Skip to content

Is This Website AI Generated? How to Tell in Five Minutes

You can tell whether a website was made with AI from two kinds of evidence, the design defaults nobody changed and the technical traces the tools left in the code. Design signs are hints. A generator tag, invisible characters, leftover chat text or a C2PA manifest in an image are proof.

8 min read

SignWhere you see itWhat it tells youFree check
Purple to blue gradient, gradient textFirst screenA hint, the default colours were keptAI design check
Hero line that fits any businessMain headlineA hintYour own eyes
Emoji as bullets, many spaced dashesBody textA hint, never proof on its ownAI watermark check
Glowing blobs and floating 3D shapesBackgroundsA hintAI design check
Three identical cards with an icon eachAlmost every sectionA hintAI design check
Generator name or builder script in the sourcePage sourceProof of the tool usedAI watermark check
Zero width characters or hidden tag textCopied textProof that text was pastedAI watermark check
C2PA manifest or IPTC source type in an imageImage fileProof for that one imageAI watermark check

To tell if a website is AI generated, look for two things. First the defaults, such as purple gradients, a hero line that fits any business, emoji bullets and three identical cards. Then the traces in the code, such as a generator tag, invisible characters and provenance data in images. Defaults are hints, traces are proof.

We sell AI website templates, so we look at a lot of AI-built sites, including our own drafts. This guide sorts the signs by how much they actually prove, which is the part most lists skip. It also explains what our two free checks look for, and where they reach their limits.

Visual signs, the defaults nobody changed

An AI website builder or a chatbot that writes code has to pick colours, fonts and layouts when nobody tells it what to do. It picks what appears most often in its training data and in the popular frameworks. That is why so many generated pages look like siblings. None of these signs proves AI on its own, because people keep the same defaults by hand. Together they tell you that nobody made a decision.

Purple gradients and gradient text

The most visible sign is a gradient from violet to indigo, often on the headline itself. Values such as #8b5cf6, #a855f7 and #6366f1 come unchanged from the default palette of Tailwind, a framework that generated code uses very often. Gradient text has a second problem, its contrast can no longer be measured cleanly, and in small sizes it gets hard to read.

Generic hero copy

“Transform your business with innovative solutions” could stand on a law firm, a bakery or a software startup. Generated hero lines promise a transformation and name nothing, no place, no price, no customer. Swap in a competitor’s name. If the sentence stays true, it says nothing about you.

Emoji bullets and dash-heavy text

Rows of rocket, sparkle and check mark emoji as bullet points are typical chat output pasted onto a page. So is the dash between two spaces, the favourite punctuation of language models. One dash is fine. Forty on one page say something about who typed the text. Both are patterns, which is why we only ever count them as a hint.

Stock 3D blobs and glow

Soft coloured spots with heavy blur behind the content, floating glossy 3D shapes and spheres that belong to no product. They add light and nothing else. Because they look the same everywhere, they make sections interchangeable, and they cost loading time.

Identical card grids

Three cards side by side, each with an icon, a heading and two lines of text, repeated in every section. It is the most common building block of generated pages. It works, but it tells nothing a list would not, and when it appears three times on one page, nobody thought about the layout. Often every card also carries the same soft shadow and a coloured bar on the left edge.

Technical signs, what the code gives away

Here the evidence gets hard. These traces are either in the page or they are not, and when they are, you can point at the exact spot.

Generator meta tags and builder scripts

Many tools write their own name into a generator line in the head of the page. Open the page source with Ctrl+U or Cmd+Option+U and search for “generator”. Where the line is missing, the address of a loaded file often gives the tool away, since a site built with Lovable, for example, loads a script from the tool’s own servers. Be fair with builders. Wix, Webflow or WordPress in the source tell you which platform hosts the site, not that an AI designed it. A hand-built Webflow site is not an AI site.

Invisible characters

Zero width characters such as U+200B have no visible width and survive every copy and paste. On a normal web page they have no business in body text, and they usually arrive when someone copies an answer out of a chat window. A rarer and stranger case is the Unicode tag block from U+E0000, which can hide whole readable sentences that no visitor sees. Either way they break search and text comparison, so they are worth removing even if nobody cares about the origin.

C2PA, IPTC and watermarks in images

Images often carry their origin with them. DALL·E, Sora, Adobe Firefly, Google Imagen and Gemini embed a signed C2PA manifest, also known as Content Credentials. Midjourney writes the IPTC value trainedAlgorithmicMedia into the metadata. Stable Diffusion setups such as Automatic1111 store the full prompt as a text field in the PNG. This is the strongest proof there is, because it is a statement by the maker, not a guess. Google’s SynthID sits in the pixels themselves and can only be read with Google’s own tools.

Leftover chat text

Sentences like “As an AI language model”, placeholders like “[Your Company Name]” and visible Markdown, asterisks around a word or a hash at the start of a line. These are not style questions, they are accidents. Someone pasted an answer and did not look again. It is live more often than you would think.

Why it matters for trust and conversion

People judge the look of a website in a fraction of a second. A study by Gitte Lindgaard and colleagues at Carleton University, published in 2006, found that visitors rate the visual appeal of a page after 50 milliseconds, and that this first rating matches the one they give after a longer look. A page that looks like a default is filed as a default before anyone reads a word.

The word AI itself can cost sales. In 2024 researchers at Washington State University ran experiments with more than 1,000 US adults and found that mentioning artificial intelligence in a product description lowered purchase intention, more strongly for risky purchases such as medical or financial services. For a law firm, a practice or a financial advisor, a site that looks machine-made raises the same doubt.

Search is the other side. Google has said since 2023 that it rewards helpful content however it was produced. What it pushes down is thin, interchangeable content, which is exactly what an AI writes when it has no facts and no rules. Our guide on whether an AI website is good for SEO covers this in detail.

How to check a website in five minutes

  1. Look at the first screen. Gradient, glowing blobs, a headline without a place or a number? Note it, but do not judge yet.
  2. Open the page source. Search for “generator”, for the names of known builders and for script addresses that point to a tool.
  3. Run the AI design check. It reads the public source of a home page and checks eight defaults, from the standard palette and gradient text to blur spots, a single stock font, spaced dashes and empty phrases. The score out of ten is not a quality grade, it shows how much of the page is a default. One check per day and device, and the address is not stored.
  4. Run the AI watermark check. It loads exactly one page and up to six of its images and runs seven checks, from invisible characters and hidden tag text to C2PA, IPTC, builder traces and leftover chat text. Five count as proof and show the location, language patterns count only as a hint, and a label on the page counts as the good finding. It works without a language model, so the result is the same every time. Three checks per day and device.
  5. Check a draft before you publish it. The text checker finds and removes invisible characters in pasted text, and it runs entirely in your browser.

How to make an AI-built site look hand-made

The goal is not to hide that you used AI. The goal is a site that could only belong to your business.

  • Mix one colour of your own and use it everywhere. Grey in five shades can carry the rest.
  • Set the accent with one italic word instead of a gradient in the headline.
  • Use real photos of your work, your team and your room. Keep the C2PA label on any image you did generate, it is honest and, under the EU AI Act, increasingly expected.
  • Write sentences with a number, a name or a case in them. “Four weeks from enquiry to live” beats “tailored solutions”.
  • Vary the sections. What matters more gets more space, the rest gets quieter.
  • Remove emoji bullets and stray dashes, and type text over instead of pasting it out of a chat window.
  • Write your rules down. An AI falls back to defaults when nobody tells it otherwise. Our templates carry a guideline file and a knowledge folder for exactly that reason, so the twentieth page follows the same rules as the first. You will find more on the idea in our guide to the AI website and in the list of the best AI website templates.

What no check can tell you

Every honest check has limits, and they belong in the result. Metadata disappears as soon as a content system or image service resizes a picture, so a clean image can still be generated. Text that someone read in a chat window and typed again leaves no trace, which is the right way to work with AI anyway. Style detectors that print a percentage are guessing, and they hit people who write plainly most often. A finding tells you which tool was used, not whether the site is good. A clean, fast, honest site built with AI is better than a hand-made one that talks past its customers.

Frequently asked questions

Often, yes. The reliable signs are technical, such as a generator tag or a builder script in the source, invisible characters in the text and provenance data in the images. Design signs such as purple gradients and generic headlines are only hints, because people use the same defaults without any AI.

Our AI watermark check reads one page and up to six of its images and shows every trace it finds with its exact location, without a language model and without a guess. The AI design check reads the source of a home page and shows which of eight design defaults it carries. Both are free.

Not for proof. Style detectors estimate how likely a text is under a language model and err in both directions, especially with people who write plainly. A zero width character or a C2PA manifest either is in the file or is not, which is why we check for those instead.

Google says it rewards helpful content however it was produced. What loses rankings is thin, interchangeable content, and that appears easily when an AI writes without a brief or real facts. Origin is not the problem, generic content is.

It depends on what it shows. Under the EU AI Act, providers of generative tools must mark their output in a machine-readable way, and deepfakes must be disclosed. Ordinary marketing copy that you edited and published under your name usually needs no label. If you are unsure, ask a lawyer.

Mix one colour of your own, use real photos, write sentences with numbers and names, vary the sections, remove emoji bullets and stray dashes, and type over chat text instead of pasting it. A template with written design rules keeps the AI from falling back to defaults.

Keep reading