Check whether AI systems can actually read, interpret, and recommend your site.
GEO Fix team15 min read
Updated:
Your writer pasted a Claude draft into WordPress. The page looks the same as always. Since 2 August 2026 that text can carry a machine-readable provenance mark — invisible, woven into the words themselves.
The question owners are asking is whether that mark now costs them visibility in AI answers. Here is what the documentation actually says, and where the real risk sits instead.
Anthropic has signed the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content, as a provider of both generative AI models and generative AI systems. Its help page describes two techniques. Every quote below is from that page as accessed on 11 August 2026 — it was last modified on 10 August, and Anthropic says three separate times that it will update the page as things ship.
| Output | What Anthropic says it attaches | What a reader sees |
|---|---|---|
| Generated text | An “imperceptible watermark” woven into the text itself, which Anthropic says does not change “the meaning, quality, or readability” | Nothing visible |
| Generated files | Signed provenance metadata following the C2PA open standard — the help page gives “.svg, .png, or .jpg” as examples, not as a complete list | Nothing in the picture itself |
The rollout is not finished, and Anthropic is direct about it. “Claude models launched on or after August 2, 2026 support marking at launch.” Models released before that date sit inside a legal transition period where Anthropic is “working to add marking support” — with no completion date published. Some platforms and features “may not support certain marking types”.
The text mark travels with copy and paste and “may persist through some editing”. It may not survive heavy editing, paraphrase, translation, mixing into other writing, or passages too short to carry a reliable signal.
The interesting part is not the policy, it is the mechanism — and the mechanism settles the question better than an absence of evidence ever could.
| Property of the mark | Anthropic’s wording | Why it matters here |
|---|---|---|
| Applied at the model level | “Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from.” | The mark is baked into generation, not added by your site — so it says nothing about your domain |
| Marks processing, not authorship | “Detecting a Claude mark tells you that the content may have been processed by Claude. It does not, on its own, confirm the full provenance of the content.” | A human draft that Claude proofread can carry a mark; a blunt filter would hit edited expert copy |
| Absence proves nothing | “Lack of a detected mark doesn’t mean the content wasn’t AI-generated or processed.” | Unmarked never meant human-written — and as marking spreads, it will mean even less |
| Detection is not shipped | Anthropic is “working to enable” detection and will share details “in forthcoming technical documentation” | Today nobody outside Anthropic can reliably check a page for a Claude mark |
Read the first row again. The watermark is applied when the model generates the text, and it rides with that text wherever it goes. It is not attached by your publishing stack, it says nothing about your domain, and it is identical whether the sentence ends up on a strong site or a spam farm. A search system could in principle detect it — but detecting it would tell that system something about the sentence, not about you.
This is the question that matters commercially: when a buyer asks an assistant for a vendor, does a provenance mark decide whether your business gets named? Here is what each provider publishes, and what is verifiably absent from those same pages.
| System | What its own documentation says decides inclusion | Watermarks or provenance mentioned? |
|---|---|---|
| Google AI Overviews and AI Mode | A page “must be indexed and eligible to be shown in Google Search with a snippet”, and “[t]here are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary” | No occurrence of watermark, provenance, C2PA or SynthID on the page |
| ChatGPT search | Ranking “is based on a number of factors designed to help users find reliable, relevant information”, with “no way to guarantee top placement”; inclusion requires allowing OAI-SearchBot and letting your CDN pass OpenAI’s published IP addresses | No occurrence of watermark on the page |
| Perplexity | Hybrid lexical and semantic retrieval, prefilters that drop “clearly non-responsive or stale content”, rerankers, a freshness model, and prioritising “documents from authoritative domains” in hot storage | No provenance, watermark or content-credential signal appears in the write-up |
| Anthropic | Marking is framed purely as transparency under the EU AI Act — Anthropic makes the mark, it does not rank or cite anything | No occurrence of rank, citation, SEO or visibility on the page |
Two of those rows deserve quoting in full. Google’s AI-features documentation states that “[t]here are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary” beyond being indexed and snippet-eligible. And OpenAI’s ChatGPT search help page says that to be included “it is important to allow OAI-Searchbot to crawl your site, and ensure your site host and/or content delivery network allows traffic from our published IP addresses”, while adding that “[t]here is no way to guarantee top placement”.
That second clause is the one that catches people out. A correct robots.txt does not save you if your CDN, WAF or bot manager is quietly refusing the traffic. We see that gap far more often than we see anyone harmed by a watermark.
On Perplexity, be precise about what is missing. Its engineering write-up on the Search API — published September 2025, as retrieved 11 August 2026 — documents plenty of quality signals: retrieval scoring, rerankers, a freshness model, and prioritising “documents from authoritative domains” in hot storage. What it never documents is a provenance, watermark or content-credential signal. “Perplexity documents no quality signals” would be false; “no provenance signal” is what the page actually supports.
Keep three things apart that get blurred together constantly:
| What people mix up | What it actually is | Does any provider document it as a citation switch? |
|---|---|---|
| Claude marking its own chat replies | Transparency on the assistant’s output | No — that mark sits on the answer, not on your website |
| A mark on a page you published | A signal that Claude may have processed that text | No — not in any of the four documents reviewed here |
| Whether AI can retrieve your site at all | Crawlers allowed, pages fetchable and parseable | Yes — Google and OpenAI both state it as a condition of inclusion |
A correct robots.txt does not settle that. The free Express Check tests live access in about a minute.
Not for being AI-written. Google’s guidance answers it directly: “Appropriate use of AI or automation is not against our guidelines.” The limit sits in the same post: “Using automation—including AI—to generate content with the primary purpose of manipulating ranking in search results is a violation of our spam policies.”
And on whether it helps: “Using AI doesn’t give content any special gains. It’s just content. If it is useful, helpful, original, and satisfies aspects of E-E-A-T, it might do well in Search. If it doesn’t, it might not.” The production method is not the axis — quality is. We walk through that split in is AI content bad for SEO? and in Google’s scaled content abuse policy.
One correction worth making, because the opposite circulates widely: SynthID is not a media-only technology. Google says SynthID watermarks “images, audio, text or video”, and it has open-sourced its text watermarker. What is currently limited to images, video and audio is the consumer-facing verification surface — the “is this made with AI?” lookup. SynthID Text is applied by Google’s own models to text they generate; it is not applied retroactively to prose written elsewhere, and Google has published no ranking or citation consequence for any modality.
Here is where most coverage of this topic gets it backwards. The marking duty is not coming — Article 50 of the EU AI Act applies from 2 August 2026, nine days before this was published.
Article 50(2) puts machine-readable marking on providers of generative AI systems — that is Anthropic’s job, not yours. But publisher-side disclosure is not a separate legal question, however often it is described as one. It is the second subparagraph of Article 50(4), in the very same article.
Deployers of an AI system that generates or manipulates text which is published with the purpose of informing the public on matters of public interest shall disclose that the text has been artificially generated or manipulated.EU AI Act, Article 50(4)
Two things follow. The trigger is narrow: text published to inform the public on matters of public interest, not your product descriptions or advertising copy. And the same subparagraph carves out an exception — the duty does not apply where the content “has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication”.
The Code of Practice itself is voluntary, and signing is one recognised route to demonstrating compliance rather than compliance in itself — organisations that comply by other means “will have to demonstrate that those measures are adequate”. The Commission’s signatory list is updated continuously; as of 11 August 2026 its illustrative examples of provider signatories include Aleph Alpha, Anthropic, Black Forest Labs, Cohere, Google, Meta, Microsoft, Mistral, OpenAI and Synthesia. Signing means a company committed to the Code. It does not mean every signatory already watermarks text the way Anthropic describes.
Anthropic adds a line aimed straight at businesses building on its models: if you deploy Claude inside your own product, you “should independently assess what Article 50 requires of your products and services”.
Do not rebuild your content process around beating a watermark that no documented system uses against you, and that nobody outside Anthropic can currently detect anyway. Rebuild it around pages only your business could publish, on a site AI systems can actually reach. That is the core of generative engine optimization.
| GEO factor | Why it decides whether you can be cited at all |
|---|---|
| Crawler access | AI cannot cite what it cannot fetch. OpenAI is explicit that a correct robots.txt is not enough — your host or CDN also has to pass its published IP addresses |
| Clear entity information | An assistant has to identify the business: name, services, location, who you serve. A vague homepage gets skipped |
| Structured data | Labels help machines interpret facts that already appear in the visible text. Not a ranking cheat, and not a substitute for the facts |
| llms.txt | An optional machine-readable file some teams add. Useful in a fuller readiness setup; not established as a citation factor and not required by Google for AI features |
| Original evidence | Prices, process, constraints, results — the things an answer engine has a reason to reuse. Interchangeable copy is easy to skip, mark or no mark |
GEO Fix does not look for watermarks, and no honest tool would sell you a watermark score. It checks whether AI systems can reach you, and where the report shows gaps it generates the files — robots instructions, an llms.txt-style file, schema — plus copy-paste steps for WordPress, Shopify, Wix and Webflow. Access does not guarantee a citation; platforms choose their sources. It removes a blocker that ends the conversation before it starts.
Chasing an unmarked page while your CDN quietly blocks OpenAI is solving the wrong problem.
Not in any documentation we could find. Reading the published ranking, crawler and AI-feature docs of Google, OpenAI, Perplexity and Anthropic on 11 August 2026, none names watermarking or provenance metadata as a factor. The mark is also applied at the model level, so it describes the text rather than your site.
There is no public way to check a page for a Claude mark today. Anthropic says it will support users and third parties in detecting its marks “as the Code requires”, but it is still “working to enable” that and commits only to publishing details “in forthcoming technical documentation”, with no date. OpenAI’s ChatGPT search help page describes crawler access and ranking factors and never mentions watermarks.
No, and Anthropic says so plainly: a detected mark means the content “may have been processed by Claude” and “does not, on its own, confirm the full provenance of the content”. People use Claude to proofread, translate and summarise — the output can carry a mark while the ideas and text came from someone else.
No — that inference is even weaker. Anthropic states that a “[l]ack of a detected mark doesn’t mean the content wasn’t AI-generated or processed”: the model may predate marking support, the text may have been edited or translated, the passage may be too short, or metadata may have been stripped by a format conversion or a screenshot.
No. Treat it as a drafting tool. Google’s published position is that appropriate use of AI is not against its guidelines, and that AI use gives content no special gains either way. What to stop is using any model as a publish button for thin, interchangeable pages.
Google does not require an “AI-written” badge to rank. Under EU law it depends on what you publish: Article 50(4) requires deployers to disclose artificially generated or manipulated text published to inform the public on matters of public interest, unless it went through human review or editorial control with a person holding editorial responsibility. That is a question about your specific publishing and worth putting to counsel — but it is the same article, not a separate one.
Key takeaways:
Free · 2 minutes · no card
The free Express Check tests whether AI systems can actually fetch your pages: crawler access, robots.txt, llms.txt and structure gaps. A clear HTML report by email in under a minute. No card, no account.
No card, no account.
