Does this landing page carry an AI watermark?
Paste the address of a single page. We read its text, its images and its source, and show every trace an AI left behind, together with the spot it sits in. We do not guess.
Free · no sign-up · no account
Three checks per day and device. We check this one page, not your domain, and we do not store the address.
Loading and checking the page …
Seven checks, and every one shows its evidence
Five of them can prove something, one is explicitly only a hint, and one goes looking for something good.
-
Invisible characters in the text
- What it finds
- Thirty-six characters with no width, from the Zero Width Space to the variation selector. They sit in the visible text and show up on no screen, not even for the person who pasted them in.
- Why it counts
- They are the only real watermark a text can carry. Some providers set them on purpose, most arrive by accident when someone copies out of a chat window. Either way they do damage, because they break every search and every comparison.
- What the result says
- Every character found, with its code point, how often it occurs and the spot it sits in.
- Weight
- Proof
-
Hidden plain text
- What it finds
- Characters from the Unicode tag block. They mirror the entire ASCII range, they are invisible, and they can therefore be strung together into readable sentences nobody sees.
- Why it counts
- In the wild that is either a disclosure or an instruction aimed at the language model that reads the page later. Both belong to the things you want to know about your own page.
- What the result says
- The hidden text, decoded and written out in plain sight.
- Weight
- Proof
-
Where the images come from
- What it finds
- Up to six images on the page are loaded and read at the byte level. We look for a C2PA manifest, the IPTC value for generated images, the name of the image tool, and the text fields some programs use to store the full prompt.
- Why it counts
- An image carries its origin with it if its maker put it there, and the big ones do. This is the strongest evidence of all, because it is a statement by the manufacturer and not a guess by us.
- What the result says
- The file, the marker, the tool we recognised and, if it is in there, the prompt.
- Weight
- Proof
-
Build tool in the source
- What it finds
- The generator entry in the head and the addresses of every file the page loads. We look for sixteen AI generators and ten site builders.
- Why it counts
- Anyone building with an AI generator almost always loads a file from its servers, and that address sits in the source. Builders such as Wix or Webflow are reported separately and explicitly not as proof of AI. A hand-built Webflow site is not an AI site.
- What the result says
- The name of the tool and the evidence we recognised it by.
- Weight
- Proof and plain fact
-
Leftovers from a chat window
- What it finds
- Sentences only an assistant writes, placeholders nobody replaced, and visible markdown, meaning asterisks around a word or a hash at the start of a line.
- Why it counts
- These are not questions of style, they are accidents. Anyone with "As an AI language model" or "[Your Company Name]" on their home page pasted an answer and stopped looking. It is live more often than you would think.
- What the result says
- The sentence we found and the context it sits in.
- Weight
- Proof
-
Language patterns
- What it finds
- Dashes between spaces, filler phrases from a list of thirty, emoji used as bullets, and the "not just X, but Y" sentence shape. We measure density, not occurrence.
- Why it counts
- This is the one check that proves nothing, and it is never counted as proof. A dash is allowed. Forty of them on one page still say something about who typed the text.
- What the result says
- The count and the density per thousand words, so that "it occurs" becomes "it stands out".
- Weight
- Hint only
-
Disclosure
- What it finds
- Does the page say for itself that AI was involved? We look for a note in the visible text, a declaration in the head of the page, and a machine-readable marking in the images.
- Why it counts
- Since 2 August 2026 that is no longer a matter of taste. It is the only check here where a finding is good news and the absence is what raises the flag.
- What the result says
- Whether a disclosure exists for people, whether one exists for machines, and what separates the two.
- Weight
- The good finding
Watermark or trace
Two things that get mixed up constantly. One is placed so it will be found. The other was simply left behind.
| Watermark | Trace |
|---|---|
| Set on purpose, so that it can be found | Left behind because nobody cleared up |
| A C2PA manifest in the image, signed by the maker | The prompt sitting in a PNG text field |
| The IPTC value trainedAlgorithmicMedia | The site builder's script in the source |
| Zero-width characters, sprinkled deliberately | "As an AI language model" in the middle of the home page |
| Survives copying and sharing | Disappears the moment someone rewrites the sentence |
| This is called a watermark | This is called a trace |
Which tool leaves what
The state of play we check against. Where something is uncertain, the row says so.
| Tool | What it leaves behind | Where it sits |
|---|---|---|
| ChatGPT, text | For a while a narrow no-break space, U+202F | In the text. OpenAI says it is not a watermark, and it has been rare since spring 2025 |
| DALL·E and Sora | A signed C2PA manifest | In the image file |
| Adobe Firefly | C2PA on images that were fully generated | In the image file |
| Google Imagen and Gemini | SynthID in the image itself, plus C2PA | In the pixels and in the file |
| Midjourney | The IPTC value trainedAlgorithmicMedia, no C2PA | In the metadata |
| Stable Diffusion with Automatic1111 | The full prompt including settings | As a text field in the PNG |
| ComfyUI | The entire workflow as JSON | As a text field in the PNG |
| Lovable | A script from cdn.gpteng.co, sometimes attributes in the markup | In the source |
| v0, Bolt, Durable, Mixo and twelve more | Their name in the generator entry or in a file address | In the source |
What we cannot see
Six things this check does not do. They are listed here because a tool that hides its limits becomes useless at the first counterexample.
- The watermark inside the image itself Google's SynthID lives in the pixels, not in the metadata, and only Google's own detector can read it. At most we see the note beside it in the file. Anyone promising you otherwise is selling you something.
- Metadata that gets lost on the way Content systems, image services and caches re-encode images and throw away everything that is not a pixel. So an image can be generated and still look clean. Here, no finding never means "not an AI image".
- Text that was typed again Anyone who reads a draft from a chat window and then types it themselves leaves nothing behind. That is, by the way, exactly the right way to work with one, not the loophole it looks like.
- The writing style We do not guess who wrote a sentence. Style detection is probability, it errs in both directions, and it hits people who write plainly hardest of all. That is why language patterns here are marked "hint only".
- Content that only appears in the browser We read the HTML as delivered. Whatever a page loads afterwards is not in there. On an application that assembles all of its content from a script, we see less than on a classic page.
- The rest of the website We check exactly the one address you paste, plus the images it embeds. No subpage, no link, no crawl of the domain. That is a promise, not a technical limit.
Disclosure, since 2 August 2026
That is the day Article 50 of the AI Act took effect, the part about transparency. It requires providers of generative systems to make their output machine-readable as artificially generated. Systems already on the market before that date have until 2 December 2026 to meet it.
For a website operator the second half matters most. Anyone publishing image, audio or video that resembles real people or events has to disclose it, and something similar applies to published text on matters of public interest. For a joinery's product page nothing follows from this. For a portrait that was never photographed, quite a lot does.
In practice that means two things. A C2PA manifest in the image is a disclosure for machines, and it appears on its own as long as you do not strip it while resizing. A sentence on the page is a disclosure for people. The check tells you whether each one is there. It does not tell you whether you need it.
As of 15 September 2026. This is an explanation, not legal advice. The code of practice on transparency was finalised on 10 June 2026, and the binding text is published by the European Commission.
The Commission's guidelines How language models read your page
What this check does with your address
You are typing someone's address into someone else's form. These are the six promises attached to that.
- One page, not a domain The check fetches the one address and up to six images it embeds. It follows no link and loads no subpage. That keeps the tool from becoming a scanner someone points at other people's servers.
- No language model in the way The check compares bytes and code points, it does not read in the sense of understanding. Text from a foreign page can therefore never turn into an instruction, and the result costs nothing and is the same every time for the same page.
- Every fetch validated on its own The target, every redirect and every image address run through the same validation, and the connection goes to exactly the address that was validated. Private networks, a cloud metadata service and anything that is not http or https are ruled out.
- The address does not stay behind It is used for the check and discarded afterwards. All we count is how often the tool runs per day, as a number with no address next to it.
- Three checks per day and device Enough to change something and see whether it is gone. Too few to use the check as a mass tool against other people's servers. The counter resets at midnight UTC.
- Evidence stays text Whatever comes back from a foreign page is stripped of control characters and inserted as text only, never as markup. A page under test can execute nothing in this result list.
Reference
The terms this is about
Nine entries, each with a statement you can verify. They explain what the check looks for and why the line between proof and assumption is everything here.
Zero Width Space
A character with zero width, code point U+200B. It has its purpose in breaking long words and in writing systems without spaces. In an English paragraph it has no business being there, and because nobody sees it, it survives every copy and paste. It is the most common invisible character on the web.
The tag block, and how text hides in it
Unicode has a range from U+E0000 that mirrors the entire ASCII set, invisibly. A whole sentence can be hidden inside a headline that way. The check lifts those characters back down and writes out the plain text they spell.
C2PA and Content Credentials
A standard for provenance. The image gets a signed manifest embedded that records what it was made with and what was done to it afterwards. DALL·E, Sora, Firefly, Imagen and Gemini put it in, and so does Photoshop on generative fill. It is the strongest statement an image can make about itself.
SynthID
Google's watermark. It does not live in the metadata but in the pixels, and it survives cropping, compression and moderate editing. Which is exactly why nobody outside Google can read it reliably. We do not see it, and we do not claim we do.
IPTC Digital Source Type
A field from the news industry's metadata standard. The value trainedAlgorithmicMedia means fully generated by a trained model. compositeWithTrainedAlgorithmicMedia means real material with generated parts. Platforms such as Pinterest and LinkedIn read this field to label images.
Why text detectors guess
A style detector computes how likely a sequence of words is under a model. It has no access to the truth, only to probability. So it takes plainly written text for machine output and reworked AI text for human. A code point, by contrast, is either there or it is not.
Why metadata disappears
As soon as an image is resized, converted or pushed through an image service, it is written anew and everything but the pixels falls away. A content system does that on every upload. Which is why a finding is proof and a non-finding is not an acquittal.
The generator entry
A line in the head of every page that many tools fill with their own name. It is not mandatory and it can be deleted, but hardly anyone deletes it. Where it is missing, the address of some loaded file usually gives the builder away instead.
Why a finding is not a verdict
A trace says what was used, not whether it is any good. A page built with AI can be clean, fast and honest. A hand-built one can talk straight past its customer. Take the result as a list of places where somebody left something behind.
What to do with the result
Three routes, depending on what the check found.
Common questions
No, and that is deliberate. We check exactly the one address you paste, plus up to six images that page embeds. We follow no link and load no subpage. If you want several pages checked, enter them one after another.
Because every run fetches someone else's servers. Three is enough to change something and see whether it is gone. More would turn a tool into a mass scanner, and that is not what it is built for. The counter resets at midnight UTC.
No. It is used for the check and discarded afterwards. All we count is how often the tool runs per day, without the address beside it.
It means this exact trace sits in this exact spot. What follows from that is written next to every finding. A C2PA manifest in an image is the maker's own statement that the image was generated, and that is hard evidence. A dash is a hint and is never counted as proof here.
No, and the result says so too. Metadata falls away the moment an image is resized, and anyone who retypes a draft leaves nothing behind. A finding is proof, a non-finding is not an acquittal.
For text the character checker does it: paste, clean, copy back. It runs entirely in your browser. Images are a different question: a C2PA manifest is a disclosure, and since the AI Act there are good reasons to leave it in place rather than strip it.
Yes, the check reads the publicly available source and the embedded images, whatever the page was built with. What it was built with is something the check tells you in passing, kept separate from the question of AI. A site builder is not proof of AI.
Because this is where the examples it looks for are written down. Go ahead and check this page: you will get "As an AI language model" and the placeholders from the tables above reported back. That is not a bug, it is precisely the promise. We report what is there, not what we suspect.
Nothing and nothing. There is no sign-up, no account and no model call in the background, so there is no query that costs us money either. What the site knows about you is in the privacy policy.
Clear out the invisible characters and the leftovers, which takes minutes. On the images, decide whether to leave the disclosure in place, and usually you should. If the result shows you that the whole page came out of a default, a template or the AI Website Service will help.
Read on
- Invisible characters in text The counterpart for a draft you paste rather than publish. It works in your browser, so your text never leaves the device.
- The AI design check Eight marks that give away stock design in the source, from the default palette to the gradient headline.
- The AEO check Whether a language model is allowed to read your page, and whether it finds the answers prospects ask for.
- All tools Six free tools, each for exactly one question about your website, all without an account.
- The design chapters How we decide colour, type, spacing and typography before a page gets built. Free to read, no purchase.
Other free tools
Each answers one question about your website, in under a minute, without an account.
Your next web project
Build on a strong foundation.
Website, design rules and seven skills in one package. Explore templates for your next project.
- Complete page structure
- Seven skills included
- German and English
All of this is already done in a template
48 finished industry websites, each with built-in knowledge and seven skills. Structure, copy, structured data and llms.txt are set before you change a single thing.
- A complete website with every subpage, fully written
- Seven skills to guide your AI tool
- The complete code as a ZIP, yours forever
One payment, everything included. We show this at most once.