it rewards helpful, original, people-first content, and it focuses on the quality of content rather than how it was produced.

Then look at the document that actually shapes evaluation. Google's Search Quality Rater Guidelines run to 182 pages. Search them for watermark, SynthID, C2PA, Content Credentials or provenance and you get zero results. Not vague language, not a soft mention. Nothing at all.

What the guidelines do say is the opposite of the panic: the use of generative AI tools alone does not determine the level of effort or the page quality rating. Raters are told to judge the page, not guess at the tooling.

The independent data is messier than either camp admits, and it is worth being straight about that. Ahrefs research analyzing a million pages found AI use correlated with somewhat lower search performance, though the gradient is slight: AI-detected content averaged 27.1 percent of the page at position one and 30.9 percent at position ten. Other analyses have found the relationship to be effectively zero. Studies in this area also lean on AI detectors, which we have just established are unreliable, so treat all of these numbers as directional at best.

Here is what none of them show: a watermark-driven penalty. No study has found one, because no search engine has documented such a signal existing. The plausible reading of the mild negative correlation is not that Google detects machine writing and demotes it, but that pages produced without much human effort tend to be less useful, and Google has always been in the business of measuring that.

What Google and Bing do target is scaled, unsupervised output. Bing's guidance warns that large-scale content generated without oversight, quality control, or editorial review often lacks usefulness, accuracy and originality, and may be excluded from indexing. Read that carefully too: the trigger is the absence of oversight, not the presence of a machine.

The Clause That Does Affect a Business Blog

Here is the part worth your attention, and it is not the provider watermarking rule at all.

Article 50 also places a duty on deployers, meaning the people publishing. AI-generated or manipulated text published to inform the public on matters of public interest must be disclosed as artificially generated.

Now the exemption, which is the whole ballgame for most publishers. That duty does not apply where the content underwent human review or editorial control and a natural or legal person holds editorial responsibility for it.

That is a remarkably practical piece of drafting. It says, in effect: if a person read it, stands behind it, and is accountable for it, it is your publication, not a machine's output. A locally focused business blog with a named author who edits and takes responsibility generally sits outside the disclosure duty on that basis.

Two honest caveats. This is European law, so its direct reach depends on whether you are targeting or serving an EU audience, and the Commission's final guidance indicates that purely incidental or unforeseeable access from the EU should not by itself pull you in. And none of this is legal advice. If you publish at scale into European markets, ask an attorney rather than a blog.

But the shape of the rule is worth absorbing regardless of jurisdiction, because it points the same direction every search and AI system already points: human editorial responsibility is the thing that counts.

What This Means for GEO and AEO

Generative Engine Optimization and Answer Engine Optimization are about getting cited by AI systems and surfacing in direct answers. The tempting story is that watermarks will create a trust hierarchy where marked content gets cited less. That is a reasonable hypothesis. It is not currently an observed fact, and it is worth keeping those apart.

What is actually documented:

• No major AI engine publishes provenance-based citation criteria. Perplexity's public explanation of how it works says only that it gathers information from authoritative sources and cites them. There is no published watermark or provenance rule from any of them.
• Synthetic sources are already being cited at scale. A 2026 audit of four AI search engines across 712 queries found roughly 16 percent of cited sources appeared to be AI-generated. Whatever filtering exists today, it is clearly not excluding machine-written pages.
• AI citation does not simply mirror the rankings. One large study of trending queries found around 30 percent of domains cited in AI Overviews did not appear in the first page of results shown alongside them, which suggests a source selection mechanism distinct from ordinary ranking.
• Overlap with traditional results is shrinking. Ahrefs measured 38 percent of pages cited in AI Overviews also ranking in the top 10 as of early 2026, down from roughly 76 percent the prior year.
• The tactics are not proven. A survey of 45 GEO studies concluded that no reviewed technique shows a stable, longitudinal, cross-platform causal effect on discoverability. Anyone selling you a guaranteed GEO formula is ahead of the evidence.

The practical read: search visibility still feeds AI visibility, which is the argument we made in SEO is the foundation of AI visibility, but the two are drifting apart and need measuring separately. Google now reports AI performance in Search Console, which we covered in our breakdown of the AI report, and it is the only first-party measurement any of us have. If the difference between the three disciplines is still fuzzy, our GEO versus SEO explainer and the SEO, AEO and GEO guide cover the distinctions.

What to Actually Do

Nothing on this list is new, which is rather the point. The transparency era rewards the things that were already working.

• Put a real, named human on every post. An author with credentials, a bio and accountability satisfies the EU exemption, matches what Google's raters are told to look for, and is the single highest-leverage change most sites can make.
• Keep editorial oversight and be able to show it. The legal exemption and the search guidance both hinge on review, not on tooling. Whatever your process is, make it real.
• Publish what a model cannot. Your own project photos, your pricing, your client outcomes, your regional knowledge. A language model can produce a competent article about roofing. It cannot produce your job from last Tuesday in Amherst.
• Do not buy detector-based tooling as a compliance product. With a documented 9 percent false positive rate on the best-known attempt, you would be flagging your own writers.
• Disclose AI assistance if you want to, for trust rather than fear. A short editorial note costs nothing and reads as confidence.
• Measure AI visibility separately from rankings. They are diverging. Track both.
• Do not block the AI crawlers by reflex. If you want to be cited, you have to be readable, a tradeoff we worked through in blocking ChatGPT in robots.txt.

The Bottom Line

The underlying instinct in the watermark story is right, even though the facts supporting it are mostly wrong. We are moving toward an information ecosystem that sorts by verifiability, and content nobody will put their name to is going to have a harder time.

But that sorting is not being done by a hidden statistical marker in your paragraphs. It is being done the way it has been done for years: by whether a real person stands behind the work, whether the page contains something that only you could have written, and whether anyone finds it useful enough to come back.

The watermark is not a filter on your rankings. Editorial responsibility is. That was true before August 2 and it will be true long after the December transition closes.

Sources

• EU AI Act, Article 50 transparency obligations (full text)
• European Commission, regulatory framework for AI
• Google DeepMind, SynthID
• OpenAI, AI text classifier and its withdrawal
• Google Search Central, creating helpful content
• Google Search Quality Rater Guidelines (PDF)
• Ahrefs, Google does not punish AI content (1,000,000 pages analyzed)
• Bing Webmaster Guidelines

Frequently Asked Questions

Does the EU AI Act require AI watermarking?

No. Article 50 requires that outputs be marked in a machine-readable format and detectable as artificially generated or manipulated. The word watermark does not appear anywhere in Article 50. Watermarking is listed in Recital 133 as one option among several, alongside metadata identification, cryptographic provenance methods, logging, fingerprints and other techniques.

Did the AI transparency rules take effect on August 2, 2026?

Yes. The Article 50 transparency obligations applied from August 2, 2026. There is a detail most coverage misses: a four month transitional period runs to December 2, 2026 for the machine-readable marking duty, and it applies only to generative AI systems that were already on the market before August 2. Anything placed on the market on or after that date had to comply immediately.

Will Google penalize my website for using AI to write content?

Google has never said that it does. Its published position is that it rewards helpful, original content however it is produced. Google's Search Quality Rater Guidelines contain no mention of watermarks, SynthID, C2PA or provenance of any kind, and they tell raters directly that the use of generative AI tools alone does not determine the level of effort or the page quality rating.

Can AI detectors reliably tell whether text was written by AI?

Not reliably. OpenAI withdrew its own AI text classifier in July 2023 because of what it called a low rate of accuracy. That tool correctly identified 26 percent of AI written text while incorrectly flagging 9 percent of human written text as AI. Treat any vendor claiming near perfect detection, especially after paraphrasing, with real skepticism.

Does my business blog need an AI disclosure label?

For most American small businesses, no. The rule that covers published text applies to AI generated text put out to inform the public on matters of public interest, and it exempts content that went through human review or editorial control where a person or company holds editorial responsibility. A locally focused business blog that a human edits and stands behind generally sits outside it. This is general information, not legal advice.

What should I actually change about my content strategy?

Very little, and none of it is about watermarks. Keep a named human author who reviews and stands behind every piece, publish things a model cannot generate such as your own data, photos and client results, and keep earning the search visibility that AI systems draw from. The measurable factors have not changed just because a labeling rule arrived.